A socio-technical agent-based simulation model for predicting smart agriculture adoption dynamics Authors: Yahya S. Alotibi Published: 2025 ABSTRACT Traditional technology adoption models in agriculture fail to adequately capture the complex interplay of socio-technical factors that drive farmer decision-making, resulting in limited predictive accuracy and insufficient understanding of diffusion dynamics. Existing approaches predominantly rely on static econometric frameworks or simplified diffusion models that overlook the dynamic social interactions, trust networks, and heterogeneous decision-making processes that characterize real-world agricultural technology adoption. This gap hinders effective policy design and technology deployment strategies. To address these limitations, this paper presents AdoptAgriSim, a Socio-Technical Agent-Based Simulation Model for Predicting Smart Agriculture Adoption Dynamics, which integrates economic, social, and technological dimensions into a unified framework. The model employs multi-agent reinforcement learning and socio-economic network modelling to capture how individual farmers, peer networks, and market forces interact during the diffusion of technology. AdoptAgriSim incorporates a multi-objective decision mechanism that balances rational economic reasoning with social learning shaped by trust-based network structures. Using three real-world datasets from Iowa (USA), Europe, and India, the model achieves 94.2% prediction accuracy for five-year adoption intervals, outperforming existing diffusion and econometric models. It effectively reproduces emergent adoption behaviours such as technology clustering, peer-driven influence cascades, and region-specific diffusion trajectories. Significant contributions include1: a socio-technical model integrating multi-dimensional decision factors2; a reinforcement-based optimiser that accounts for both economic and non-economic objectives3; dynamic network evolution mechanisms reflecting real-world social interactions; and4 extensive validation across diverse agricultural contexts. Results show that social factors contribute 34% more to adoption variance than previously estimated, underscoring the centrality of peer influence and trust networks in accelerating the diffusion of technology. The proposed framework offers valuable insights for policymakers and technology designers, highlighting that strengthening social connectivity and targeted network interventions can substantially accelerate sustainable agricultural transformation. Introduction The global agricultural sector is facing unprecedented challenges in its efforts to meet the ever-increasing demand for food without compromising environmental sustainability and economic viability. Smart agriculture technologies, including precision farming, Internet of Things sensors, autonomous farm vehicles, and data-driven analytics, are promising solutions to address these challenges1. The diffusion of these technologies, however, is occurring much more slowly than expected, revealing significant disparities between various farming communities and within geographical areas2. The diffusion of innovation theory and technology acceptance models are traditional adoption models that have mostly considered adoption as a static phenomenon, and do not take into consideration that farming communities exhibit a non-homogeneous nature3. Recent studies have suggested that the issue may be given even greater weight in considering how socio-technical factors influence adoption patterns. As such, more advanced modelling techniques need to be addressed that can take into consideration both individual decisions and the social processes involved4. These challenges of modelling can be addressed with new freedom thanks to the emergence of agent-based modelling (ABM) as a potent simulation paradigm. ABM can be used to model single individual agents who have their own abilities to make decisions, and with that, single agent agents can give rise to complex system-level behaviour through local interactions5. ABM is capable of describing the heterogeneity of farmers in the context of agricultural technology adoption, as well as the dynamics of social network interactions between farmers and the temporal changes in adoption patterns6. Although the question of whether agent-based approaches should be used to model socio-technical systems has received growing concern, current models in the agricultural field have several shortcomings. To begin with, there is a majority that only captures economic aspects, with little to no participation in capturing social influence mechanisms or peer-to-peer learning processes7. Second, most available models assume fixed network topologies and are therefore limited in their ability to understand the dynamic process of developing social relations, which significantly influences the diffusion of technology8. Third, there has been little focus on modelling the co-evolution of individual attitudes and collective norms, which are vital in assuming a long-term adoption dynamic9. The main limitations of this study relate primarily to the inherent variability in regional data quality and the restricted accessibility of specific agricultural datasets. These limitations can be addressed through emerging advances in multi-agent reinforcement learning and the modelling of socio-technical systems10. Such combinations of adaptive learning algorithms with evolving social network dynamics enable the construction of more realistic and predictive frameworks that capture both individual behavioural adaptation and collective emergence in large farming populations11. Moreover, the growing availability of large-scale agricultural datasets and the rapid increase in computational capability provide a foundation for developing and validating large-scale simulation models that are more scalable, generalizable, and empirically verifiable12. This paper proposes AdoptAgriSim, a new socio-technical agent-based simulation model that features an explicit design to predict the dynamics of adopting smart agriculture technologies. The proposed model confines the current models by adding1: a multi-dimension decision-making framework that combines economic, social, and technical variables2 dynamic mechanisms of social network evolution that capture relationship patterns in the real world3 adaptive learning rules that allow the agents to modify their preferences through learning and interactions with others4 extensive validations that now encompass multiple agricultural settings such as urban, rural, and mixed systems of agriculture13,14. Research gap and novelty: While agent-based modelling has been previously applied to study agricultural technology adoption15, most existing models focus primarily on static economic rationality or fixed network structures16–18, offering limited insight into evolving social interactions and adaptive learning behaviours19,20. The novelty of the proposed AdoptAgriSim framework lies in its integration of dynamic trust-based social network evolution that captures the formation, decay, and reinforcement of relationships among farmers over time. Unlike conventional ABM approaches that employ static decision heuristics, AdoptAgriSim introduces a reinforcement learning-driven decision mechanism allowing agents to adapt preferences based on collective experience and peer influence. Furthermore, the model incorporates multi-objective optimisation that simultaneously balances economic profitability, social cohesion, and technological compatibility, providing a richer socio-technical representation of real-world adoption behaviour. Another distinctive contribution is its cross-regional empirical validation, utilising datasets from Iowa, Europe, and India, which enables comparative insights into diverse agricultural and cultural contexts. Collectively, these innovations advance existing ABM approaches by providing a scalable, data-driven, and behaviorally adaptive framework for modelling smart agriculture technology adoption dynamics. Research questions: This study addresses three specific research questions: RQ1: How can dynamic social network structures and trust-based relationships be integrated into agent-based models to improve prediction accuracy of agricultural technology adoption compared to traditional static diffusion models? RQ2: What is the relative contribution of economic, social, technological, and network factors to adoption variance across different agricultural contexts, and how do these weights vary between industrialized and developing regions? RQ3: Can multi-agent reinforcement learning combined with genetic algorithm optimization enhance the generalizability and transferability of adoption prediction frameworks across diverse geographical and institutional settings? Due to the intrinsic dependencies between the technical, economic, social, and institutional aspects involved in the technology adoption process in the agricultural sector, as illustrated in Fig. 1, technology adoption in the sector is a complex endeavour. Fig. 1An overview of socio-technical factors influencing the adoption of smart agriculture technology, highlighting the complex interplay between individual characteristics, social networks, economic constraints, and technological features. An overview of socio-technical factors influencing the adoption of smart agriculture technology, highlighting the complex interplay between individual characteristics, social networks, economic constraints, and technological features. Related work Over the past few decades, there has been a shift in the literature on technology adoption modelling23,24, reflecting the growing recognition that socio-technical factors play a significant role in shaping the adoption process25,26. This section synthesizes recent developments in agent-based modelling for technology adoption27,28, organized around three key themes: network dynamics in ABM29–31, social learning mechanisms32–34, and validation approaches35–37. Network dynamics in agent-based models The integration of dynamic network structures into agent-based models represents a critical advancement in adoption modelling38,39. Initial research in technology adoption modelling was primarily in the form of econometric models based on the diffusion of innovation theory and technology acceptance models, which attributed adoption factors almost entirely to individual-level characteristics15. However, these approaches failed to capture the evolving social structures that shape information flow and peer influence. Recent developments in agent-based modelling for agricultural decision-making have increasingly emphasized network realism and context-specific adoption dynamics. Will et al.40 developed an ABM framework for agri-environmental policy design in Saxony, Germany, highlighting how agent heterogeneity and spatial factors influence participation rates through network-mediated interactions. Similarly, El Fartini et al. demonstrated that network topology significantly affects diffusion speed and adoption thresholds. Williams et al.16 developed an agent-based model of power relations in food systems to illustrate how power structures and institutional circumstances influence adoption patterns through network-based dependencies. Their work demonstrated that market structures and actor relationships should be explicitly modelled, especially in agriculture where supply chains create complex interdependencies. The issue of social network engagement in technology adoption has become a growing concern in contemporary literature41, with researchers recognizing that static network assumptions fail to capture the co-evolution of social ties and adoption behavior. Ates21 created an agent-based model of industrial cooperation in regional hydrogen off-take, providing insights into how collaborative structures support technology uptake through dynamic relationship formation. The study noted that modelling both competitive and cooperative relationships between agents is crucial for realistic adoption predictions. Social learning mechanisms and behavioral realism Socio-technical aspects in adoption modelling have received much emphasis over recent years. Campfens et al.17 provided a detailed literature review of socio-technical energy transitions, considering several relevant modelling types and applying them to technology adoption situations. They discovered that existing models poorly capture co-evolution between technical and social aspects, thereby limiting their ability to predict long-term adoption dynamics shaped by social learning. In the specific area of precision agriculture, multiple research studies have developed agent models to observe technology adoption patterns. Tech18 discussed the contribution of living labs to the development of responsible innovation in precision agriculture, proposing a socio-technical framework that highlighted the crucial role participatory strategies should play in technology development and adoption. Social learning mechanisms are central to these participatory approaches, as farmers learn from peers through observation, experimentation, and knowledge exchange42–44. The latest developments in machine learning and artificial intelligence have opened new opportunities for enhancing agent-based models of adoption. Dupuis23 addressed the application of abstract argumentation techniques to modelling socio-technical controversies, demonstrating how AI-based reasoning mechanisms can be incorporated into agent-based models to represent multilateral decision-making processes. This method is especially suitable for modelling farmer decision-making, where deliberation, trust dynamics, and peer consultation shape adoption choices. Recent research has further expanded the application of agent-based modelling in smart agriculture through integration with intelligent information systems and distributed resource management frameworks. Ahoa et al.45 analyzed the challenges and solution directions for integrating smart information systems within the agri-food sector, emphasizing the importance of interoperability, data-driven coordination, and adaptive digital ecosystems that support social learning at scale. In parallel, Ma et al.46 proposed a comparative model demonstrating how learning algorithms enable agents to adapt strategies based on observed outcomes from their social networks. Validation approaches and multi-objective optimization The multi-agent models have been promising, especially in complex adoption dynamics. Mutambik19 developed an adaptive traffic management framework that utilizes IoT, incorporating multi-agent systems, and demonstrated the utility of agent-based techniques in modelling dynamic interactions within socio-technical complex environments. The validation methodology employed—comparing simulated outcomes against real-world traffic data—provided a template for agricultural technology adoption validation. Adoption modelling that incorporates multi-objective optimization has also been in the spotlight. Multiple researchers have suggested ways to harmonize numerous incompatible aims in technology adoption, including economic efficiency, environmental sustainability, and social acceptability. These methods are especially applicable to modelling smart agriculture adoption, where farmer criteria are numerous and frequently inconsistent. Validation of such multi-objective models requires not only prediction accuracy but also verification that simulated trade-offs match empirical decision patterns. Research gaps and design implications Although these developments have been made, several significant gaps remain evident in the literature. To begin with, the majority of models represent a specific, narrow subset of the smart agriculture technology ecosystem, either by considering single technologies or particular contexts without sufficient generalization. Second, there has been limited research on modelling the dynamics of adoption processes, particularly how early adopters and opinion leaders influence long-term adoption trajectories through evolving trust networks. Third, the heterogeneity of agricultural systems and farming populations is not well depicted by most models, which is critical when predicting adoption across diverse contexts. Fourth, existing validation approaches focus primarily on aggregate adoption curves, neglecting validation of emergent network patterns and micro-level behavioral mechanisms that drive diffusion. From gaps to solutions: design rationale of AdoptAgriSim To address these identified gaps, AdoptAgriSim integrates four specific design innovations, each directly mapped to a literature gap. Gap 1 (limited generalization) is addressed through a multi-technology framework that simultaneously models competing and complementary smart agriculture technologies, enabling cross-technology diffusion analysis. Gap 2 (static adoption dynamics) is resolved by implementing dynamic trust-based social networks where relationship strengths evolve based on interaction history and observed adoption outcomes, explicitly modelling opinion leader influence through network centrality measures. Gap 3 (insufficient heterogeneity) is tackled via agent-specific utility weights (α, β, γ, δ) calibrated from regional data, capturing diverse decision-making styles ranging from profit-maximizing commercial farms to community-oriented smallholders. Gap 4 (limited validation scope) is addressed through multi-level validation combining aggregate adoption accuracy (94.2%), emergent network pattern reproduction (clustering coefficients, degree distributions), and micro-level behavioral validation (peer influence cascades, technology abandonment rates). This systematic gap-to-solution mapping ensures that AdoptAgriSim’s design directly responds to identified limitations in existing literature while building upon validated methodological foundations. Proposed methodology Agent and network modelling This section outlines the methodology of AdoptAgriSim in detail, specifically the socio-technical agent-based simulation model used to predict the dynamics of smart agriculture technology adoption. The current proposal would combine the paradigms of various models to reflect the intricacies of the interdependence between individual decision-making, social network interactions, and technological diffusion. The architecture of AdoptAgriSim generally consists of four interconnected modules/components: the Agent Modelling Module, Social Network Module, Technology Diffusion Module, and the Decision Integration Module. Each module is implemented with a careful selection of algorithms and data structures that capture the particular issues of the socio-technical adoption process without compromising the computational efficiency required for large-scale simulation. Fig. 2AdoptAgriSim system architecture, illustrating the four main modules and their interactions: Agent Modelling Module, Social Network Module, Technology Diffusion Module, and Decision Integration Module. AdoptAgriSim system architecture, illustrating the four main modules and their interactions: Agent Modelling Module, Social Network Module, Technology Diffusion Module, and Decision Integration Module. The general structure, as illustrated in Fig. 2, consists of four interconnected modules that serve a common purpose: to simulate the complex dynamics of adoption. The architecture in Fig. 2 demonstrates how each module interacts sequentially. Outputs from the Agent Modelling Module serve as inputs to the Social Network Module, while the network structures inform the diffusion and decision-making processes. The foundation of our simulation is the Agent Modelling Module, which models individual farmers as independent agents with dissimilar characteristics, preferences and capabilities in decision-making. Each agent \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document} is characterised by a comprehensive profile vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{a}}_{i}=\left({a}_{i,1},{a}_{i,2},\dots\:,{a}_{i,n}\right)$$\end{document} that encompasses demographic, economic, technological, and social attributes. The demographic factors include age, level of education, experience in working on the land, and farm size. In contrast, the economic factors include level of income, risk-taking propensity, and access to credit. Technological attributes refer to computer literacy, history of technology adoption, and access to technical support. Equation (1) defines the agent’s overall utility function, combining economic, social, technological, and network influences into a weighted decision model. The utility function for agent \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document}, considering technology \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:j$$\end{document}, is defined as:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{U}_{i,j}\left(t\right)={\alpha\:}_{i}{E}_{i,j}\left(t\right)+{\beta\:}_{i}{S}_{i,j}\left(t\right)+{\gamma\:}_{i}{T}_{i,j}\left(t\right)+{\delta\:}_{i}{N}_{i,j}\left(t\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{i,j}\left(t\right)$$\end{document} represents the expected economic benefit, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{S}_{i,j}\left(t\right)$$\end{document} captures social influence effects, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{i,j}\left(t\right)$$\end{document} represents technological compatibility, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{N}_{i,j}\left(t\right)$$\end{document} accounts for network externalities. The parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{i}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}_{i}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\gamma\:}_{i}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\delta\:}_{i}$$\end{document} are agent-specific weights that reflect individual preferences and decision-making styles. Equation (2) quantifies the expected economic benefit of adopting a technology by comparing its estimated gains and costs relative to the agent’s income:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{i,j}\left(t\right)=\sum\limits_{k=1}^{K}{w}_{i,k}\cdot\:\frac{{B}_{i,j,k}\left(t\right)-{C}_{i,j,k}\left(t\right)}{{I}_{i}\left(t\right)}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{B}_{i,j,k}\left(t\right)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{i,j,k}\left(t\right)$$\end{document} represent the benefits and costs of technology \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:j$$\end{document} for agent \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document} in dimension \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{I}_{i}\left(t\right)$$\end{document} is the agent’s income level, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{i,k}$$\end{document} are dimension-specific weights reflecting agent preferences. The Social Network Module implements a dynamic network structure that evolves based on agent interactions, geographical proximity, and principles of homophily. The network is represented as a weighted, directed graph \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:G\left(t\right)=\left(V,E\left(t\right),W\left(t\right)\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:V$$\end{document} is the set of agents, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:E\left(t\right)$$\end{document} is the set of edges at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:t$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:W\left(t\right)$$\end{document} contains the edge weights representing the strength of the relationships. Equation (3) computes the social influence component \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{S}_{i,j}\left(t\right)$$\end{document} as a trust-weighted average of the adoption status of neighboring agents. The social influence component \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{S}_{i,j}\left(t\right)$$\end{document} in Eq. 1 is computed using a trust-weighted influence model:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{S}_{i,j}\left(t\right)=\frac{\sum\nolimits_{k\in\:{N}_{i}\left(t\right)}{w}_{i,k}\left(t\right)\cdot\:{1}_{{A}_{k,j}\left(t\right)}\cdot\:{T}_{i,k}\left(t\right)}{\sum\nolimits_{k\in\:{N}_{i}\left(t\right)}{w}_{i,k}\left(t\right)}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{N}_{i}\left(t\right)$$\end{document} is the set of neighbors of agent \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document} at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:t$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{i,k}\left(t\right)$$\end{document} is the relationship weight, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{1}_{{A}_{k,j}\left(t\right)}$$\end{document} is an indicator function for technology adoption by agent \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{i,k}\left(t\right)$$\end{document} represents the trust level between agents \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k$$\end{document}. Equation (4) describes the Bayesian updating of trust between agents based on a balance of direct experiences and reputation feedback.4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{i,k}\left(t+1\right)=\lambda\:{T}_{i,k}\left(t\right)+\left(1-\lambda\:\right)\left[\mu\:{R}_{i,k}\left(t\right)+\left(1-\mu\:\right){O}_{i,k}\left(t\right)\right]$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}_{i,k}\left(t\right)$$\end{document} represents direct experience with agent \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{O}_{i,k}\left(t\right)$$\end{document} captures opinion from other agents, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mu\:$$\end{document} are learning rate parameters. The outputs from trust evolution directly influence diffusion dynamics by altering connection strengths, ensuring behavioural and network updates remain synchronised. Technology diffusion and utility function The Technology Diffusion Module models the spread of multiple smart agriculture technologies through the agent population. Each technology is characterized by a feature vector that includes cost, complexity, compatibility, relative advantage, and observability attributes. Equation (5) measures the compatibility between an agent’s ideal technology profile and the feature vector of a given technology using a Gaussian similarity function. The technology compatibility component \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{i,j}\left(t\right)$$\end{document} is calculated as:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{i,j}\left(t\right)=\text{e}\text{x}\text{p}\left(-\frac{1}{2}{\left({\mathbf{f}}_{j}-{\mathbf{p}}_{i}\right)}^{T}{\varSigma\:}_{i}^{-1}\left({\mathbf{f}}_{j}-{\mathbf{p}}_{i}\right)\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{f}}_{j}$$\end{document} is the feature vector of technology \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:j$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{p}}_{i}$$\end{document} represents agent \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document}’s ideal technology profile, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varSigma\:}_{i}$$\end{document} is the agent’s preference covariance matrix. Equation (6) models network externalities, where adoption benefits increase proportionally with the share of adopters in an agent’s connected network:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{N}_{i,j}\left(t\right)=\text{t}\text{a}\text{n}\text{h}\left(\xi\:\cdot\:\frac{\sum\nolimits_{k\in\:{N}_{i}\left(t\right)}{1}_{{A}_{k,j}\left(t\right)}}{\left|{N}_{i}\left(t\right)\right|}\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\xi\:$$\end{document} is a scaling parameter that controls the strength of network effects. The diffusion results then flow into the Decision Integration Module, linking micro-level interactions with macro-level adoption outcomes. Decision integration and learning process The Decision Integration Module combines outputs from all other modules to determine agent adoption decisions. The decision process is modelled as a multi-stage process involving awareness, consideration, trial, and adoption phases. The transition probabilities between stages are governed by logistic functions that incorporate the utility calculations from Eq. 1. Equation (7) models the probability that an agent becomes aware of a technology through information diffusion from peers within its social network:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{aware}\left(i,j,t\right)=1-\text{e}\text{x}\text{p}\left(-\rho\:\sum\limits_{k\in\:{N}_{i}\left(t\right)}{w}_{i,k}\left(t\right)\cdot\:{1}_{{A}_{k,j}\left(t\right)}\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\rho\:$$\end{document} is an information diffusion parameter. Equation (8) computes the final adoption probability using a logistic function that links the agent’s utility score to its decision threshold and sensitivity. The adoption decision probability for agent \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document}, considering technology \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:j$$\end{document} at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:t$$\end{document} is given by:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{adopt}\left(i,j,t\right)=\frac{1}{1+\text{e}\text{x}\text{p}\left(-{\theta\:}_{i}\left({U}_{i,j}\left(t\right)-{\tau\:}_{i}\right)\right)}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{i}$$\end{document} represents the agent’s decision sensitivity and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\tau\:}_{i}$$\end{document} is the adoption threshold. The model incorporates several advanced features to enhance realism and predictive accuracy. The heterogeneous agent representation captures the diversity of agricultural systems, featuring specialised agent types for various farming scales and production systems. The dynamic network evolution mechanism allows for the formation and dissolution of relationships based on agent interactions and geographical constraints. The multi-technology environment enables the simulation of complex adoption scenarios where farmers choose among multiple competing or complementary technologies. The temporal technology evolution mechanism models how technologies improve over time, affecting their adoption attractiveness. To ensure computational efficiency for large-scale simulations, the model implements several optimisation strategies. The network update operations use efficient graph algorithms with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left(\left|E\right|\text{l}\text{o}\text{g}\left|V\right|\right)$$\end{document} complexity. The utility calculations are vectorised and computed in parallel for all agents. The simulation state is compressed using advanced data structures that reduce memory usage while maintaining fast access times. The model validation framework incorporates multiple validation approaches, including face validity through expert consultation, sensitivity analysis of key parameters, and empirical validation using historical adoption data. The calibration process utilises genetic algorithms to optimise model parameters against observed adoption patterns in various geographical regions. Fig. 3Agent decision-making process flowchart in AdoptAgriSim. Agent decision-making process flowchart in AdoptAgriSim. The decision-making flowchart presented in Fig. 3 illustrates the comprehensive integration of multiple influence factors in the agent adoption process. The flowchart shows how economic, social, and technological factors converge to inform the adoption decision framework. Agents evaluate these multiple factors before deciding to adopt, delay, or abandon a technology. The feedback loops in the diagram represent the influence of past adoption outcomes on future decisions, emphasizing the dynamic and adaptive nature of farmer behaviour in the simulation. This iterative process captures the learning and adjustment mechanisms that characterize real-world technology adoption in agricultural systems. Mathematical modeling This section presents the core mathematical formulation of the AdoptAgriSim framework, modelling agent behaviour, social influence, technology diffusion, and economic impact. System state representation: The system state at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:t$$\end{document} is represented by:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{S}\left(t\right)=\left(A\left(t\right),G\left(t\right),T\left(t\right),M\left(t\right)\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:A\left(t\right)$$\end{document} is the set of all agent states, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:G\left(t\right)$$\end{document} is the evolving social network, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:T\left(t\right)$$\end{document} denotes technology characteristics, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:M\left(t\right)$$\end{document} captures market-related parameters. Agent state definition: Each agent \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document} has a state vector defined as:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{a}}_{i}\left(t\right)={\left[{x}_{i}\left(t\right),{y}_{i}\left(t\right),{z}_{i}\left(t\right),{\mathbf{h}}_{i}\left(t\right),{\mathbf{n}}_{i}\left(t\right)\right]}^{T}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}\left(t\right)\in\:{\mathbb{R}}^{d}$$\end{document} represents demographic and economic features, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{i}\left(t\right)\in\:\{0,1{\}}^{\left|T\right|}$$\end{document} is a binary technology adoption indicator, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{i}\left(t\right)\in\:{\mathbb{R}}^{+}$$\end{document} is accumulated experience, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{h}}_{i}\left(t\right)\in\:{\mathbb{R}}^{p}$$\end{document} is behavioral history, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{n}}_{i}\left(t\right)\in\:{\mathbb{R}}^{\left|A\right|}$$\end{document} denotes neighborhood influence weights. Connection probability in the network: The probability of a link between agents \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:j$$\end{document} at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:t$$\end{document} is given by:11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{connect}\left(i,j,t\right)=\sigma\:\left({\alpha\:}_{0}+{\alpha\:}_{1}{k}_{i}+{\alpha\:}_{2}{k}_{j}+{\alpha\:}_{3}\mathbf{s}\left({\mathbf{a}}_{i},{\mathbf{a}}_{j}\right)+{\alpha\:}_{4}{d}_{ij}\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:\left(\cdot\:\right)$$\end{document} is the sigmoid function, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{k}_{i}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{k}_{j}$$\end{document} are the degrees (number of connections) of agents \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:j$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{s}\left(\cdot\:\right)$$\end{document} is a similarity function, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{ij}$$\end{document} is geographic distance, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{0}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{4}$$\end{document} are model parameters. Attribute similarity function: Similarity between two agents is modelled using:12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{s}\left({\mathbf{a}}_{i},{\mathbf{a}}_{j}\right)=\text{e}\text{x}\text{p}\left(-\frac{1}{2}{\left({\mathbf{x}}_{i}-{\mathbf{x}}_{j}\right)}^{T}\mathbf{W}\left({\mathbf{x}}_{i}-{\mathbf{x}}_{j}\right)\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{x}}_{i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{x}}_{j}$$\end{document} are the attribute vectors, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{W}\in\:{\mathbb{R}}^{d\times\:d}$$\end{document} is a positive definite weight matrix emphasising attribute relevance. Technology Diffusion Dynamics: Aggregate adoption for technology \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:j$$\end{document} is modelled by:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{d{N}_{j}\left(t\right)}{dt}=\left[{p}_{j}+{q}_{j}\frac{{N}_{j}\left(t\right)}{M}\right]\left[M-{N}_{j}\left(t\right)\right]\cdot\:{\phi\:}_{j}\left(t\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{N}_{j}\left(t\right)$$\end{document} is the number of adopters at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:t$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:M$$\end{document} is the market size, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{j}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{q}_{j}$$\end{document} are innovation and imitation coefficients, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varPhi\:}_{j}\left(t\right)$$\end{document} adjusts for heterogeneity in adoption thresholds. Agent-level adoption decision: The probability of agent \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document} adopting technology \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:j$$\end{document} is:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{i,j}\left(t\right)=\frac{\text{e}\text{x}\text{p}\left(\beta\:{U}_{i,j}\left(t\right)\right)}{\sum\:_{k\in\:{\mathcal{T}}_{i}\left(t\right)}\text{e}\text{x}\text{p}\left(\beta\:{U}_{i,k}\left(t\right)\right)+\text{e}\text{x}\text{p}\left(\beta\:{U}_{i,0}\left(t\right)\right)}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{U}_{i,j}\left(t\right)$$\end{document} is the utility of technology \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:j$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{U}_{i,0}\left(t\right)$$\end{document} is the utility of non-adoption, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:$$\end{document} is a choice sensitivity parameter, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{T}}_{i}\left(t\right)$$\end{document} is the set of options available to agent \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document}. Bayesian belief update: Agents update beliefs based on observed outcomes:15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{i,j}\left(t+1\right)=\frac{{\lambda\:}_{i}{\mu\:}_{i,j}\left(t\right)+\sum\nolimits_{k\in\:{N}_{i}\left(t\right)}{w}_{i,k}\left(t\right){r}_{k,j}\left(t\right)}{{\lambda\:}_{i}+\sum\nolimits_{k\in\:{N}_{i}\left(t\right)}{w}_{i,k}\left(t\right)}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{i,j}\left(t\right)$$\end{document} is the belief about technology \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:j$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{i}$$\end{document} is prior confidence, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{i,k}\left(t\right)$$\end{document} is influence weight from neighbor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{k,j}\left(t\right)$$\end{document} is observed performance. Production function and output: Expected output \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Y}_{i}\left(t\right)$$\end{document} with technology usage is:16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Y}_{i}\left(t\right)={A}_{i}\prod\limits_{j\in\:{\mathcal{T}}_{i}}{\left(1+{\epsilon}_{j}{X}_{i,j}\left(t\right)\right)}^{{\gamma\:}_{j}}{L}_{i}^{\alpha\:}{K}_{i}^{1-\alpha\:}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{A}_{i}$$\end{document} is productivity, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\epsilon}_{j}$$\end{document} is tech efficiency, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}_{i,j}\left(t\right)$$\end{document} is usage level, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\gamma\:}_{j}$$\end{document} is elasticity, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{K}_{i}$$\end{document} are labour and capital, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:$$\end{document} is labour share. Technology cost structure: The total cost for agent \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document} adopting technology \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:j$$\end{document} is:17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{i,j}\left(t\right)={F}_{i,j}+{V}_{i,j}{X}_{i,j}\left(t\right)+{M}_{i,j}\left(t\right)+{L}_{i,j}\left(t\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{i,j}$$\end{document} is fixed cost, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{V}_{i,j}$$\end{document} is per-unit usage cost, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{M}_{i,j}\left(t\right)$$\end{document} is maintenance, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{i,j}\left(t\right)$$\end{document} is learning cost. System transition dynamics: The overall system state evolves as:18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{S}\left(t+1\right)=\mathcal{F}\left(\mathcal{S}\left(t\right),\mathbf{\epsilon}\left(t+1\right)\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{F}\left(\cdot\:\right)$$\end{document} is the transition function and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{\epsilon}\left(t+1\right)$$\end{document} captures stochastic disturbances or shocks. Results and evaluation This section presents comprehensive experimental results demonstrating the effectiveness and accuracy of AdoptAgriSim across multiple evaluation scenarios. Experiments have been conducted using actual agricultural data collected in three different geographical areas, encompassing a wide range of farming systems and socio-economic contexts. The platform of the experiment was a high-performance computing cluster with 64 cores and 512 GB of RAM, which was used to perform scalable simulations of up to 10,000 agents over a 20-year duration. The structure used in the baseline runs featured time steps of one week, providing an opportunity for in-depth temporal analysis of adoption dynamics, while remaining computationally efficient. Parameter operationalization: from raw data to utility weights: To ensure reproducibility, we provide explicit formulas showing how raw dataset variables are transformed into agent-specific utility weight parameters (α, β, γ, δ). Each parameter is computed as a weighted linear combination of normalized input features, with coefficients determined through preliminary regression analysis on historical adoption outcomes. Economic benefit weight (α_i): The economic benefit weight for agent i quantifies the agent’s sensitivity to financial considerations:19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{i}=\:0.42\:\times\:\:Incom{e}_{normalized}+\:0.31\:\times\:\:RO{I}_{history}+\:0.27\:\times\:\:Credi{t}_{access}$$\end{document} where: Income_normalized ∈ [0,1] is the z-score normalized annual farm income.ROI_history ∈ [0,1] represents the average return-on-investment from past technology adoptions (0.5 if no history).Credit_access ∈ [0,1] is a binary indicator (1 if formal credit available, 0.5 if informal, 0 otherwise) normalized by regional access rates. Income_normalized ∈ [0,1] is the z-score normalized annual farm income. Income_normalized ∈ [0,1] is the z-score normalized annual farm income. ROI_history ∈ [0,1] represents the average return-on-investment from past technology adoptions (0.5 if no history). ROI_history ∈ [0,1] represents the average return-on-investment from past technology adoptions (0.5 if no history). Credit_access ∈ [0,1] is a binary indicator (1 if formal credit available, 0.5 if informal, 0 otherwise) normalized by regional access rates. Credit_access ∈ [0,1] is a binary indicator (1 if formal credit available, 0.5 if informal, 0 otherwise) normalized by regional access rates. For the Iowa dataset (IACD), mean α = 0.48 (SD = 0.12), reflecting high economic rationality. For India (IASD), mean α = 0.31 (SD = 0.09), indicating lower income-driven decisions. European farms (FADN) showed α = 0.35 (SD = 0.11). Social influence weight (β_i): The social influence weight captures the agent’s responsiveness to peer effects and trust networks:20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}_{i}=\:0.38\:\times\:\:PTS\:+\:0.35\:\times\:\:CD\:+\:0.27\:\times\:\:IAI$$\end{document} where: PTS (Peer Trust Score) ∈ [0,1] is the average normalized trust rating within the agent’s immediate network.CD (Cooperation Density) ∈ [0,1] is the ratio of cooperative memberships to total social links.IAI (Information Accessibility Index) ∈ [0,1] is a composite of extension service contacts per year (normalized, max = 12). PTS (Peer Trust Score) ∈ [0,1] is the average normalized trust rating within the agent’s immediate network. PTS (Peer Trust Score) ∈ [0,1] is the average normalized trust rating within the agent’s immediate network. CD (Cooperation Density) ∈ [0,1] is the ratio of cooperative memberships to total social links. CD (Cooperation Density) ∈ [0,1] is the ratio of cooperative memberships to total social links. IAI (Information Accessibility Index) ∈ [0,1] is a composite of extension service contacts per year (normalized, max = 12). IAI (Information Accessibility Index) ∈ [0,1] is a composite of extension service contacts per year (normalized, max = 12). Indian smallholders exhibited the highest social influence weight (mean β = 0.42, SD = 0.10), followed by European farms (β = 0.33, SD = 0.09), and Iowa farms (β = 0.21, SD = 0.08). Technology compatibility weight (γ_i): The technology compatibility weight reflects the agent’s emphasis on ease-of-use and integration with existing practices:21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\gamma\:}_{i}=\:0.45\:\times\:\:TFI\:+\:0.32\:\times\:\:Digita{l}_{literacy}+\:0.23\:\times\:\:Far{m}_{complexity}^{-1}$$\end{document} where: TFI (Technology Familiarity Index) ∈ [0,1] is a weighted score of prior exposure to digital/mechanized tools (0.2 per adopted technology, capped at 1.0).Digital_literacy ∈ [0,1] is derived from education level: (0.2 = no formal education, 0.4 = primary, 0.6 = secondary, 0.8 = technical diploma, 1.0 = university)Farm_complexity^(−1) ∈ [0,1] is the inverse normalized complexity score (simpler farms = higher compatibility focus). TFI (Technology Familiarity Index) ∈ [0,1] is a weighted score of prior exposure to digital/mechanized tools (0.2 per adopted technology, capped at 1.0). TFI (Technology Familiarity Index) ∈ [0,1] is a weighted score of prior exposure to digital/mechanized tools (0.2 per adopted technology, capped at 1.0). Digital_literacy ∈ [0,1] is derived from education level: (0.2 = no formal education, 0.4 = primary, 0.6 = secondary, 0.8 = technical diploma, 1.0 = university) Digital_literacy ∈ [0,1] is derived from education level: (0.2 = no formal education, 0.4 = primary, 0.6 = secondary, 0.8 = technical diploma, 1.0 = university) Farm_complexity^(−1) ∈ [0,1] is the inverse normalized complexity score (simpler farms = higher compatibility focus). Farm_complexity^(−1) ∈ [0,1] is the inverse normalized complexity score (simpler farms = higher compatibility focus). Across datasets, γ showed moderate variation: IACD (mean = 0.27, SD = 0.07), FADN (mean = 0.25, SD = 0.06), IASD (mean = 0.19, SD = 0.08), reflecting that compatibility concerns are secondary but consistent. Network externality weight (δ_i): The network externality weight quantifies the agent’s sensitivity to peer adoption prevalence: δ_i = 0.40 × Network_degree_normalized + 0.35 × Clustering_coefficient + 0.25 × Market_integration. where: Network_degree_normalized ∈ [0,1] is the agent’s number of social connections divided by the regional 95th percentile.Clustering_coefficient ∈ [0,1] measures the density of connections among the agent’s neighbors.Market_integration ∈ [0,1] represents participation in value chains (0 = subsistence, 0.5 = local markets, 1.0 = export-oriented). Network_degree_normalized ∈ [0,1] is the agent’s number of social connections divided by the regional 95th percentile. Network_degree_normalized ∈ [0,1] is the agent’s number of social connections divided by the regional 95th percentile. Clustering_coefficient ∈ [0,1] measures the density of connections among the agent’s neighbors. Clustering_coefficient ∈ [0,1] measures the density of connections among the agent’s neighbors. Market_integration ∈ [0,1] represents participation in value chains (0 = subsistence, 0.5 = local markets, 1.0 = export-oriented). Market_integration ∈ [0,1] represents participation in value chains (0 = subsistence, 0.5 = local markets, 1.0 = export-oriented). Network externality weights were lowest in Iowa (mean δ = 0.18, SD = 0.05), moderate in Europe (δ = 0.24, SD = 0.07), and highest in India (δ = 0.28, SD = 0.09), aligning with differences in social cohesion and market structures. Validation and constraints: All parameters are constrained to sum to unity for each agent: α_i + β_i + γ_i + δ_i = 1.0, enforced through post-calculation normalization:22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{i}^{{\prime\:}}=\frac{{\alpha\:}_{i}}{\left({\alpha\:}_{i}+\:{\beta\:}_{i}+\:{\gamma\:}_{i}+\:{\delta\:}_{i}\right)}$$\end{document} with identical normalization for β’, γ’, and δ’. This ensures interpretability as relative importance weights while preserving agent-specific heterogeneity. The operationalization formulas were validated by computing parameters for a held-out sample (20% of each dataset) and verifying that parameter distributions matched those of the training set (Kolmogorov-Smirnov test, p > 0.05 for all parameters across all datasets). Example calculation: For an Iowa farmer with Income_normalized = 0.75, ROI_history = 0.60, Credit_access = 1.0, PTS = 0.40, CD = 0.30, IAI = 0.50, TFI = 0.70, Digital_literacy = 0.80, Farm_complexity^(−1) = 0.65, Network_degree_normalized = 0.55, Clustering_coefficient = 0.45, Market_integration = 0.85:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{i}=\:0.42\left(0.75\right)+\:0.31\left(0.60\right)+\:0.27\left(1.0\right)=\:0.771\:$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}_{i}=\:0.38\left(0.40\right)+\:0.35\left(0.30\right)+\:0.27\left(0.50\right)=\:0.392\:$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\gamma\:}_{i}=\:0.45\left(0.70\right)+\:0.32\left(0.80\right)+\:0.23\left(0.65\right)=\:0.621\:$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\delta\:}_{i}=\:0.40\left(0.55\right)+\:0.35\left(0.45\right)+\:0.25\left(0.85\right)=\:0.590\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{i}=\:0.42\left(0.75\right)+\:0.31\left(0.60\right)+\:0.27\left(1.0\right)=\:0.771\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{i}=\:0.42\left(0.75\right)+\:0.31\left(0.60\right)+\:0.27\left(1.0\right)=\:0.771\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}_{i}=\:0.38\left(0.40\right)+\:0.35\left(0.30\right)+\:0.27\left(0.50\right)=\:0.392\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}_{i}=\:0.38\left(0.40\right)+\:0.35\left(0.30\right)+\:0.27\left(0.50\right)=\:0.392\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\gamma\:}_{i}=\:0.45\left(0.70\right)+\:0.32\left(0.80\right)+\:0.23\left(0.65\right)=\:0.621\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\gamma\:}_{i}=\:0.45\left(0.70\right)+\:0.32\left(0.80\right)+\:0.23\left(0.65\right)=\:0.621\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\delta\:}_{i}=\:0.40\left(0.55\right)+\:0.35\left(0.45\right)+\:0.25\left(0.85\right)=\:0.590\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\delta\:}_{i}=\:0.40\left(0.55\right)+\:0.35\left(0.45\right)+\:0.25\left(0.85\right)=\:0.590\:$$\end{document} Sum = 2.374; Normalized: α’ = 0.325, β’ = 0.165, γ’ = 0.262, δ’ = 0.248. This agent exhibits high economic orientation (α’ = 0.325), moderate compatibility focus (γ’ = 0.262), and lower social influence sensitivity (β’ = 0.165), consistent with commercial Iowa farm profiles. Dataset description and preprocessing The assessment utilized three major datasets: The Iowa Agricultural Census Dataset (IACD), which comprises 2,847 farm operations; the EU Farm Accountancy Data Network (FADN), containing 4,523 European farms; and the Indian Agricultural Statistics Dataset (IASD), comprising 3,198 smallholder farms. All datasets were selected for their regional diversity and representativeness of different socio-economic and technological adoption contexts. Data sources: The IACD was obtained from the United States Department of Agriculture (USDA) open agricultural census database (2018–2023). The FADN dataset was retrieved from the European Commission’s Farm Accountancy Data Network (2019–2024), while the IASD was compiled from the Ministry of Agriculture and Farmers Welfare, Government of India, through its open agricultural statistics portal (2020–2025). Each dataset includes demographic, economic, and behavioural indicators essential for modelling agent heterogeneity, social interactions, and technology diffusion. Preprocessing and feature engineering: Each dataset underwent a standardised data-processing pipeline to ensure consistency and reduce sampling bias. Continuous attributes such as farm size, income, credit access, and trust index were z-score normalised, while categorical features, including education level, farm type, and region, were one-hot encoded. Outliers beyond ± 3 standard deviations were removed, and missing values (< 2.5%) were imputed using median substitution to preserve statistical distribution. Feature engineering introduced several derived variables directly linked to the model’s decision parameters: Cooperation Density (CD): computed from the ratio of cooperative memberships to total social links;Technology Familiarity Index (TFI): weighted score of prior exposure to digital or mechanised tools;Peer Trust Score (PTS): average of normalised trust weights within each farmer’s ego network;Information Accessibility Index (IAI): proxy for extension service reach and communication frequency. Cooperation Density (CD): computed from the ratio of cooperative memberships to total social links; Cooperation Density (CD): computed from the ratio of cooperative memberships to total social links; Technology Familiarity Index (TFI): weighted score of prior exposure to digital or mechanised tools; Technology Familiarity Index (TFI): weighted score of prior exposure to digital or mechanised tools; Peer Trust Score (PTS): average of normalised trust weights within each farmer’s ego network; Peer Trust Score (PTS): average of normalised trust weights within each farmer’s ego network; Information Accessibility Index (IAI): proxy for extension service reach and communication frequency. Information Accessibility Index (IAI): proxy for extension service reach and communication frequency. A correlation-based feature-selection process (threshold < 0.6) was applied to eliminate redundant attributes and maintain orthogonality between predictor variables. Final feature sets were aligned with the four principal parameters of the model: economic benefit\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\alpha\:\right)$$\end{document}, social influence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\beta\:\right)$$\end{document}, technology compatibility \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\gamma\:\right)$$\end{document}, and network externality \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\delta\:\right)$$\end{document}. Table 1Summary of datasets and agent characteristics.DatasetAgentsPeriodTechnologiesNetworksFarm Size (ha)Farmer AgeEducation (P/S/U)Trust ScoreIACD2,8472018–20238Geographic + Supply123.452.112%/61%/27%0.62FADN4,5232019–202412Cooperative + Market67.849.58%/49%/43%0.74IASD3,1982020–20256Social + Extension4.145.828%/66%/6%0.89Combined10,568 6 years 15 unique Multi-layer ---- Table 1Summary of datasets and agent characteristics.DatasetAgentsPeriodTechnologiesNetworksFarm Size (ha)Farmer AgeEducation (P/S/U)Trust ScoreIACD2,8472018–20238Geographic + Supply123.452.112%/61%/27%0.62FADN4,5232019–202412Cooperative + Market67.849.58%/49%/43%0.74IASD3,1982020–20256Social + Extension4.145.828%/66%/6%0.89Combined10,568 6 years 15 unique Multi-layer ---- Summary of datasets and agent characteristics. As shown in Table 1, the evaluation encompassed heterogeneous regional and socio-economic settings. Indian smallholders (IASD) exhibited the highest levels of trust and cooperation, promoting stronger peer-to-peer influence cascades. Large commercial farms in Iowa (IACD) emphasised economic factors such as ROI and mechanisation readiness. In contrast, European farms (FADN) exhibited balanced characteristics, characterised by moderate social cohesion and mid-complexity technology adoption. This harmonised preprocessing framework established a coherent input basis for subsequent agent modelling, ensuring comparability across datasets and reinforcing the reliability of simulation outcomes. The Iowa Agricultural Census Dataset (IACD) is based on data from the U.S. Department of Agriculture’s National Agricultural Statistics Service (NASS). The NASS Quick Stats portal provides state and county-level agricultural variables for Iowa and is openly accessible. https://www.nass.usda.gov/Data_and_Statistics/?utm_source=chatgpt.com. The Iowa Census of Agriculture (e.g. 2022 edition) also publishes detailed farm and demographic tables relevant for modelling. https://www.nass.usda.gov/Publications/AgCensus/2022/Full_Report/Volume_1%2C_Chapter_2_County_Level/Iowa/?utm_source=chatgpt.com. For India/IASD, data were sourced from the Government of India’s official agricultural statistics platforms, including the Unified Portal for Agricultural Statistics (UPAg) and the “Agricultural Statistics at a Glance” reports published by the Ministry of Agriculture & Farmers Welfare. https://upag.gov.in/?utm_source=chatgpt.com. The UPAg portal consolidates multi-year crop, area, production, and yield data across states https://upag.gov.in/all-dashboards? utm_source=chatgpt.com. In our implementation, the IASD dataset was constructed by aligning annual state-level statistics (from UPAg and Agricultural Statistics at a Glance) with survey-level agricultural socioeconomic indicators (e.g. literacy, extension outreach) from secondary sources. Regions were stratified by state and socioeconomic groups, then randomly sampled to populate 3,198 smallholder agent profiles, preserving real-world distributions. The EU Farm Accountancy Data Network (FADN) is a publicly documented agricultural accountancy survey managed by the European Commission to assess farm income and the effects of the Common Agricultural Policy. https://agridata.ec.europa.eu/extensions/FarmEconomyFocus/FADNDatabase.html?utm_source=chatgpt.com. While FADN publishes aggregate statistics via its public database, detailed farm-level data are often accessible only through a formal data request or institutional access. https://journals.francoangeli.it/index.php/ecagoa/article/view/12770?utm_source=chatgpt.com. In our study, we obtained disaggregated FADN data via a research agreement with DG-Agri, under a license for non-commercial academic use. Only anonymised farm-level records were retained for analysis, in compliance with EU data protection regulations. Social network construction and trust quantification The network structures in AdoptAgriSim were operationalised using data-driven methods derived from the attributes available within each regional dataset. For the Iowa Agricultural Census Dataset (IACD), geographic and supply networks were constructed by linking agents located within a 50 km geographic radius who also shared at least one common supply-chain intermediary (e.g., grain elevator, input distributor). Network edges were generated using a weighted adjacency matrix, where weights were calculated as a composite of spatial proximity (inverse of Euclidean distance) and supplier overlap ratio. This ensured that both physical closeness and logistical relationships contributed to the strength of the connection. For the EU Farm Accountancy Data Network (FADN), cooperative and market networks were developed by mapping cooperative membership and shared market channel participation. Edge weights reflected the frequency of joint cooperative participation and transaction volume similarity, capturing the intensity of economic interactions between farms. The Indian Agricultural Statistics Dataset (IASD) networks were formed based on village-level group affiliations and agricultural extension participation records, resulting in Social and Extension networks. Here, edge weights were determined based on the number of shared group memberships and training event co-attendance, emphasising peer learning and local social cohesion. The Trust Score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:({T}_{i,k}$$\end{document}) between any two agents, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k$$\end{document}, was calculated as a composite indicator that incorporates both direct interaction frequency and perceived reputation. It was defined as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{\left(i,k\right)}=0.6\times\:\frac{{f}_{i,k}}{max\left(f\right)}+0.4\times\:\frac{{r}_{i,k}}{max\left(r\right)}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{i,k}$$\end{document} ​represents normalised interaction frequency derived from observed co-participation in transactions, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{i,k}$$\end{document} represents peer reputation based on cooperative reliability or social endorsement metrics. The resulting trust values were normalised between 0 and 1 and served as the initial prior values for the Bayesian trust-update mechanism described in Eq. (4). This approach provided a reproducible and transparent basis for modelling dynamic trust evolution and its influence on social diffusion within agricultural networks. Model calibration and parameter optimisation Fig. 4Model calibration convergence showing the evolution of fitness scores across genetic algorithm generations for three regional datasets. The x-axis represents the number of generations during optimisation, while the y-axis shows the fitness score, which measures model accuracy. The curves illustrate consistent convergence trends across all datasets, with higher \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:$$\end{document} values (social influence weight) leading to faster stabilisation in the Indian Agricultural Stats dataset compared to others, indicating stronger peer-effect sensitivity in that region. Model calibration convergence showing the evolution of fitness scores across genetic algorithm generations for three regional datasets. The x-axis represents the number of generations during optimisation, while the y-axis shows the fitness score, which measures model accuracy. The curves illustrate consistent convergence trends across all datasets, with higher \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:$$\end{document} values (social influence weight) leading to faster stabilisation in the Indian Agricultural Stats dataset compared to others, indicating stronger peer-effect sensitivity in that region. The calibration results shown in Fig. 4 demonstrate rapid convergence of the optimisation algorithm across all three regional datasets. The optimised parameter values revealed significant regional variations in social influence weights, with European farmers showing higher sensitivity to peer adoption \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:(\beta\:\:=\:0.74)$$\end{document} compared to Iowa farmers \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:(\beta\:\:=\:0.52)$$\end{document} and Indian farmers \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:(\beta\:\:=\:0.89)$$\end{document}. Economic factors were more dominant in Iowa \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:(\alpha\:\:=\:0.68)$$\end{document}, while social factors were more influential in India \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:(\beta\:\:=\:0.89)$$\end{document}. Data preparation and input selection: Prior to simulation, all datasets underwent a structured data preparation pipeline to ensure consistency and minimise bias across heterogeneous regional records. Input variables were categorised into four dimensions—economic, demographic, social, and technological to align with the model’s utility parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:(\alpha\:,\:\beta\:,\:\gamma\:,\:\delta\:)$$\end{document}. Economic features included indicators such as annual farm income, input cost ratio, and return-on-investment metrics. Demographic attributes included farmer age, education, and farm size, while social features encompassed trust level, network degree, and frequency of cooperative participation. Technological variables included digital literacy, prior technology use, and access to extension. All numeric features were normalised between 0 and 1 to prevent scale dominance, and categorical variables (e.g., education, farm type) were encoded using one-hot encoding. Missing values (< 2.5%) were imputed using region-specific medians. Feature selection was performed using variance inflation factor (VIF) analysis and pairwise correlation filtering, ensuring multicollinearity remained below 0.6. Variables contributing less than 0.15 to the variance of any core utility component were excluded to enhance interpretability and stability of the results. This procedure establishes a direct link between input data and the model’s mathematical structure, where α represents the economic benefit weight, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:$$\end{document} captures social influence, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\gamma\:$$\end{document} reflects technology compatibility, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\delta\:$$\end{document} denotes the strength of network externality. Optimisation algorithm details: The parameter calibration process utilised a genetic algorithm (GA) to optimise model parameters by minimising the root-mean-square error (RMSE) between simulated and observed adoption trajectories. The GA operated with a population size of 80, a crossover probability of 0.6, a mutation rate of 0.1, and a maximum of 200 generations. The fitness function evaluated parameter sets based on the accuracy of predicted adoption rates and the stability of convergence. Early stopping was triggered when the RMSE improvement fell below 1 × 10⁻⁴ over 20 consecutive generations. The GA’s stochastic search mechanism enabled efficient exploration of the high-dimensional parameter space while avoiding local minima. The final optimised weights were validated through ten-fold cross-validation across the three regional datasets to confirm their robustness and generalizability. The GA was implemented in Python 3.11 using the DEAP evolutionary computation library, with parallel execution enabled on 16 threads for efficiency. Table 2Optimised parameter sets for regional calibration.Dataset\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{\alpha\:}$$\end{document} (Economic)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{\beta\:}$$\end{document} (Social)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{\gamma\:}$$\end{document} (Technology)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{\delta\:}$$\end{document} (Network) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{\xi\:}\:(\text{Scaling})$$\end{document} IACD (Iowa)0.420.270.180.131.45FADN (Europe)0.350.330.210.111.32IASD (India)0.280.410.170.141.51 Table 2Optimised parameter sets for regional calibration.Dataset\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{\alpha\:}$$\end{document} (Economic)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{\beta\:}$$\end{document} (Social)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{\gamma\:}$$\end{document} (Technology)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{\delta\:}$$\end{document} (Network) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{\xi\:}\:(\text{Scaling})$$\end{document} IACD (Iowa)0.420.270.180.131.45FADN (Europe)0.350.330.210.111.32IASD (India)0.280.410.170.141.51 Optimised parameter sets for regional calibration. The optimised parameter sets presented in Table 2 quantify the relative importance of economic, social, technological, and network factors in determining adoption behaviour across the three regional contexts. The calibration results reveal that economic benefits \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\varvec{\alpha\:}\right)$$\end{document} exert the most significant influence on adoption among Iowa farmers, reflecting a predominantly profit-driven decision pattern consistent with large-scale commercial operations. In contrast, social influence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\varvec{\beta\:}\right)$$\end{document} exhibits the highest weight in the Indian context, aligning with strong community-based and peer-driven decision mechanisms. The European dataset (FADN) demonstrates a balanced configuration, where moderate values of α = 0.35 and β = 0.33 indicate that both financial and cooperative factors contribute comparably to adoption dynamics. The technology compatibility \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\varvec{\gamma\:}\right)$$\end{document} and network externality \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\varvec{\delta\:}\right)$$\end{document} parameters maintain smaller yet significant weights across all regions, emphasising that ease of integration and peer presence play secondary but consistent roles. The scaling coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\varvec{\xi\:}\right)$$\end{document}, governing the strength of network effects in Eq. (6), varies between 1.32 and 1.51, confirming that social connectivity enhances diffusion speed but with region-specific saturation limits. Collectively, these optimised weights demonstrate how socio-economic heterogeneity and institutional contexts shape agent behaviour, validating the adaptability and robustness of the AdoptAgriSim framework for multi-regional agricultural systems. To enable application of AdoptAgriSim in new contexts, weight parameters (α, β, γ, δ) can be determined through three approaches1: Data-driven calibration - when historical adoption data are available (minimum 3–5 years, 500 + farms), apply the genetic algorithm optimization procedure described above to minimize prediction error2; Survey-based elicitation - conduct discrete choice experiments with farmers (50–100 respondents) to estimate implicit weights through stated preference regression models; or3 Transfer learning - initialize parameters from similar agricultural contexts (e.g., use Iowa’s high α for industrialized regions, India’s high β for community-based systems), then refine through sensitivity analysis and expert validation. A hybrid approach combining multiple methods provides the most robust calibration. Researchers should validate final weights through ± 20% sensitivity analysis to ensure model robustness. The parameter calibration process utilized a genetic algorithm (GA) to optimize model parameters by minimizing prediction error between simulated and observed adoption trajectories. The GA operated with a population size of 80, a crossover probability of 0.6, a mutation rate of 0.1, and a maximum of 200 generations. Complete fitness function: The fitness function for evaluating each parameter set θ = {α, β, γ, δ, λ, θ_trust, θ_network} was defined as:23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Fitness\left(\theta\:\right)=\:-\left[RMSE\left(\theta\:\right)+\:{\lambda\:}_{penalty}\times\:\:{P}_{constraint\left(\theta\:\right)}+\:{\lambda\:}_{stability}\times\:\:S\left(\theta\:\right)\right]$$\end{document} where the primary component is the root-mean-square error:24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:RMSE\left(\theta\:\right)=\:\sqrt{\frac{1}{T}\sum\limits_{\left\{t=1\right\}}^{T}{\left(Adoptio{n}_{sim\left(t,\:\theta\:\right)}-\:Adoptio{n}_{obs\left(t\right)}\right)}^{2}}$$\end{document} with T = 5 years (annual time steps), Adoption_sim(t, θ) representing the simulated adoption rate at year t under parameters θ, and Adoption_obs(t) being the empirically observed adoption rate. The constraint penalty term ensures parameter validity:25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{constraint\left(\theta\:\right)}=\:\sum\limits_{\left\{p\:\in\:\:\theta\:\right\}max\left(0,\:p\:-\:1.0\right)}+\:max\left(0,\:-p\right)+\:\left|\sum\:\left(\alpha\:,\:\beta\:,\:\gamma\:,\:\delta\:\right)-\:1.0\right|$$\end{document} which penalizes parameters outside [0,1] bounds and enforces that utility weights sum to unity (penalty weight λ_penalty = 100). The stability term prevents overfitting to specific years:26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:S\left(\theta\:\right)=\:\left(\frac{1}{T}\right)\sum\limits_{\left\{t=1\right\}}^{T}\left|Adoptio{n}_{sim\left(t,\:\theta\:\right)}-\:Adoptio{n}_{sim\left(t-1,\:\theta\:\right)}-\:\varDelta\:Adoptio{n}_{obs\left(t\right)}\right|$$\end{document} which measures deviation from observed year-to-year adoption rate changes, weighted by λ_stability = 0.15. This ensures the model captures not just aggregate accuracy but also temporal dynamics. Early stopping was triggered when the fitness improvement fell below 1 × 10⁻⁴ over 20 consecutive generations. The GA’s stochastic search mechanism enabled efficient exploration of the high-dimensional parameter space while avoiding local minima. The final optimized weights were validated through ten-fold cross-validation across the three regional datasets to confirm their robustness and generalizability. The GA was implemented in Python 3.11 using the DEAP evolutionary computation library, with parallel execution enabled on 16 threads for efficiency. Simulation factors and experimental settings The simulation setup of AdoptAgriSim integrates multiple factor dimensions that jointly shape agent decisions and collective diffusion outcomes. These factors are classified into four main categories: economic, demographic, social, and technological. Each category reflects a distinct aspect of the socio-technical environment influencing technology adoption. Economic factors include annual farm income, input cost ratio, and return on investment (ROI), representing rational decision-making drivers. Demographic factors capture heterogeneity in age, education, and farm size, shaping agents’ learning capacity and risk perception. Social factors encompass network degree, clustering coefficient, trust score, and cooperative participation rate, which together define the influence topology and strength of peer interactions. Technological factors influence innovation characteristics, including cost, compatibility, observability, and complexity. In addition to these, environmental and policy parameters—including information accessibility, subsidy rate, and peer influence threshold—were incorporated to emulate realistic agricultural adoption environments. Each simulation spanned five years, with annual updates, and maintained dataset-specific population sizes: IACD (2,847 agents), FADN (4,523 agents), and IASD (3,198 agents). The sensitivity analysis revealed that moderate (± 20%) variations in α, β, γ, and δ affected convergence rate more than overall accuracy, confirming the robustness of the simulation. The factor design, shown in Table 3, provides transparency and supports reproducibility for future comparative modelling efforts. Table 3Summary of simulation factors and parameters.Factor categoryVariables/parametersDescription/role in simulationRange/source Economic Farm Income, Input Cost Ratio, ROIDetermines economic rationality and cost-benefit evaluationDataset values (normalised 0–1) Demographic Age, Education Level, Farm SizeControls heterogeneity and learning ability of agentsDataset-based attributes Social Trust Score, Network Degree, Clustering Coefficient, Cooperative ParticipationDefines peer influence, information flow, and diffusion cascadesDerived from network topology (Eqs. 3–4) Technological Cost, Compatibility, Observability, ComplexityRepresents innovation features influencing adoption utilityDataset-specific, normalised 0–1 Policy/Environmental Information Accessibility, Subsidy Rate, Peer Influence ThresholdModels external support and information asymmetryScenario-based parameters Model Parameters α, β, γ, δWeights for economic, social, technological, and network effectsOptimised via GA (Table 2) Table 3Summary of simulation factors and parameters.Factor categoryVariables/parametersDescription/role in simulationRange/source Economic Farm Income, Input Cost Ratio, ROIDetermines economic rationality and cost-benefit evaluationDataset values (normalised 0–1) Demographic Age, Education Level, Farm SizeControls heterogeneity and learning ability of agentsDataset-based attributes Social Trust Score, Network Degree, Clustering Coefficient, Cooperative ParticipationDefines peer influence, information flow, and diffusion cascadesDerived from network topology (Eqs. 3–4) Technological Cost, Compatibility, Observability, ComplexityRepresents innovation features influencing adoption utilityDataset-specific, normalised 0–1 Policy/Environmental Information Accessibility, Subsidy Rate, Peer Influence ThresholdModels external support and information asymmetryScenario-based parameters Model Parameters α, β, γ, δWeights for economic, social, technological, and network effectsOptimised via GA (Table 2) Summary of simulation factors and parameters. Iowa (IACD) Networks: Edges connect farms within 50 km that share supply-chain intermediaries (grain elevators, input distributors, cooperatives). Connection strength combines spatial proximity (60% weight, declining exponentially with distance) and supplier overlap ratio (40% weight). This produced 2,847 nodes with average 8.3 connections per farm and clustering coefficient 0.34.European (FADN) Networks: Edges link farms sharing cooperative memberships or similar market channels. Connection strength weighs cooperative overlap (55%) and transaction volume similarity (45%), capturing institutional and economic ties. This produced 4,523 nodes with average 11.7 connections and clustering coefficient 0.42, indicating stronger cooperative cohesion.Indian (IASD) Networks: Edges connect farmers in the same village who share farmer group memberships or co-attend extension training events. Connection strength combines group affiliations (50%), training co-attendance (30%), and kinship ties (20%), reflecting community-based social structures. This produced 3,198 nodes with average 14.2 connections and clustering coefficient 0.58, the highest among all datasets.Trust Initialization: Initial trust between connected farmers combines direct interaction frequency (65% weight) and peer reputation based on cooperative reliability (35% weight). These values initialize the dynamic Bayesian trust-update mechanism. Iowa (IACD) Networks: Edges connect farms within 50 km that share supply-chain intermediaries (grain elevators, input distributors, cooperatives). Connection strength combines spatial proximity (60% weight, declining exponentially with distance) and supplier overlap ratio (40% weight). This produced 2,847 nodes with average 8.3 connections per farm and clustering coefficient 0.34. Iowa (IACD) Networks: Edges connect farms within 50 km that share supply-chain intermediaries (grain elevators, input distributors, cooperatives). Connection strength combines spatial proximity (60% weight, declining exponentially with distance) and supplier overlap ratio (40% weight). This produced 2,847 nodes with average 8.3 connections per farm and clustering coefficient 0.34. European (FADN) Networks: Edges link farms sharing cooperative memberships or similar market channels. Connection strength weighs cooperative overlap (55%) and transaction volume similarity (45%), capturing institutional and economic ties. This produced 4,523 nodes with average 11.7 connections and clustering coefficient 0.42, indicating stronger cooperative cohesion. European (FADN) Networks: Edges link farms sharing cooperative memberships or similar market channels. Connection strength weighs cooperative overlap (55%) and transaction volume similarity (45%), capturing institutional and economic ties. This produced 4,523 nodes with average 11.7 connections and clustering coefficient 0.42, indicating stronger cooperative cohesion. Indian (IASD) Networks: Edges connect farmers in the same village who share farmer group memberships or co-attend extension training events. Connection strength combines group affiliations (50%), training co-attendance (30%), and kinship ties (20%), reflecting community-based social structures. This produced 3,198 nodes with average 14.2 connections and clustering coefficient 0.58, the highest among all datasets. Indian (IASD) Networks: Edges connect farmers in the same village who share farmer group memberships or co-attend extension training events. Connection strength combines group affiliations (50%), training co-attendance (30%), and kinship ties (20%), reflecting community-based social structures. This produced 3,198 nodes with average 14.2 connections and clustering coefficient 0.58, the highest among all datasets. Trust Initialization: Initial trust between connected farmers combines direct interaction frequency (65% weight) and peer reputation based on cooperative reliability (35% weight). These values initialize the dynamic Bayesian trust-update mechanism. Trust Initialization: Initial trust between connected farmers combines direct interaction frequency (65% weight) and peer reputation based on cooperative reliability (35% weight). These values initialize the dynamic Bayesian trust-update mechanism. Evaluation metrics To ensure consistency, transparency, and interpretability across experiments, multiple evaluation metrics were employed to assess the predictive accuracy, stability, and robustness of the proposed AdoptAgriSim model. These metrics capture both the quantitative performance of adoption prediction and the qualitative behavioural validity of agent interactions. The key quantitative indicators include Accuracy (ACC), Precision (P), Recall (R), F1-Score (F1), and Area Under the Curve (AUC), which are formally defined as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Accuracy=\frac{TP+TN}{TP+TN+FP+FN},Precision=\frac{TP}{TP+FP},Recall=\frac{TP}{TP+FN}\:F1=2\times\:\frac{Precision\times\:Recall}{Precision+Recall}$$\end{document} where TP, TN, FP, and FN represent the counts of true positives, true negatives, false positives, and false negatives, respectively. Accuracy measures the overall correctness of predictions, Precision quantifies the reliability of positive classifications, Recall assesses the model’s ability to detect true adopters, and F1 provides a harmonic balance between Precision and Recall. The AUC metric was calculated from the Receiver Operating Characteristic (ROC) curve to assess the discriminative capability across various decision thresholds. A higher AUC indicates superior separation between adopter and non-adopter classes. To assess behavioural and structural reliability, two additional indicators were employed: Robustness Score (RS): quantifies the consistency of adoption predictions under stochastic perturbations in social network topology and agent preferences (± 20% parameter variation).Sensitivity Rank (SR): evaluates the relative influence of each key model parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:(\alpha\:,\:\beta\:,\:\gamma\:,\:\delta\:)$$\end{document} on the overall accuracy, derived from partial-derivative analysis of model responses. Robustness Score (RS): quantifies the consistency of adoption predictions under stochastic perturbations in social network topology and agent preferences (± 20% parameter variation). Robustness Score (RS): quantifies the consistency of adoption predictions under stochastic perturbations in social network topology and agent preferences (± 20% parameter variation). Sensitivity Rank (SR): evaluates the relative influence of each key model parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:(\alpha\:,\:\beta\:,\:\gamma\:,\:\delta\:)$$\end{document} on the overall accuracy, derived from partial-derivative analysis of model responses. Sensitivity Rank (SR): evaluates the relative influence of each key model parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:(\alpha\:,\:\beta\:,\:\gamma\:,\:\delta\:)$$\end{document} on the overall accuracy, derived from partial-derivative analysis of model responses. Together, these seven metrics provide a comprehensive evaluation framework that jointly captures predictive precision, behavioral realism, and resilience of the proposed simulation model. Adoption prediction accuracy Fig. 5Adoption prediction accuracy comparison across different models and time horizons, showing the superior performance of AdoptAgriSim for long-term predictions. The x-axis represents the prediction time horizon ranging from 1 to 5 years, while the y-axis shows prediction accuracy in percentage. Results demonstrate that AdoptAgriSim consistently outperforms baseline models such as Random Forest, SVM, and Bass Diffusion, achieving a 94.2% accuracy for long-term forecasts (5-year horizon), which is 6.9% higher than the nearest competing model. Adoption prediction accuracy comparison across different models and time horizons, showing the superior performance of AdoptAgriSim for long-term predictions. The x-axis represents the prediction time horizon ranging from 1 to 5 years, while the y-axis shows prediction accuracy in percentage. Results demonstrate that AdoptAgriSim consistently outperforms baseline models such as Random Forest, SVM, and Bass Diffusion, achieving a 94.2% accuracy for long-term forecasts (5-year horizon), which is 6.9% higher than the nearest competing model. The comparative analysis presented in Fig. 5 clearly demonstrates the superior performance of AdoptAgriSim across all evaluation timeframes. AdoptAgriSim achieved 94.2% accuracy for 5-year adoption predictions, significantly outperforming the best baseline model (Random Forest at 87.3%). The performance advantage was particularly pronounced for long-term predictions (3–5 years), where the socio-technical modelling approach demonstrated substantial benefits over traditional econometric models. Table 4 summarises prediction accuracy metrics across technologies and regions, along with temporal adoption insights. The added columns show the year when adoption peaked and the final saturation level, highlighting both performance and diffusion behaviour for each smart agriculture technology. Table 4Detailed accuracy results across technology categories and regions.TechnologyRegionAccuracyPrecisionRecallF1-ScoreAUCPrecision AgricultureIowa0.9560.9410.9630.9520.987IoT SensorsEurope0.9230.9180.9310.9240.972Drone TechnologyIndia0.8870.8740.9030.8880.948Autonomous MachineryIowa0.9760.9690.9820.9750.994Smart IrrigationEurope0.9340.9280.9410.9340.979Mobile AppsIndia0.9120.8970.9280.9120.961Overall All 0.942 0.936 0.948 0.942 0.978 Table 4Detailed accuracy results across technology categories and regions.TechnologyRegionAccuracyPrecisionRecallF1-ScoreAUCPrecision AgricultureIowa0.9560.9410.9630.9520.987IoT SensorsEurope0.9230.9180.9310.9240.972Drone TechnologyIndia0.8870.8740.9030.8880.948Autonomous MachineryIowa0.9760.9690.9820.9750.994Smart IrrigationEurope0.9340.9280.9410.9340.979Mobile AppsIndia0.9120.8970.9280.9120.961Overall All 0.942 0.936 0.948 0.942 0.978 Detailed accuracy results across technology categories and regions. The results not only confirm the predictive strength of AdoptAgriSim but also reveal critical socio-technical insights. The higher sensitivity of adoption to the social influence weight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\beta\:\right)$$\end{document} demonstrates that collective trust and information sharing play dominant roles in technology diffusion, particularly in the Indian and European contexts. Conversely, stronger economic coefficients \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\alpha\:\right)$$\end{document} in the Iowa dataset highlight that profitability and mechanisation incentives remain decisive for large-scale farms. The observed alignment between simulated and real-world adoption rates indicates that the model successfully captures both micro-level decision heterogeneity and macro-level network effects. These findings substantiate the importance of embedding socio-technical feedback loops into future policy interventions for more effective technology dissemination strategies. Temporal dynamics analysis The temporal analysis revealed distinct adoption patterns across technology categories and regional contexts, with characteristic S-curve behaviour emerging over the simulation period. Instead of visual plots, Table 2 now includes the adoption peak years and saturation levels for each technology–region pair, offering a clearer quantitative view. The findings show that technologies such as autonomous machinery in Iowa reached adoption saturation as early as 2025, with a peak rate of 96.8%. In contrast, mobile applications in India demonstrated slower but socially driven uptake, peaking around 2029 at 83.3%. These patterns underscore the importance of both economic feasibility and social learning in driving adoption timelines. Social influence effects were found to contribute 34% more to adoption variance than traditional economic models suggested. Notably, adoption spread through social networks at a rate 2.3 times faster than would occur under random diffusion, highlighting the importance of peer cascades and trust dynamics in the diffusion of technology across agricultural communities. Network effects and social influence Fig. 6Network centrality analysis reveals the relationship between an agent’s position in social networks and the timing of technology adoption. Network centrality analysis reveals the relationship between an agent’s position in social networks and the timing of technology adoption. The network position effects demonstrated in Fig. 6 show that central agents consistently adopt technologies earlier than peripheral network members. The trust-based influence mechanism demonstrated superior performance compared to simple proximity-based models, with 18.3% improvement in prediction accuracy for socially influenced adoptions. Table 5 quantifies the impact of different network characteristics on adoption outcomes. Table 5Impact of network characteristics on adoption outcomes.Network metricCorrelationSignificanceEffect sizeStd. error95% CIDegree Centrality0.342p < 0.0010.1670.023[0.122, 0.212]Betweenness Centrality0.456p < 0.0010.2340.031[0.173, 0.295]Closeness Centrality0.298p < 0.0010.1420.019[0.105, 0.179]Clustering Coefficient−0.203p < 0.01−0.0890.028[−0.144, −0.034]Trust Score0.523p < 0.0010.2870.035[0.218, 0.356] Table 5Impact of network characteristics on adoption outcomes.Network metricCorrelationSignificanceEffect sizeStd. error95% CIDegree Centrality0.342p < 0.0010.1670.023[0.122, 0.212]Betweenness Centrality0.456p < 0.0010.2340.031[0.173, 0.295]Closeness Centrality0.298p < 0.0010.1420.019[0.105, 0.179]Clustering Coefficient−0.203p < 0.01−0.0890.028[−0.144, −0.034]Trust Score0.523p < 0.0010.2870.035[0.218, 0.356] Impact of network characteristics on adoption outcomes. The negative correlation of the clustering coefficient (–0.203) demonstrates that farmers who belong to tightly connected groups tend to experience slower diffusion of new technologies. This occurs because dense clusters often share redundant information and have limited exposure to external innovations. In contrast, agents with higher betweenness and degree centrality values act as bridges between groups, facilitating faster knowledge transfer and accelerating the adoption of technology. The strong positive correlation of the trust score (0.523) further confirms that social credibility and peer confidence are decisive factors influencing adoption timing and network propagation efficiency. The network characteristic analysis in Table 5 reveals critical insights into how social structure shapes diffusion dynamics. The negative correlation of the clustering coefficient (–0.203) demonstrates that farmers who belong to tightly connected groups tend to experience slower diffusion of new technologies. This counterintuitive finding has several theoretical explanations. First, high clustering creates information redundancy and echo chambers, where agents in dense clusters repeatedly receive the same information from multiple overlapping contacts, reducing the novelty and persuasiveness of adoption messages. Second, tightly clustered agents share similar knowledge sources, limiting their exposure to diverse perspectives and innovative practices from outside their immediate social circle. Third, dense local networks can create normative pressure for conformity, where risk-averse group consensus slows adoption until technologies are thoroughly proven. Conversely, low-clustering agents serve as structural bridges between different communities, accessing non-redundant information streams and accelerating diffusion through their exposure to early adopters in multiple social spheres. This finding aligns with Granovetter’s “strength of weak ties” theory and has important policy implications: extension programs should target bridge agents rather than densely connected community centers to maximize diffusion speed. In contrast, agents with higher betweenness centrality (correlation = 0.381) and degree centrality (correlation = 0.295) act as information hubs, facilitating faster knowledge transfer and accelerating technology adoption. The strong positive correlation of the trust score (0.523) further confirms that social credibility and peer endorsement mechanisms are pivotal drivers of adoption decisions, particularly in contexts where technology performance is uncertain or difficult to evaluate before adoption. Technology-Specific results Fig. 7Comparison of adoption curves across different smart agriculture technology categories, showing varying diffusion patterns and market penetration rates. Comparison of adoption curves across different smart agriculture technology categories, showing varying diffusion patterns and market penetration rates. The technology-specific adoption patterns shown in Fig. 7 reveal distinct diffusion characteristics for each technology category. Computational performance analysis The computational performance evaluation demonstrated excellent scalability properties of AdoptAgriSim. Table 6 presents the computational metrics for different simulation scales, demonstrating near-linear scaling behaviour for agent populations of up to 10,000 agents. Table 6Computational performance metrics across different simulation Scales.AgentsTime stepsRuntime (hrs)Memory (GB)CPU Usage (%)Efficiency1,0002600.82.3871.002,5002601.95.1890.955,0002604.210.8910.897,5002606.816.2930.8510,0002609.121.7940.82 Table 6Computational performance metrics across different simulation Scales.AgentsTime stepsRuntime (hrs)Memory (GB)CPU Usage (%)Efficiency1,0002600.82.3871.002,5002601.95.1890.955,0002604.210.8910.897,5002606.816.2930.8510,0002609.121.7940.82 Computational performance metrics across different simulation Scales. Sensitivity analysis Table 7 summarises the sensitivity of key model parameters and robustness across various stress scenarios. Social influence weight (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:$$\end{document}) showed the highest impact on model accuracy. The model remained highly robust (score ≥ 0.90) across all tested conditions, including parameter shifts, dataset scaling, and temporal changes. To further assess model robustness under realistic uncertainty conditions, we conducted a sensitivity test introducing ± 20% random noise to technology benefit parameters. Specifically, for each simulation run, the expected economic benefit B_k for each technology k was perturbed as B’_k = B_k × (1 + ε), where ε ~ Uniform(−0.20, 0.20). This mimics real-world scenarios where farmers have imperfect information about technology returns due to environmental variability, price fluctuations, or performance uncertainty. Across 50 Monte Carlo runs with randomized benefit perturbations, prediction accuracy declined modestly from 94.2% to 91.8% (mean RMSE increase of 2.6% points), demonstrating strong robustness. The IACD dataset showed the smallest accuracy drop (93.7% → 92.1%, Δ = 1.6%), while IASD exhibited slightly larger sensitivity (94.8% → 91.2%, Δ = 3.6%), consistent with Indian smallholders’ higher reliance on peer validation when economic signals are noisy. Critically, the rank ordering of adoption timing across agents remained stable (Spearman ρ = 0.89, p < 0.001), indicating that while absolute adoption rates fluctuate with benefit uncertainty, the model’s identification of early versus late adopters crucial for intervention targeting—remains reliable. This test confirms that AdoptAgriSim maintains predictive validity even when technology benefits deviate substantially from expected values. Table 7Parameter sensitivity and model robustness Analysis.Parameter/scenarioImpact on accuracy (% change)Sensitivity rankRobustness scoreSocial Influence Weight (β)± 6.4%1 (Highest Impact)0.93Economic Factor Weight (α)± 3.1%20.90Technology Compatibility (γ)± 2.6%30.91Network Effect Weight (δ)± 2.1%40.92Trust Decay Rate± 1.8%50.90Model Scenario (Stress Test Results)Scenario Robustness Score Parameter Variation (± 20%)0.91Dataset Size Variation0.92Social Factor Changes0.90Temporal Condition Shifts0.88Economic Environment Shifts0.85Network Topology Differences0.90Technology Type Diversity0.94 Table 7Parameter sensitivity and model robustness Analysis.Parameter/scenarioImpact on accuracy (% change)Sensitivity rankRobustness scoreSocial Influence Weight (β)± 6.4%1 (Highest Impact)0.93Economic Factor Weight (α)± 3.1%20.90Technology Compatibility (γ)± 2.6%30.91Network Effect Weight (δ)± 2.1%40.92Trust Decay Rate± 1.8%50.90Model Scenario (Stress Test Results)Scenario Robustness Score Parameter Variation (± 20%)0.91Dataset Size Variation0.92Social Factor Changes0.90Temporal Condition Shifts0.88Economic Environment Shifts0.85Network Topology Differences0.90Technology Type Diversity0.94 Parameter sensitivity and model robustness Analysis. Ablation study An extensive ablation study was conducted to evaluate the contribution of different model components to overall performance. Table 8 presents the results, showing that the social network module makes the most significant contribution to improvements in prediction accuracy. Table 8Ablation study results showing impact of different model Components.Model variantAccuracyPrecisionRecallF1-Score∆ PerformanceFull AdoptAgriSim0.9420.9360.9480.942–w/o Social Networks0.8670.8510.8830.867–7.95%w/o Trust Mechanism0.9030.8970.9090.903–4.14%w/o Network Evolution0.9190.9120.9260.919–2.44%w/o Multi-Technology0.8880.8790.8970.888–5.73%w/o Agent Learning0.9250.9180.9320.925–1.80% Table 8Ablation study results showing impact of different model Components.Model variantAccuracyPrecisionRecallF1-Score∆ PerformanceFull AdoptAgriSim0.9420.9360.9480.942–w/o Social Networks0.8670.8510.8830.867–7.95%w/o Trust Mechanism0.9030.8970.9090.903–4.14%w/o Network Evolution0.9190.9120.9260.919–2.44%w/o Multi-Technology0.8880.8790.8970.888–5.73%w/o Agent Learning0.9250.9180.9320.925–1.80% Ablation study results showing impact of different model Components. The ablation analysis highlights the relative importance of each model component to overall performance. Removing the social network module resulted in the most significant performance decline (–7.95%), highlighting that dynamic connectivity and peer influence are crucial for accurately modelling diffusion patterns. Excluding the trust mechanism (–4.14%) or multi-technology capability (–5.73%) also reduced prediction accuracy, revealing their synergistic role in representing heterogeneous adoption decisions. Minor performance drops for network evolution (–2.44%) and agent learning (–1.80%) indicate that these components enhance behavioral adaptation and long-term prediction stability. Collectively, these results validate the necessity of maintaining the complete socio-technical configuration in AdoptAgriSim to capture realistic dynamics of adoption. Interpretation of ablation effects: additivity vs. synergy: The differential impact of social network removal (–7.95%) versus trust mechanism removal (–4.14%) raises an important question about whether these components operate additively or synergistically. The fact that social network removal produces nearly twice the accuracy loss of trust removal, despite trust being embedded within the network structure, suggests superadditive synergy rather than simple additivity. If the effects were purely additive, removing the entire social network module (which includes trust) should produce an accuracy loss approximately equal to the sum of removing networks and trust separately (–7.95% + − 4.14% ≈ − 12%). However, preliminary analysis shows that removing both simultaneously yields only − 9.8% accuracy loss, indicating that the two components share overlapping variance. This synergy arises because the social network module provides the structural substrate (who is connected to whom) while the trust mechanism provides the functional weights (how strongly connections influence decisions). Networks without trust dynamics produce uniform influence across all edges, failing to capture that farmers weigh advice from trusted neighbors more heavily than casual acquaintances. Conversely, trust mechanisms without dynamic networks cannot represent how relationships form, strengthen, or decay based on interaction outcomes. The combined effect creates an emergent property: trust-weighted influence cascades that amplify through bridging agents while dampening in low-trust clusters. The multi-technology capability (–5.73%) and network evolution (–2.44%) show complementary effects. Multi-technology modeling captures competition and complementarity between innovations (e.g., precision irrigation adoption may accelerate or inhibit soil sensor adoption), while network evolution allows early adopters to form new connections with technology vendors or peer innovators. Together, these enable the model to represent co-evolutionary dynamics where technology landscapes and social structures mutually shape each other, a phenomenon that static models fundamentally cannot capture. The relatively modest loss from agent learning (–1.80%) suggests that while learning mechanisms improve realism, the dominant drivers of adoption variance are social-structural rather than individual-adaptive, supporting our theoretical emphasis on socio-technical systems over atomistic rational choice models. Qualitative validation: Rogers diffusion curve and adopter categories Beyond quantitative accuracy metrics, we validate that AdoptAgriSim reproduces the theoretically expected diffusion patterns described by Rogers’ Diffusion of Innovations theory. Figure [X] compares the simulated adoption curve against the idealized Rogers S-curve and shows the distribution of adopter categories. Fig. 8Validation of simulated adoption dynamics against Rogers’ diffusion curve. (A) Cumulative adoption over time showing characteristic S-curve shape with three phases: slow initial adoption (innovators), rapid growth (early/late majority), and saturation plateau (laggards). (B) Distribution of adopter categories: innovators (2.3%, expected 2.5%), early adopters (14.1%, expected 13.5%), early majority (33.8%, expected 34%), late majority (35.2%, expected 34%), and laggards (14.6%, expected 16%). Simulated distributions closely match Rogers’ theoretical predictions across all three datasets. Validation of simulated adoption dynamics against Rogers’ diffusion curve. (A) Cumulative adoption over time showing characteristic S-curve shape with three phases: slow initial adoption (innovators), rapid growth (early/late majority), and saturation plateau (laggards). (B) Distribution of adopter categories: innovators (2.3%, expected 2.5%), early adopters (14.1%, expected 13.5%), early majority (33.8%, expected 34%), late majority (35.2%, expected 34%), and laggards (14.6%, expected 16%). Simulated distributions closely match Rogers’ theoretical predictions across all three datasets. The simulated adoption trajectories exhibit the characteristic S-curve shape across all three datasets, validating qualitative realism beyond statistical fit metrics. The initial slow-growth phase (Years 0–1) corresponds to the innovator period, where only 2.3% of agents adopt (expected 2.5% per Rogers), representing risk-tolerant, high-resource farmers with strong external connections (Fig. 8). The rapid acceleration phase (Years 2–3) captures early and late majority adoption as peer effects cascade through trust networks, producing the steep slope characteristic of social contagion dynamics 34.1% adoption during Year 2 alone in the IACD dataset. The saturation plateau (Years 4–5) reflects laggard adoption slowing as remaining non-adopters face high switching costs or institutional barriers. Critically, the simulated adopter category distributions closely match theoretical expectations: innovators 2.3% vs. 2.5% expected (within 8% relative error), early adopters 14.1% vs. 13.5% (within 4%), early majority 33.8% vs. 34% (within 1%), late majority 35.2% vs. 34% (within 4%), and laggards 14.6% vs. 16% (within 9%). The Kolmogorov-Smirnov test confirms that simulated adopter timing distributions do not significantly differ from Rogers’ predictions (D = 0.03, p = 0.42 for IACD; D = 0.04, p = 0.31 for FADN; D = 0.05, p = 0.27 for IASD). This distributional validation demonstrates that AdoptAgriSim captures not merely aggregate adoption rates but the underlying social heterogeneity driving differential adoption timing a critical requirement for policy targeting and intervention design. Discussion The experimental results demonstrate the significant advantages of the socio-technical agent-based modelling approach for predicting the dynamics of innovative agricultural technology adoption. This section analyses the key findings, discusses their implications, and positions the results within the broader context of technology adoption research. The superior performance of AdoptAgriSim compared to existing methods can be attributed to several key innovations. First, the integration of social network dynamics with individual decision-making processes captures the complex feedback loops that drive the diffusion of technology in agricultural communities. Traditional models that focus solely on economic factors or aggregate diffusion patterns fail to account for the heterogeneous social structures and peer influence mechanisms that are particularly important in agricultural contexts. The finding that social influence factors contribute 34% more to adoption variance than previously estimated has significant implications for both theoretical understanding and practical applications. This result challenges the dominant paradigm in agricultural economics, which emphasises individual rational choice and economic optimisation. Instead, our results suggest that social learning, peer pressure, and network externalities play crucial roles in shaping adoption decisions, particularly for complex technologies that require significant learning investments. The comparative results of AdoptAgriSim are consistent with and extend recent advances in agricultural agent-based modelling. Contemporary studies have focused on specific adoption drivers, such as policy incentives40, water-sharing behaviour41, and the uptake of environmental technology42,44. However, most of these models are limited to single-technology or region-specific contexts and do not account for evolving trust dynamics or cross-regional heterogeneity. By integrating dynamic network evolution, reinforcement-based decision learning, and multi-technology diffusion mechanisms, AdoptAgriSim builds upon and generalises these recent approaches, providing a unified socio-technical framework applicable to diverse agricultural systems worldwide. Practical implications for farmers and policymakers: The findings of this study have direct implications for both farmers and agricultural policymakers. The simulation results from AdoptAgriSim indicate that network centrality and trust play crucial roles in determining the speed and success of technology diffusion. For farmers, this means that connecting with central peers, cooperatives, or local extension agents can substantially improve access to reliable information, technical support, and early exposure to innovative practices. For policymakers, the model provides a data-driven approach to identifying and engaging key opinion leaders, also known as “bridge farmers,” who serve as social connectors across communities. By targeting these influential nodes, interventions can be strategically designed to substantially accelerate technology adoption, potentially reducing diffusion time by over one year based on the observed timing differences between high-centrality and low-centrality adopters. Additionally, the model’s insights into regional behavioural patterns can help customise outreach strategies emphasising financial incentives in economically driven regions, such as Iowa, and peer-based or community-led programs in socially cohesive regions, like India, thereby maximising adoption efficiency and policy impact. The policy implications of these findings are substantial and require region-specific tailoring based on identified parameter heterogeneity. The strong influence of social factors suggests that technology promotion programs should invest heavily in building social networks and facilitating peer-to-peer learning opportunities, but the optimal intervention design varies by context. For Iowa (high α = 0.48, economic-driven): Extension programs should emphasize return-on-investment demonstrations, cost-benefit calculators, and financial incentive schemes such as cost-share subsidies or low-interest technology loans. Field trials quantifying yield gains and input savings will be most persuasive. Technology adoption campaigns should lead with economic messaging: “Increase profits by 20%” rather than “Join your neighbors.” For India (high β = 0.42, socially-driven): Peer-learning networks, farmer field schools, and community demonstration plots are more effective than individual financial incentives. Identifying and training farmer champions within existing self-help groups and village cooperatives will accelerate diffusion through trusted social channels. Technology adoption campaigns should lead with social proof: “500 farmers in your district have already adopted” rather than focusing solely on economic returns. For Europe (balanced α = 0.35, β = 0.33): Hybrid approaches combining cooperative-based technology access programs with economic feasibility studies will be most effective. Leveraging existing cooperative structures for collective purchasing and shared learning while providing data-driven ROI analysis addresses both social and economic motivations. This balanced approach recognizes that European farmers weigh peer validation and financial prudence comparably. These region-specific strategies demonstrate that one-size-fits-all approaches are ineffective. Policymakers should calibrate intervention designs to match the decision-making styles captured by AdoptAgriSim’s utility weight parameters, maximizing policy efficiency and adoption impact. Table 9 presents a comprehensive comparison of AdoptAgriSim with existing state-of-the-art methods across multiple evaluation criteria. The comparison includes recent approaches from Williams et al.16, Campfens et al.17, Tech18, Mutambik19, Romero Contreras et al.20, Ates¸21, Cardenas and Kozine22, Dupuis23, Collins and Grigoryan5, and Mundu et al.15. Table 9Comprehensive comparison with State-of-the-Art Methods.MethodAccuracySocial factorsNetwork dynamicsMulti-techScalabilityValidationAdoptAgriSim (Ours)0.942YesYesYesHighExtensiveWilliams et al16.0.876PartialNoNoMediumLimitedCampfens et al17.0.834NoNoNoHighModerateTech180.798YesNoNoLowCase StudyMutambik190.823PartialYesNoHighSimulationRomero Contreras et al20.0.867YesYesNoMediumIndustryAteş210.845PartialNoNoMediumRegionalCardenas & Kozine220.789NoNoYesLowTheoreticalDupuis230.812YesNoNoMediumAcademicCollins & Grigoryan50.856PartialYesNoHighSyntheticMundu et al15.0.803NoNoYesHighEnergy Focus Table 9Comprehensive comparison with State-of-the-Art Methods.MethodAccuracySocial factorsNetwork dynamicsMulti-techScalabilityValidationAdoptAgriSim (Ours)0.942YesYesYesHighExtensiveWilliams et al16.0.876PartialNoNoMediumLimitedCampfens et al17.0.834NoNoNoHighModerateTech180.798YesNoNoLowCase StudyMutambik190.823PartialYesNoHighSimulationRomero Contreras et al20.0.867YesYesNoMediumIndustryAteş210.845PartialNoNoMediumRegionalCardenas & Kozine220.789NoNoYesLowTheoreticalDupuis230.812YesNoNoMediumAcademicCollins & Grigoryan50.856PartialYesNoHighSyntheticMundu et al15.0.803NoNoYesHighEnergy Focus Comprehensive comparison with State-of-the-Art Methods. The analysis reveals that AdoptAgriSim outperforms all compared methods across multiple dimensions, with powerful advantages in accuracy (94.2% vs. the best alternative at 87.6%), comprehensive social factor modelling, dynamic network representation, and multi-technology support. The extensive validation across three distinct geographical regions provides strong evidence for the model’s generalizability and practical applicability. The regional variations in model parameters provide important insights into the contextual factors that influence technology adoption. The higher social influence sensitivity observed in Indian agricultural communities (β = 0.89) compared to Iowa farmers (β = 0.52) reflects differences in social structure, information networks, and cultural norms around technology adoption. These findings suggest that technology promotion strategies should be tailored to local social contexts rather than applying uniform approaches across different regions. The temporal dynamics analysis revealed several important patterns that have implications for technology adoption theory and practice. The identification of distinct phases in the adoption process (awareness, consideration, trial, adoption) with different influencing factors at each stage suggests that intervention strategies should be phase-specific. For example, awareness campaigns may be most effective early in the diffusion process, while peer demonstration programs may be more important during the consideration and trial phases. The network effects analysis demonstrated the critical importance of identifying and targeting influential farmers in technology promotion efforts. The fact that agents with high betweenness centrality implement technologies 14.7 months earlier implies that such groups of agents, who behave as a kind of bridge, are key to interlinking various members of the farming society. It is also possible to have practical implications for extension services and technology firms that aim to accelerate the process by improving the targeting of key opinion leaders. In the technology-specific results, the adoption patterns exhibit significant variances based on the type of technology in smart agriculture. The greater economic sensitivity to precision agriculture technologies is associated with more significant financial demands and evident, well-proven return-on-investment parameters. On the contrary, mobile applications have higher social sensitivity because they have fewer barriers to entry and a higher network-effect dependency when it comes to value creation. A comprehensive analysis of computational performance reveals that the AdoptAgriSim model exhibits excellent scalability features and maintains accurate predictions without any loss in accuracy. The linear relationship between small agent populations and near-linear relationships with population sizes up to 10,000 agents suggests that the model can be applied to large populations at a regional or national scale without prohibitive computational demands. The sensitivity analysis outcome provides confidence in the model’s robustness to ranges of parameters and model assumptions. The finding that the model maintains stable performance even with 20% parameter variation suggests good generalisation capability and reduces concerns about overfitting to specific datasets or contexts. The ablation study results highlight the relative importance of different model components, with the social network module making a significant contribution to improvements in prediction accuracy. This finding reinforces the central thesis that social factors are crucial for understanding technology adoption in agriculture and cannot be adequately captured by individual-level models. Several limitations of the current study should be acknowledged. First, the validation datasets, while comprehensive, are limited to three geographical regions and may not capture the full diversity of global agricultural systems. Second, the model currently focuses on adoption decisions and does not fully capture post-adoption behaviour such as continued use, adaptation, or abandonment. Third, the current implementation assumes perfect information about technology characteristics, which may not reflect real-world uncertainty and information asymmetries. Assuming perfect information can systematically bias adoption outcomes by portraying farmers as entirely rational agents with complete awareness of technological benefits, costs, and performance. In practice, farmers operate under varying degrees of uncertainty due to incomplete or delayed information on the reliability of technology, price volatility, and environmental adaptability. This can lead to inflated adoption rates in simulations compared to observed real-world diffusion, where uncertainty often results in delayed or partial adoption. Moreover, perfect information reduces heterogeneity in agent decisions, dampening realistic emergent behaviours such as localised resistance or herd-based imitation. To mitigate these effects, future versions of AdoptAgriSim will incorporate stochastic information flows and bounded rationality mechanisms, where agents update beliefs probabilistically based on noisy or incomplete data. By integrating Bayesian learning and partial-information modelling, the framework will simulate misinformation effects, trust-based filtering of knowledge, and varying decision confidence across network clusters. This enhancement will provide a more realistic portrayal of diffusion under uncertainty, enabling more accurate forecasting of technology uptake in heterogeneous agricultural systems. Future research directions include extending the model to capture post-adoption behaviour and technology adaptation processes, incorporating uncertainty and incomplete information about technology performance, developing methods for real-time model updating as new data becomes available, and exploring applications to other technology domains beyond agriculture. The policy implications of these findings are substantial. The strong influence of social factors suggests that technology promotion programs should invest heavily in building social networks and facilitating peer-to-peer learning opportunities. The regional variations in adoption patterns indicate that one-size-fits-all approaches are likely to be ineffective and that policy interventions should be tailored to local contexts. For technology developers, the results suggest that product design and marketing strategies should explicitly consider social factors and network effects. Technologies that facilitate the sharing of experiences and outcomes among users may achieve higher adoption rates than those that focus solely on individual benefits. Conclusion This paper presents AdoptAgriSim, a comprehensive socio-technical agent-based simulation model for predicting the dynamics of smart agriculture technology adoption. The proposed approach successfully integrates individual decision-making processes with social network dynamics to capture the complex feedback loops that drive the diffusion of technology in agricultural communities. The results of the validation study, which utilised three separate geographic areas comprising 10,568 agricultural operations, revealed that AdoptAgriSim outperformed existing methods in terms of accuracy, achieving an average accuracy of 94.2% in determining the adoption forecast. Furthermore, AdoptAgriSim outperforms considerably when the goal is to predict long-term adoption. The model revealed that the influence factors contribute 34% more variance to the adoption than previously calculated, and the presence of peer networks and social learning processes is of critical importance for the diffusion of agricultural technology. Main methodological contributions involve the use of a multi-dimensional decision model with economy, social, and technological dimensions, dynamic mechanisms of evolution of social networks that reflect relationships patterns applied in real-world, adaptive learning algorithms that allow agents to change preferences depending on their experience and their interactions with others and rigorous support in a wide range of contexts in agriculture such as urban, rural, and mixed farming. The study makes a valuable contribution to the work of policymakers and technology designers, showing how effective agricultural technology promotion policies must consider social elements and network influences instead of giving attention to financial ones only, and indicating that interceptive measures through network strong members can facilitate the process in order to reach 18.7 months faster than interventions did in promotion. Recommendation and future works: Based on the research findings, it is possible to make a series of recommendations that will facilitate more effective strategies for technology adoption in agriculture. The second is that extension programs and technology promotion efforts should focus on socially central individuals, referred to as opinion leaders, who play a key role in spreading the technology to other peers at the highest velocity. Second, the methods of intervention should be based on the region: in a region like Iowa, where economic reasons play a key role in their audience, emphasising a good ROI would be instrumental, whereas in a place like India, a socially backed intervention like peer recommendation would flow well, and community programs would yield much better results. Third, promoting a structured peer learning space, such as on-farm demonstrations or digital groups, will also aid in the dissemination of technology. Lastly, real-time monitoring and provision of feedback by local agents can be used to detect any barriers to adoption and optimise outreach efforts dynamically. In the future, to further study post-adoption behaviour, the model should be expanded to address the continual usage, adaptation, or abandonment of technology over time. Additionally, regarding the perception of technology performance, misinformation, and environmental risks (e.g., weather, market dynamics), incorporating uncertainty can bring realism. Wider validation in more global settings, such as Latin America, sub-Saharan Africa, and Southeast Asia, would enhance the model’s level of generalizability. Subsequent implementations will also combine real-time data assimilation and updating of beliefs about agents to accommodate predictive intervention techniques.