STDDN: A Physics-Guided Deep Learning Framework for Crowd Simulation

Zijin Liu, Xu Geng, Wenshuai Xu, Xiang Zhao, Yan Xia, You Song · 2026

Read STDDN: A Physics-Guided Deep Learning Framework for Crowd Simulation by Zijin Liu, Xu Geng, Wenshuai Xu, Xiang Zhao, Yan Xia, You Song (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-02756.txt.

Introduction

Liu and five colleagues propose STDDN, a physics-guided framework for crowd simulation. Treating crowds as lone paths invites error buildup and unstable runs, since lone paths miss large-scale physical laws. The framework imposes the fluid continuity equation as a strong constraint. A neural ordinary differential equation models how crowd density evolves from individual motion, keeping path forecasts physically sound. The work targets safety planning, evacuation planning, and traffic systems.

Catalog record

CREATORZijin Liu, Xu Geng, Wenshuai Xu, Xiang Zhao, Yan Xia, You Song
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE63906
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.02756

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

STDDN: A Physics-Guided Deep Learning Framework for Crowd Simulation Zijin Liu, Xu Geng, Wenshuai Xu, Xiang Zhao, Yan Xia, You Song Abstract Accurate crowd simulation is crucial for public safety management, emergency evacuation planning, and intelligent transportation systems. However, existing methods, which typically model crowds as a collection of independent individual trajectories, are limited in their ability to capture macroscopic physical laws. This microscopic approach often leads to error accumulation and compromises simulation stability. Furthermore, deep learning-driven methods tend to suffer from low inference efficiency and high computational overhead, making them impractical for large-scale, efficient simulations. To address these challenges, we propose the Spatio-Temporal Decoupled Differential Equation Network (STDDN), a novel framework that guides microscopic trajectory prediction with macroscopic physics. We innovatively introduce the continuity equation from fluid dynamics as a strong physical constraint. A Neural Ordinary Differential Equation (Neural ODE) is employed to model the macroscopic density evolution driven by individual movements, thereby physically regularizing the microscopic trajectory prediction model. We design a density-velocity coupled dynamic graph learning module to formulate the derivative of the density field within the Neural ODE, effectively mitigating error accumulation. We also propose a differentiable density mapping module to eliminate discontinuous gradients caused by discretization and introduce a cross-grid detection module to accurately model the impact of individual cross-grid movements on local density changes. The proposed STDDN method has demonstrated significantly superior simulation performance compared to state-of-the-art methods on long-term tasks across four real-world datasets, as well as a major reduction in inference latency. Code: https://github.com/liuzjin/STDDN . 1 Introduction Crowd simulation ( 18 ; 8 ; 27 ; 41 ; 32 ) has become a cornerstone technology in domains such as public safety management and intelligent transportation systems. It holds both significant research value and practical importance for real-world problems, such as crowd evacuation in densely populated areas and passenger flow organization in transportation hubs. Accurate simulation of crowd dynamics not only aids in optimizing spatial design and management strategies but also offers critical decision support for emergency responses to unforeseen events ( 41 ) . Existing mainstream approaches to crowd simulation generally fall into three categories: physics-based methods, data-driven deep learning approaches, and physics-guided deep learning methods ( 32 ; 43 ; 35 ; 2 ; 44 ) . Physics-based methods simulate local behaviors like avoidance and following by characterizing physical interactions among individuals ( 12 ; 34 ) . These methods are grounded in solid theoretical foundations and offer strong physical interpretability. However, they are typically built upon simplified linear mechanics assumptions, making it difficult to capture the nonlinear and stochastic characteristics inherent in crowd behaviors. As a result, their performance drops significantly in high-density or highly interactive environments. With the advancement of deep learning, data-driven methods (e.g., recurrent neural networks ( 3 ) , graph neural networks ( 40 ) , and diffusion models ( 14 ) ) have achieved remarkable progress in trajectory prediction by learning motion patterns from large-scale trajectory datasets. For example, STGCNN ( 28 ) learns trajectory evolution from historical data, while MID ( 11 ) uses a diffusion mechanism to estimate the probability distribution of future trajectories. However, the absence of physical constraints in these methods often leads to unrealistic behaviors that violate fundamental physical laws, such as unnatural crowd congestion or collisions that ignore avoidance principles. To bridge the gap between physical consistency and expressive modeling capacity, physics-guided deep learning approaches have emerged. These methods aim to integrate physical models into neural network architectures or training processes as inductive biases (e.g., PCS ( 43 ) , SPDiff ( 2 ) ). While improving physical consistency, most focus on microscopic interactions and struggle to capture macroscopic density evolution, leading to error accumulation and suboptimal results. Additionally, SPDiff’s diffusion-based paradigm suffers from slow inference and high computational cost, limiting its scalability. Based on these insights, we propose a novel Spatio-Temporal Decoupled Differential Equation Network (STDDN) for crowd simulation. By incorporating the continuity equation from fluid dynamics as a structural constraint, STDDN guides the trajectory prediction process from a global perspective of density evolution. In contrast to existing physics-guided methods that primarily rely on individual-level modeling, STDDN treats the crowd as a continuous medium and reformulates trajectory evolution as a transport process in the density field. It uses neural networks to learn the coupling between the density and velocity fields, enabling macroscopic modeling of collective dynamics. Meanwhile, STDDN retains the strong representation capability of data-driven methods, enhancing both physical consistency and global interpretability while supporting end-to-end training. Experiments on multiple benchmark datasets demonstrate that the proposed method achieves higher prediction accuracy and greater inference efficiency in long-term simulation tasks, showing strong generalization ability and engineering applicability. Key contributions of this paper include: • A unified macro-micro coupled modeling framework. We incorporate the continuity equation as a differentiable physical constraint, achieving a tight, end-to-end integration of differential equations with deep neural networks for the task of crowd trajectory prediction. This design significantly enhances the physical consistency and global stability of the simulation. • Physically interpretable design of a dynamic graph network. A cross-timestep dynamic graph neural network is constructed by using current velocity as incoming edges and future velocity as outgoing edges, enabling explicit modeling of density flux over time and improving the interpretability of the simulation process. • Two differentiable structures for enhanced physical consistency. We propose a differentiable density mapping module based on radial basis functions and a continuous cross-grid detection module. These designs mitigate gradient discontinuities caused by the discretization process and ensure both mass conservation and smooth backpropagation. • Superior accuracy and efficiency. Experimental results on four real-world datasets demonstrate that STDDN outperforms existing state-of-the-art methods in long-term crowd simulation accuracy, while also achieving a substantial increase in inference speed. 2 Related Work Crowd simulation. As discussed in the previous section, there are three main approaches in crowd simulation research: physics-based methods, data-driven deep learning methods, and physics-guided deep learning approaches. Physics-based methods were among the earliest frameworks studied ( 6 ; 12 ; 30 ; 5 ) . For example, Social Force Model (SFM) ( 12 ) pioneered dynamic modeling of sparse crowds by formulating a resultant force system between pedestrians, the environment, and their goals. However, its underlying linear force assumptions exhibit clear limitations in high-density scenarios. To account for speed differences among pedestrians, a fine-grid cellular automaton model ( 34 ) was proposed to enable more realistic crowd simulations. With advancements in data acquisition and modeling techniques, data-driven deep learning methods have gained wide adoption. PECNet employs a Conditional Variational Autoencoder (CVAE) ( 15 ) to generate multiple possible trajectories and samples a plausible endpoint based on observed data. STGCNN ( 28 ) leverages graph neural networks to model interactions with surrounding pedestrians at each time step. MID ( 11 ) directly models the distribution of future trajectories using a diffusion model and designs a Transformer-based architecture to capture temporal dependencies in historical data. Although these methods achieve good performance in trajectory prediction, their lack of physical constraints limits generalization under extreme or unforeseen scenarios, often resulting in physically implausible behaviors. In recent years, researchers have begun exploring physics-guided deep learning methods for crowd simulation ( 21 ) . PCS ( 43

Read the complete text