A neural network model for managing renewable resources with population growth Authors: Sadique Ahmad, Israr Ahmad, Ala Saleh Alluhaidan, Mohammed A ElAffendi Published: 2026 ABSTRACT AbstractThis paper presents a fractional-order modelling framework for population-resource dynamics using the -Hilfer derivative with neural network-based validation. The model captures logistic population growth coupled to renewable resource biomass while incorporating memory effects through fractional calculus. Novel contributions include the application of the -Hilfer operator to population-resource systems, representing hereditary dynamics via fractional order and type parameter . A detailed qualitative analysis establishes existence, uniqueness, and Ulam-Hyers (UH) stability with explicit parameter conditions. A linearised quadrature numerical scheme for the -Hilfer problem is developed and validated through neural networks, achieving . Simulations reveal that fractional orders () produce smoother transients than integer-order counterparts, with governing memory strength and modulating response patterns. Critically, we identify as a sustainable harvesting threshold, beyond which resource collapse occurs for integer-order systems, while fractional memory () provides partial damping. These findings offer actionable policy insights, including optimal harvest limits and stabilisation strategies, advancing both theoretical understanding and practical tools for sustainable resource management.Keywords: Renewable resource, -Hilfer derivative, Qualitative study, Linearized quadrature method, Neural network approximationsSubject terms: Engineering, Mathematics and computing Introduction Population growth and the sustainable exploitation of renewable resources are major concerns in environmental management, ecological modelling and applied mathematics1. The increase in population size is putting pressure on finite natural resources like water, forests, and energy sources. These resources regenerate at a specific rate defined by the ecological boundaries and historical utilisation patterns. This imbalance between population growth and resource regeneration may cause ecological instability, depletion and unsustainability2. Therefore, the development of a robust mathematical model that accurately represents the interconnected dynamics of population growth and renewable resources is an urgent scientific and managerial priority. Recent advances in fractional calculus have provided novel options for modelling complex biological and ecological systems. Neural networks have been successfully applied across diverse domains, including environmental modelling of chemostat systems3 and medical prediction of conjunctivitis infections4, showcasing the broad applicability of fractional calculus combined with neural networks. In epidemiological modelling, hybrid neural network architectures combining logistic sigmoid, radial basis, and hyperbolic tangent functions have been successfully applied to model M-pox disease dynamics5, while hybrid GA-IPA and neural network-based solutions have been developed for fractional-order nonlinear predator-prey dynamical systems6. Furthermore, radial basis and hyperbolic tangent neural network combinations have been utilised to solve herpes dynamics7, establishing the efficacy of neural networks in capturing complex biological behaviours. It is very clear from the literature that integer order population-resource models have been studied in detail8. The integer order model describes the status of the dynamics of the model only at the current time while failing to account for memory and hereditary effects. The dynamics of actual populations and resources are path-dependent in real-world ecological systems9. Their behaviours depend on the present but are also strongly influenced by the legacy of the past. Looking at this limitation of the integer order models, the interest in the fractional order models is increasing day by day10. The -Hilfer fractional derivative is selected over other fractional operators (Caputo, Riemann-Liouville, Caputo-Fabrizio) due to the following reasons. When , it reduces to the classical Hilfer derivative. When , it reduces to the classical Hilfer derivative. When and , it becomes the Caputo derivative. When and , it becomes the Caputo derivative. When and , it recovers the Riemann-Liouville derivative. When and , it recovers the Riemann-Liouville derivative. When , it corresponds to the Hadamard derivative. When , it corresponds to the Hadamard derivative. To develop a numerical scheme, we may use one of these conditions. This derivative gives us the generalisation of all other derivatives in a single framework. The -Hilfer operator maintains power-law kernels essential for long-term memory in ecological systems. The two parameters () provide control over fluctuations, and by setting and specific values, we recover classical operators. The considered model under the -Hilfer operator is given as follows: The detail of the variables and parameters is given as follows: Renewable resource biomass (e.g., forest volume, fish stock). Renewable resource biomass (e.g., forest volume, fish stock). Population density (human population dependent on the resource). Population density (human population dependent on the resource). K : Carrying capacity of the resource (maximum sustainable biomass). K : Carrying capacity of the resource (maximum sustainable biomass). L : Carrying capacity of the population. L : Carrying capacity of the population. r : Intrinsic growth rate of the resource (natural regeneration rate). r : Intrinsic growth rate of the resource (natural regeneration rate). s : Intrinsic growth rate of the population. s : Intrinsic growth rate of the population. E : Harvesting effort (e.g., fishing days, logging intensity). E : Harvesting effort (e.g., fishing days, logging intensity). q : Catchability coefficient measuring harvesting efficiency. q : Catchability coefficient measuring harvesting efficiency. Linear resource consumption rate (baseline per-capita consumption). Linear resource consumption rate (baseline per-capita consumption). Density-dependent consumption coefficient (nonlinear consumption effects). Density-dependent consumption coefficient (nonlinear consumption effects). Linear resource-assisted growth coefficient for the population. Linear resource-assisted growth coefficient for the population. Nonlinear (synergistic) resource benefit coefficient for the population. Nonlinear (synergistic) resource benefit coefficient for the population. To provide a clearer understanding of the proposed model, we elaborate on the ecological meaning of each interaction term. The term captures total consumption increasing with resource availability or population size. The term introduces nonlinearity at higher densities, reflecting resource competition, Allee effects, and handling time constraints. The resource-assisted growth terms, and , represent nutritional contribution to population growth scaling with both the variables, with the term capturing nonlinear synergistic effects at high resource abundance. The combination of neural networks and fractional calculus has opened a gateway for the analysis of different models, including renewable resources and population growth dynamics11. Neural networks are capable of validating numerical schemes, and with this capability we can predict the dynamics and different behaviours of models as well as hidden relationships12. However, a critical consideration is that when neural networks are trained on data generated by the same numerical scheme they are meant to validate, they primarily learn to replicate the outputs of the scheme rather than providing independent verification. Therefore, in this work we employ neural networks for the approximation and prediction of solutions, error analysis, and sensitivity analysis. Our work provides a mixture of fractional calculus, coupled population-resource models and artificial intelligence under the -Hilfer derivative. Thus, it develops a highly interesting and flexible framework for population-resource dynamics. In this analysis a thorough analysis is provided, including all qualitative aspects using non-linear functional analysis techniques. We not only present a novel numerical scheme for the -Hilfer coupled population-resource model but also provide a complete analysis of the validation of our numerical method with the help of a neural network. A detailed numerical simulation of the concerned model has been carried out using MATLAB 2024 to investigate the effect of the fractional order and type parameter on the population stability and renewable resource sustainability. The simulations contain 2D, 3D and phase portrait plots under different values of and . It is quite clear from the simulations how memory effects play a key role in regulating growth dynamics and resource regeneration. This study reinforced the importance of fractional calculus in ecological systems and depicts the practical importance of our approach. The network architecture consists of an input layer with 1 neuron, two hidden layers each containing 15 neurons with tanh activation, and an output layer with 1 neuron. The complete forward propagation for this 1–15-15-1 network is mathematically expressed as follows: More specifically, this forward propagation can be written as: where is the weight matrix and is the bias vector. The hyperbolic tangent activation function is used throughout the hidden layers. We investigated all possible aspects of our specified model with the help of a neural network. These include mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), maximum absolute error, regression analysis, best fit for the outputs of each compartment of the model, and a comparison of the approximate and neural network solutions. All these analyses are computed and illustrated graphically. The proposed structure of the neural network is presented in Fig. 1. Neural Network architecture. Neural networks have limitations such as poor extrapolation beyond training data, propagation of numerical errors, lack of interpretability, and high training demands. We counter these by using hold-out validation, multi-parameter comparisons, and detailed error metrics. We develop the Banach space such that and the norm is defined as follows: Definition 1.1 13. Let and J be a finite or infinite on the real line . Furthermore, the function be an increasing and positive monotone on (0, T). The left Riemann–Liouville fractional integral of a function with respect to another function is given as follows: Definition 1.2 13. Let and J be a finite or infinite on the real line and , increasing positive monotone function and the derivative of is not zero, for all values of t in the interval J. The -Hilfer left fractional derivative of order and type is defined in the following way: The rest of the work is organised as follows: The theoretical analysis of the model is given in Section “Theoretical analysis”, and Section “Stability” is devoted to the UH stability. Section "Numerical analysis and simulations" details the numerical analysis and simulations, including a quantitative comparison between integer-order and fractional-order versions of the model, as well as parameter sensitivity analyses. In Section "Neural network methodology", we present the neural network approximation. Finally, Section “Conclusion” summarises the key findings. Theoretical analysis The detailed qualitative analysis of Model (1) is provided in this section. The following functions are defined for the rest of the work. Therefore Model (1) may obtain the following form: Lemma 2.1 The solution of the Problem (2) obtain the following form: Where . From (3), we define an operator, : Consider the following assumptions for the remaining analysis: For all , there exists such that For all , there exist such that To verify assumptions () and (), we consider Since and are bounded quantities, there exist positive constants and such that and for all . Furthermore, we can write For the product terms, we write Therefore, for , we have Taking the maximum over and using the definition of the norm, we obtain Where Similarly, for , we have Where Now, we write Substituting the bounds for and : Since We obtain Therefore, assumption () holds with Lipschitz constant: For assumption (), we estimate for any as follows Taking the maximum over and using the definition of the norm, we obtain Similarly, for , we can write and therefore Now, we write Since and , we obtain Consider Then, we obtain Hence, we get This completes the verification of both assumptions () and (). Theorem 2.2 Under condition and the assumption , Problem (4) has a unique solution. Proof Let , we can write Obviously has a unique solution when . Theorem 2.3 Under assumptions and , there exists at least one solution to Problem (4). Proof To prove the conditions of Schauder’s theorem for the operator . Consider , such that and the convex subset is define as, Following are the steps for the proof of the theorem. Step-1: For the continuity of , consider is a sequence in , tends to , we consider Hence as . Thus is continuous on . Step-2: A bounded set has a bounded image in . For this, we need to prove that . Let , we can write Therefore in that image is uniformly bounded. One should keep this point in mind: that the expression gives maximum value as we take b close to zero. As b tends to zero, the expression approaches its maximum value. Hence the maximum value depends on the value of b. Step-3: To prove the equi-continuity of on . We take such that and we consider As limit , we get Because of this, is equi-continous and, as a result, meets every requirement of the Schaudar fixed point theorem. It follows that Problem (4) has at least one solution. Stability Stability analysis is a fundamental aspect of the qualitative study of dynamical systems. Determining the stability of system solutions allows us to characterise their long-term behaviour. More importantly, the qualitative dynamics of the system remain invariant under bounded perturbations to initial states or parameters. Therefore, stability analysis is an essential part of our work14. Definition 3.1 The considered Model (2) is UH stable if there exists , provided that for , there exists an approximate solution of the aforementioned model holds the following inequalities: Next, assuming that there exists a unique solution for the specified model. Moreover, we have Theorem 3.2 Under assumptions , Model (2) is UH-stable provided that Proof Consider a function and for , we have Let, , such that for and , which satisfies The solution of the preceding equation (5) becomes Consider as , has unique solution in our specified model then it becomes Now, consider Where Thus, our considered model is UH-Stable. Remark 3.3 The constants () depend on the ecological parameters. Consequently, the stability condition in Theorem 3.2 yields the following ecological interpretations: Larger intrinsic growth rates (r, s) increase the Lipschitz constants, requiring either a smaller time horizon T or a larger fractional order to maintain stability. Larger intrinsic growth rates (r, s) increase the Lipschitz constants, requiring either a smaller time horizon T or a larger fractional order to maintain stability. Higher harvesting effort (E) increases , thereby enlarging and potentially destabilising the system. Higher harvesting effort (E) increases , thereby enlarging and potentially destabilising the system. Larger carrying capacities (K, L) reduce the Lipschitz constants via the terms and , which promotes stability. Larger carrying capacities (K, L) reduce the Lipschitz constants via the terms and , which promotes stability. The fractional order directly affects the stability threshold through ; increasing expands the stability region. The fractional order directly affects the stability threshold through ; increasing expands the stability region. Numerical analysis and simulations Numerical techniques offer flexible approaches to solution approximation, decomposing complex problems into manageable steps. Simulations take these methods further, applying virtual models to simulate scenarios before they are actually applied. Together, they facilitate greater understanding, enhance efficiency, and promote advancement in science and technology. Numerical scheme development The Equation (3) obtain the following form with : where and is nonlinear. We partition the interval as with . On each subinterval , we approximate as linear: where . Then the integral in (6) in discritze form on is given as follows: Let . Therefore, we have Therefore, the approximation to the integral term in (6) is given as follows: Substituting into (6), we obtain the discrete scheme: Where Hence, in light of (7), the required numerical scheme for our Problem (1) is given as follows: for n = 1, 2,..., N. Numerical scheme error and stability The proposed linearized quadrature method achieves accuracy for sufficiently smooth solutions. This is justified as follows: Quadrature error: The piecewise linear approximation of on each subinterval introduces local truncation error of , leading to global error after summation. Quadrature error: The piecewise linear approximation of on each subinterval introduces local truncation error of , leading to global error after summation. Fractional integral approximation: The exact integrals and are computed without approximation, preserving the accuracy of the fractional kernel. Fractional integral approximation: The exact integrals and are computed without approximation, preserving the accuracy of the fractional kernel. Stability: For , the scheme is A-stable. This follows from two properties: (i) the fractional kernel is positive and completely monotonic, ensuring that the numerical solution inherits the contractivity of the continuous problem; and (ii) the quadrature weights and are positive and decreasing, which guarantees that the amplification factor remains bounded by unity for all step sizes . Stability: For , the scheme is A-stable. This follows from two properties: (i) the fractional kernel is positive and completely monotonic, ensuring that the numerical solution inherits the contractivity of the continuous problem; and (ii) the quadrature weights and are positive and decreasing, which guarantees that the amplification factor remains bounded by unity for all step sizes . A rigorous error bound can be established under the assumption . Following the standard analysis of fractional linear multistep methods, there exists a constant , depending only on , T, and the norm of , such that holds for all , where denotes the exact solution and the numerical approximation. Algorithm for the numerical scheme Linearized Quadrature Method for -Hilfer Population-Resource System Numerical simulations Numerical simulations of Model (1) are performed to investigate the impact of the fractional order and type parameter . We consider the initial conditions and parameters values with the step size 15. Figure 2 presents a comparison between the time evolution of resource biomass and population density . It is clear from the figure that resource biomass increases sharply compared to population density, but as the fractional order and type parameter are changed, significant variation in the solution curves becomes visible. The subplots of fractional orders are more realistic as compared to the classical order solution as given in (a). Introducing the fractional order, the solution curves exhibit more realistic and smooth transients and delayed responses in comparison with integer order solutions. Comparison of and for different values of and . Figure 3 illustrates the impact of the fractional order on the dynamic behaviour of and for fixed . It is clear from the visualisation that as the fractional order deviates from 1, the solution curves exhibit smoother and more realistic transients. The resource biomass for classical order decreases more sharply as compared to fractional orders, while the population density decreases for fractional orders. With the passage of time, both solution curves approach stable equilibrium. Visualization of and for different values of and . Figure 4 present the system dynamics under fixed fractional order and varying type parameter . As compared to Fig. 2, these figures show the same dynamics with not too much variation in the solution curves. Therefore it is clear that at this time the parameter has more influence on the solution as compared to . Visualization of and for different values of and . Figure 5 presents the solution curves at different values of and . These plots show that as the system moves away from integer order, the desire or realistic behaviour occurs and transitions toward stable, equilibrium-like dynamics over time. The solution curves clearly show that the fractional order is a key factor in suppressing the unusual dynamics and yielding stability. The second parameter is responsible for producing more gradual modification for each value of fractional order . Visualization of and for different values of and . Parametric analysis of the solution variables is provided in Figs. 6, 7. The visualization illustrates how these variables react to the two parameters and . Surface and contour representation of . Surface and contour representation of . Figure 8 presents the phase portraits in the plane. The phase portraits show that the resource-population model is asymptotically stable for the considered values of and . The decrease in the value of improves the memory effect and slows down the convergence rate towards equilibrium, while variation in parameter modifies the transient interaction patterns. The curves converge toward the interior equilibrium point, where both the solutions remain positive, showing a stable coexistence state. It shows that the chosen derivative effectively captures the hereditary dynamics in the specified system while preserving long-term stability. Phase portraits of and for different values of and . Table 1 provides a comparison between integer-order derivatives and different fractional orders. It is clear from this table and the visualisations that the final resource biomass rises from 0.48 (integer-order case ) to 1.34 (fractional-order ). Similarly, the final population density falls from 102.31 () to 94.53 (). The memory effect delays population response to resource abundance, preventing the human population from fully exploiting available resources within the finite time horizon. The maximum resource biomass increases marginally from 5.06 to 5.10, while the maximum population density decreases from 105.53 to 94.53. This inverse relationship confirms the trade-off induced by fractional memory. The time to reach maximum resource biomass remains essentially unchanged (from to ). However, the time to reach maximum population density is significantly delayed from () to (). This clearly demonstrates that fractional memory affects the response of population dynamics most profoundly. As the fractional order decreases from 1.0 to 0.5, we observe a consistent pattern: resource biomass increases, population density decreases, and the peak timing for population density is substantially delayed. Table 1Quantitative comparison of fractional vs integer orders.MetricFractional OrderFinal resource biomass0.47900.55220.66480.82281.03941.3416Final population density102.3109102.3817102.2379101.468399.330194.5282Maximum resource biomass5.05605.06605.07625.08615.09455.0999Time of0.10.10.10.10.10.1Maximum population density105.5296102.8455102.2379101.468399.330194.5282Time of2.44.220.020.020.020.0 Quantitative comparison of fractional vs integer orders. Parameter sensitivity analysis Sensitivity analysis across key parameter ranges quantifies the robustness of the fractional-order model. The results are presented in both visual and tabular formats. Figure 9 presents sensitivity heatmaps showing normalised responses to key parameter variations across fractional orders. Figure 10 compares final resource and population values across twelve management scenarios. Figure 11 displays a robustness spider plot, confirming that yields the most robust predictions. Tables 2 through 11 report the final resource biomass and population density for fractional orders . Table 3Sensitivity to resource carrying capacity (K).K500.47/102.250.51/102.260.67/101.92750.48/102.290.51/102.300.68/101.981000.48/102.310.51/102.320.68/102.011500.48/102.330.52/102.340.68/102.042000.48/102.340.52/102.350.69/102.05 Sensitivity to resource carrying capacity (K). Table 4Sensitivity to harvesting effort (E).E01.28/114.671.29/114.601.36/114.08500.91/107.630.93/107.601.03/107.231000.48/102.310.51/102.320.68/102.011500.00/100.000.14/99.980.39/99.142000.00/100.000.04/99.740.22/98.10 Sensitivity to harvesting effort (E). Table 5Sensitivity to population growth rate (s).s0.50.46/105.220.50/105.220.67/104.330.80.47/103.390.51/103.400.68/102.901.20.48/102.310.51/102.320.68/102.011.80.48/101.570.52/101.570.69/101.382.50.49/101.140.52/101.140.69/101.01 Sensitivity to population growth rate (s). Table 6Sensitivity to population carrying capacity (L).L500.97/54.371.00/54.461.16/54.97750.65/78.090.69/78.150.85/78.331000.48/102.310.51/102.320.68/102.011500.29/151.440.33/151.310.49/149.872000.20/200.970.24/200.700.39/198.04 Sensitivity to population carrying capacity (L). Table 7Sensitivity to linear benefit ().0.0000.48/101.930.52/101.910.69/101.470.0050.48/102.120.51/102.120.68/101.740.0100.48/102.310.51/102.320.68/102.010.0200.48/102.690.51/102.730.68/102.540.0500.47/103.810.50/103.920.67/104.11 Sensitivity to linear benefit (). Table 8Sensitivity to nonlinear benefit ().0.000.49/100.410.53/100.130.71/98.050.050.48/101.380.52/101.250.69/100.100.100.48/102.310.51/102.320.68/102.010.200.47/104.060.50/104.330.66/105.490.500.45/108.650.48/109.550.62/114.19 Sensitivity to nonlinear benefit (). Table 9Sensitivity to linear harvest rate ().0.0000.57/103.210.60/103.210.75/102.900.0010.53/102.740.56/102.750.72/102.440.0010.48/102.310.51/102.320.68/102.010.0020.38/101.550.43/101.570.62/101.230.0050.10/100.160.20/100.190.44/99.54 Sensitivity to linear harvest rate (). Table 10Sensitivity to nonlinear harvest rate ().0.0006.80/491.186.80/490.216.76/483.420.0050.89/107.360.92/107.551.09/108.480.0100.48/102.310.51/102.320.68/102.010.0200.25/100.710.28/100.580.43/99.380.0500.10/100.170.13/99.940.23/98.14 Sensitivity to nonlinear harvest rate (). Sensitivity heatmaps of resource and population responses to parameter variations generated using MATLAB R2024a (Software available at: https://www.mathworks.com/products/matlab.html). Resource and population final values across twelve scenarios for three fractional orders. Spider plot of robustness metrics for . Table 2Sensitivity to resource growth rate (r).r0.50.00/100.000.03/99.730.20/98.001.00.00/100.000.10/99.870.35/98.831.60.48/102.310.51/102.320.68/102.012.51.20/113.011.21/112.951.29/112.464.02.04/136.382.05/136.272.09/135.50 Sensitivity to resource growth rate (r). Table 11Sensitivity to catchability coefficient (q).q0.0001.28/114.671.29/114.601.36/114.080.0050.91/107.630.93/107.601.03/107.230.0100.48/102.310.51/102.320.68/102.010.0200.00/100.000.04/99.740.22/98.100.0500.00/100.000.01/99.700.04/97.52 Sensitivity to catchability coefficient (q). Scenario-based analysis We evaluate practical applicability by simulating twelve distinct management scenarios representing different ecological and economic conditions. Table 12 reports the final values, peak values, and peak times for each scenario across fractional orders. Table 12Scenario-based analysis results at .Scenario1: Base,,,,,,,,,2: Low,,,,,,,,,3: High,,,,,,,,,4: Low,,,,,,,,,5: High,,,,,,,,,6: Low,,,,,,,,,7: High,,,,,,,,,8: Low,,,,,,,,,9: High,,,,,,,,,10: Sustainable,,,,,,,,,11: Overexploit,,,,,,,,,12: Optimal,,,,,,,,, Scenario-based analysis results at . Robustness analysis We compute the coefficient of variation (CV) across all scenarios for each fractional order. Lower CV indicates greater robustness to parameter variations. Table 12 presents twelve management scenarios comparing fractional orders, while Table 13 presents the robustness metrics quantified by CV. Table 13CV across scenarios for different fractional orders.MetricCV()58.06%49.85%32.44%CV()27.10%27.08%27.11%CV()7.18%6.62%6.18%CV()38.03%35.46%31.62% CV across scenarios for different fractional orders. The sensitivity and scenario-based analyses reveal that fractional order provides the most robust predictions, with the lowest coefficient of variation across metrics. A critical threshold emerges at harvesting effort , beyond which resource collapse occurs for all orders. Decreasing from 1.0 to 0.6 increases final resource biomass by an average of while marginally reducing population density by , though peak timing delays substantially. Sustainable management achieves stable outcomes (–0.73, –155), whereas overexploitation drives depletion, though provides partial buffering ( vs near-zero). These results confirm that memory effects enhance model robustness, with the identified harvesting threshold of providing a clear sustainability boundary for resource managers. Neural network methodology A neural network approach is used to predict the dynamics of a fractional order resource-population model. To train the neural network, we used the data generated by a novel hybrid numerical scheme. The network architecture consists of an input layer with 1 neuron, two hidden layers each containing 15 neurons with tanh activation, and an output layer with 1 neuron. The hyperbolic tangent activation function is used throughout the hidden layers. Training employs the Levenberg-Marquardt algorithm (trainlm) with an adaptive learning rate , a maximum of 1000 epochs, a performance goal of . The data is partitioned into a training set (70%), a validation set (15%), and a test set (15%), which are respectively used for parameter optimization, generalization monitoring, and final evaluation. In our analysis the neural networks presented an excellent approximation for both and . The performance of the network is evaluated using different metrics, consist of MSE, RMSE, MAE, and the coefficient of determination (). The following Tables 14-18 provide a complete detail of neural network performance metrics, regression analysis, comparative performance of neural networks for both and . Table 15Regression analysis for prediction.ParameterValueRegression equationRegression 1.000000Number of training epochs92Final gradientFinal learning rate () Regression analysis for prediction. Table 16Performance metrics for .MetricValueMSERMSEMAEMaximum Absolute Error Score0.999998 Performance metrics for . Table 17Regression analysis for prediction.ParameterValueRegression equationRegression 1.000000Number of training epochs10Final gradientFinal learning rate () Regression analysis for prediction. Table 14Performance metrics for .MetricValueMSERMSEMAEMaximum absolute error Score0.999999 Performance metrics for . Table 18Comparative performance for and .MetricMSERMSEMAEMaximum error score0.9999990.999998Training epochs9210 Comparative performance for and . Figure 12 gives the idea about the data division for the training of the neural network and the best fit for the outputs of . Figures 13,14 present the neural network and error analysis of , respectively. Neural network analysis of . Training data distribution and model fit. Neural network analysis of . Error analysis of . Figures 15-17 present best fit for the outputs of , neural network and error analysis of respectively. Best fit for outputs. Error analysis of . Figure 18 shows a comparison of numerical and neural network solutions for and for different values of and . The figure illustrates close agreement between the numerical solutions and neural network outputs. Comparison of numerical and neural network solutions for and . Through neural networks, both and achieved exceptional accuracy in the approximation process. The solution presents higher performance with a notably lower MSE of compared to the solution with MSE . The convergence of is faster in the process of training; has 10 epochs, while has 92 epochs. It is quite evident from the regression equations that both neural networks have effectively learned the dynamics of the considered fractional model. The excellent agreement () between neural network outputs and numerical solutions confirms internal consistency, that the network successfully learned the mapping encoded by our numerical scheme, and demonstrates the approximation capability of neural networks to efficiently replicate solutions for rapid parameter evaluation. Notably, this represents consistency with the numerical scheme rather than independent validation against experimental data, which remains essential future work. Conclusion In this work, we have developed a comprehensive mathematical framework utilising advanced fractional calculus to analyse population-resource dynamics. Specifically, we have investigated a -Hilfer model that captures the logistically coupled dynamics of global population density and renewable resource biomass. The -Hilfer derivative provides enhanced flexibility for modelling this nonlinear problem through fractional order and type parameter . Our analysis has encompassed three complementary components: a qualitative study of system properties, a hybrid numerical scheme for accurate simulation, and neural network-based validation to ensure reliability. The proposed neural network approach successfully predicts system dynamics with exceptional accuracy (), enabling rapid evaluation across parameter spaces. Through this comprehensive approach, we have established the reliability and applicability of our fractional framework. This work advances theoretical understanding of fractional population-resource dynamics while delivering concrete practical implications for policy-making. Specifically, our identification of as a critical harvest threshold provides regulators with a quantifiable sustainability boundary, while Scenario 10 demonstrates that integrated policies achieve stable resource biomass (0.60–0.73.60.73) alongside healthy population levels (153–155). These actionable insights equip decision-makers with evidence-based tools for balancing human welfare with resource conservation across diverse ecological and economic contexts. Acknowledgements This research was funded by Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2026R234), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia. The author would also like to thank Prince Sultan University for their valuable support. Author contributions Sadique Ahmad: Formal analysis, Writing – review & editing; Israr Ahmad: Writing-original draft, Investigation, Methodology, Supervision; Ala Saleh Alluhaidan: Methodology, Writing-review & editing; Mohammed A. ELAffendi: Formal analysis, Writing-review & editing. Funding This research was supported by Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2026R234), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia. Data availability All data is included in the paper. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Israr Ahmad, Email: israrahmad@jc.edu.pk. Ala Saleh Alluhaidan, Email: ASALluhaidan@pnu.edu.sa. References Associated Data Data Availability Statement All data is included in the paper.