Abstract for: Incorporating Deep Learning into System Dynamics: Amortized Bayesian Inference for Likelihood-free Parameter Estimation

In system dynamics modeling, the integration of empirical data with theoretical models remains a key challenge, particularly when it comes to the estimation of model parameters and their uncertainties. Addressing this challenge necessitates advanced estimation methods that effectively incorporate qualitative and quantitative data, enabling parameter estimation and uncertainty quantification. This study explores the applicability of Amortized Bayesian Inference (ABI) methods, particularly focusing on Neural Posterior Estimation (NPE), to bridge this gap. Utilizing the BayesFlow Python package, we demonstrate the viability, scalability, and user-friendliness of these neural estimation techniques through two example models: a simple Random Walk model to systematically explore process and measurement noise, and a moderately complex SEIRb model, which integrates behavioral feedback loops into classical epidemiological dynamics and exhibits multiple reference modes. Our analysis highlights the potential of ABI to facilitate a principled, scalable, and likelihood-free inference workflow. We provide the accompanying code to streamline the application of these methods beyond our examples. This research underlines the opportunities for leveraging cutting-edge machine learning techniques to enhance the fidelity and applicability of system dynamics models, promising a more integrated and responsive approach to data-driven modeling for complex systems.