Interpretable machine-learning enhanced parametrization methodology for Pluronics–water mixtures in DPD simulations

Abstract

Dissipative particle dynamics (DPD) is an incredibly powerful tool for simulating the behavior of structured fluids. However, identifying the appropriate model parameters to accurately replicate physical properties remains a challenge. This study showcases the benefits of integrating machine learning techniques into the top-down parameterization of Pluronic systems. The proposed workflow outlines a data-driven approach to accurately determine model parameters tailored to various Pluronic systems. Gaussian process regression (GPR)-based surrogate models effectively replicate the results of DPD simulations, delivering faster responses that streamline parameter optimization and enable the calibration of Pluronic systems against experimental data. Although DPD simulations provide valuable insight, their high computational cost, due to extensive simulations and post-processing, presents a challenge. The GPR-based surrogate model addresses this by modeling the relationships between input parameters and output properties. SHAP (SHapley additive exPlanations) analysis enhances model interpretability, providing deeper insights into the relationships and causal mechanisms between the input parameters and the predicted properties. The combination of GPR and SHAP analysis provides an interpretable machine learning approach, enabling a more efficient optimization process and reducing the need for exhaustive simulations. This work lays a foundation for generalizing the parameterization process across Pluronic systems and conditions, such as varying temperatures, by incorporating additional DPD model input parameters.