Interpretable surrogate modeling for polymers: A SHAP-guided framework for descriptor-level analysis

Predicting how molecular structure and thermodynamic conditions shape polymer properties is challenging because simulations generate high-dimensional data, and the underlying couplings between conformation, packing, and relaxation are highly nonlinear. We develop a methodology that integrates molecular dynamics simulations with explainable machine-learning surrogates to analyze polymer melts across structural, thermodynamic, and viscoelastic regimes. Simulation-derived descriptors are incorporated into a surrogate representation of macroscopic responses using ensemble regression models, and SHapley Additive exPlanations (SHAP) are used to quantify the contribution of each descriptor within this representation. This combined MD-XAI framework reproduces established trends in polymer physics while organizing them into a clear hierarchy of controlling variables, enabling a mechanistic interpretation of glass transition behavior, density-mobility coupling, and pressure-dependent relaxation. The aim of this work is to employ surrogate modeling as an interpretive framework to organize and attribute the relative influence of physically coupled simulation descriptors on polymer response. Rather than relying solely on curve inspection or statistical correlations, the approach provides a systematic path to interrogate MD datasets and identify the structural or thermobaric factors most relevant to a given property. The workflow is general in formulation and demonstrated here for a representative polymer system (PTMO), providing a pathway for incorporating interpretability into polymer simulation pipelines. It enables physically grounded interrogation of surrogate models and can be extended to related data-driven analyses, including coarse-grained modeling and design-oriented studies. • Molecular dynamics is combined with SHAP to interpret polymer relaxation behavior. • Key thermodynamic and structural features governing Tg are quantitatively ranked. • Stress-relaxation parameters are linked to pressure via explainable ML surrogates. • The approach reveals physically interpretable drivers beyond empirical correlations. • A transparent MD-XAI workflow is demonstrated for polymer property analysis.