Integrating machine learning and molecular dynamics for accelerated discovery of polymers with high thermal conductivity

Abstract

The vast chemical diversity of polymers, compounded with sparse reliable characterization data, fundamentally constrains machine learning (ML)-driven exploration of advanced polymeric materials. To overcome this, we establish an integrated computational framework combining a deep neural network (DNN), Bayesian optimization (BO), and molecular dynamics (MD) simulations for the targeted design of backbone polymers with high thermal conductivity (TC ≥ 0.40 W m−1 K−1). Our workflow begins with a data set of 501 backbone polymers, whose thermal conductivities (TCs) are labeled by MD simulations. We then established a mapping between their force-field-inspired descriptors and TCs using a DNN. While sparse high-TC data limited the DNN's predictive accuracy for TC ≥ 0.40 W m−1 K−1, we leveraged it to screen 2209 high-potential candidates from a 14 706-polymer virtual library generated by small molecules according to chemical reaction rules. Subsequent closed-loop BO-MD validation enabled efficient navigation of this subspace, and 11 synthesizable polymers with high TC were identified in 500 evaluations. Crucially, mechanistic analysis reveals that intra-chain interactions dominate thermal transport, with intra-chain contributions of 79.1%–87.5% on TC for the first six polymers. This work delivers a robust paradigm for ML-driven material discovery under data constraints.