Active learning-enabled multi-objective design of thermally conductive and mechanically compliant polymers

Yuhan Liu, Jiaxin Xu, Renzheng Zhang, Meng Jiang, Tengfei Luo

arXiv:2603.23494·cond-mat.mtrl-sci·Published 2026-03-24

Polymers are attractive in applications like flexible electronics and thermal interface materials due to their mechanical compliance and processability. However, conventional polymers have low thermal conductivity (TC), limiting their heat dissipation performance. Identifying polymers that simultaneously achieve high intrinsic TC and mechanical flexibility (i.e., low modulus) remains a challenge. Here, we develop an active learning (AL) framework based on multi-objective Bayesian optimization (MOBO) to discover polymers exhibiting both high TC and low bulk modulus. Initially, a high-throughput molecular dynamics (MD) pipeline generated an initial dataset, and independent Deep Kernel Learning (DKL) surrogate models were trained for TC and bulk modulus to predict properties and uncertainties. Using the parallel noisy expected hypervolume improvement (qNEHVI) acquisition function, the framework iteratively screens a larger unlabeled polymer database, systematically recommends new polymer candidates for MD evaluation, and updates the DKL models with newly acquired data. Ultimately, six candidates were identified on the Pareto front, representing optimal trade-offs between TC and modulus. Interpretability analysis further revealed molecular features associated with these trade-offs, and synthesizability assessment supported the practical relevance of the selected candidates. By combining MD simulations with AL-enabled MOBO, our workflow mitigates data scarcity, reduces development time, and provides actionable guidance for designing multifunctional polymers tailored for different applications.

TopicsQuantum Chemistry & Force Fields

Tagsactive-learning bayesian-optimization molecular-dynamics thermal-properties

arXiv categoriescond-mat.mtrl-sci

arXiv abstract pagePDF