Automated Construction of Data-Efficient Machine-Learned Force Fields for Polymers

General-purpose machine-learned force fields (GP-MLFFs) provide near-first-principles accuracy across broad chemical spaces, but their reliability is limited by training coverage and reference level of quantum mechanics. We present an automated active-learning workflow using the open-source package aims-PAX that builds on GP-MLFF sampling to generate targeted, higher-accuracy reference data for fine-tuning or training models with task-specific architectures. We demonstrate this approach by training a single model on approximately 4,000 hybrid-DFT structures spanning seven chemically distinct polymers. The model predicts densities and glass-transition temperatures with accuracy comparable to GP-MLFFs trained on hundreds to thousands of times more data, supports stable molecular dynamics of approximately 10,000 atoms, and transfers to unseen polymers. An explicit dispersion term is indispensable in this low-data regime, providing the inter-chain cohesion that GP-MLFFs must otherwise learn implicitly from large datasets. Replay fine-tuning extends the model to silicon-containing polydimethylsiloxane while preserving accuracy on the original polymers. Beyond collective properties, the aims-PAX model better describes the local response to hydrogen bonding. The model reproduces experimentally observed frequency shifts of the nylon-6 amide group, which a polymer-specific GP-MLFF captures only partially, and their cooperative enhancement along hydrogen-bonded chains, which a classical force field cannot represent. These results establish aims-PAX as a data-efficient route from general-purpose sampling to specialized MLFFs, combining targeted chemical coverage, suitable architectures, and the reference accuracy the problem requires.