Code and data for "Polymer Property Prediction via an Automated Molecular Dynamics Pipeline and Transfer Learning"

Abstract

Code and data accompanying the manuscript Polymer Property Prediction via an Automated Molecular Dynamics Pipeline and Transfer Learning (J. N. Law, D. Lazarenko, T. Bernat, B. C. Knott, M. R. Shirts). The study builds short oligomers (trimers) directly from monomer SMILES and a polymerization mechanism, parameterizes them with OpenFF, runs high-throughput LAMMPS molecular dynamics, extracts thermophysical and structural properties, and uses those properties to pre-train a graph neural network that is then fine-tuned on experimental polymer properties (transfer learning). This archive is a snapshot of the paper repository with its git submodules expanded in place, so it is complete without network access: external/polymer_build_workflow: monomer SMILES to packed, parameterized simulation boxes (polymerist, OpenFF, Signac). external/polymer_MD_LAMMPS: LAMMPS input decks, run scripts, and property extraction; per-system MD property tables for trimers and pentamers. external/ml_transfer_learning: training data (1,820 trimer MD property vectors and 1,507 publicly available polymer property records), pre-training, transfer-learning and multi-task training notebooks, and the code for every ML figure. external/polyID_md_transfer_learning: the md_transfer_learning branch of PolyID used to train the models. data/ and figures/: paper-level data and the figure-to-code map. Start with README.md. ARCHIVE_CONTENTS.md lists the source repository and exact commit of every component. Individual components retain their own LICENSE files.