The Open Polymers 2026 (OPoly26) Dataset and Evaluations

TL;DR

This work introduces Open Polymers 2026 (OPoly26), the largest open-source DFT dataset dedicated to polymers to date, aimed at enabling universal atomistic models for polymer systems. OPoly26 provides over 6.57 million single-point DFT calculations on fragments up to 360 atoms, spanning 2,444 monomers and totaling more than 1.2 billion atoms, and covers diverse chemistries and architectures generated via MD, MLIP-MD, AFIR, and DFTB workflows. The dataset is complemented by an equivalent DFT workflow to OMol25 to support cross-domain generalization, and baseline experiments show that adding polymer-specific data improves polymer energy predictions without degrading performance on other chemical domains. By releasing OPoly26 under CC-BY-4.0 and providing accompanying code, the authors aim to accelerate open, universal atomistic modeling for polymers and related materials, with applications in polymer dynamics, degradation, and device-relevant processes.

Abstract

Polymers-macromolecular systems composed of repeating chemical units-constitute the molecular foundation of living organisms, while their synthetic counterparts drive transformative advances across medicine, consumer products, and energy technologies. While machine learning (ML) models have been trained on millions of quantum chemical atomistic simulations for materials and/or small molecular structures to enable efficient, accurate, and transferable predictions of chemical properties, polymers have largely not been included in prior datasets due to the computational expense of high quality electronic structure calculations on representative polymeric structures. Here, we address this shortcoming with the creation of the Open Polymers 2026 (OPoly26) dataset, which contains more than 6.57 million density functional theory (DFT) calculations on up to 360 atom clusters derived from polymeric systems, comprising over 1.2 billion total atoms. OPoly26 captures the chemical diversity that makes polymers intrinsically tunable and versatile materials, encompassing variations in monomer composition, degree of polymerization, chain architectures, and solvation environments. We show that augmenting ML model training with the OPoly26 dataset improves model performance for polymer prediction tasks. We also publicly release the OPoly26 dataset to help further the development of ML models for polymers, and more broadly, strive towards universal atomistic models.

Figures (6)

• Figure 1: Overview of . We generate diverse polymeric structures that cover a broad spectrum of polymer chemistries including homopolymers and highly diverse copolymers, as well as interactions between polymers and solvent molecules and ionic species. We then extract >6M substructures with <360 atoms, capping dangling bonds with hydrogens as necessary, and simulate each with DFT. We train MLIP models on both the and OMol25 datasets, which employed identical DFT settings, to ensure model generalization across molecular chemistry domains, unlocking improved simulation capabilities in a range of high-impact applications.
• Figure 2: Summary of dataset statistics. a) Distribution of DFT calculations by polymer categories. b) Distribution of force norms per atom across all DFT calculations (left) and distribution of the number of atoms in each DFT calculation in comparison to OMol25 (right). c) Comparison of the number of unique monomers across polymer datasets. d) Total number of atoms contained in polymer DFT datasets meant for MLIP training including , data from SimPolysimm_simpoly_2025, and the work of Matsumura et al.matsumura_generator_2025. **As SimPoly is not yet publicly available, we estimate the total number of atoms by assuming the highest possible number of atoms in each DFT calculation.
• Figure 3: Overview of the data generation process. Starting from a diverse collection of polymeric compositions (Section \ref{['ssec:poly_compositions']}), we apply a sequence of computational steps tailored to each class of materials. In general, we first use classical MD simulations to generate a broad distribution of polymer configurations (Section \ref{['ssec:simulating_dynamics']}). From these trajectories, we sample both the relaxed final frame (for use in further calculations) as well as sample diverse frames from across the MD trajectory, as described in Section \ref{['sec:extracting_substructures']}. All resulting structures are subsequently passed into a final DFT calculation, yielding the MLIP training and test data.
• Figure 4: Overview of the polymer families present in the Traditional Polymer compositions. Polymer families are determined by defining a SMARTS substructure pattern for each polymer family and checking if each Traditional Polymer composition matches this pattern.
• Figure 5: Illustrative examples of the fluoropolymer compositions present in , depicting polymers with a wide range of fluorine content.