PolyCG-Base: A Foundation Model for Universal, State-Aware Coarse-Graining of Linear Polymers

Khartik Uppalapati, Bora Yimenicioglu, Shakeel Abdulkareem

Keywords: Coarse Graining, Polymer Simulations, Foundation Models

TL;DR: PolyCG-Base is a conditional foundation model for polymers that learns CG mappings, interactions, and friction. Using equivariant encoders and physics-based refinement, it achieves transferable accuracy across chemistries.

Abstract: Linear repeat-unit polymers (polyolefins, styrenics, (meth)acrylates, vinyls, polyesters, polyamides) share saturated backbones with side-group--controlled physics, yet coarse-grained (CG) models are typically re-derived per chemistry and thermodynamic state, hindering transfer and scale-up. We introduce PolyCG-Base, a conditional foundation model that amortizes CG design-mapping, bonded and nonbonded interactions, and friction-from a polymer specification (BigSMILES) and state variables $(T,P,$ composition, tacticity, $M_w)$. The encoder uses E(3)-equivariant message passing to learn chemically typed embeddings from atomistic/united-atom fragments and melts. Parameters are initialized by multiscale coarse-graining/force matching and refined by relative-entropy minimization to match reference ensembles, implemented with standard coarse-graining toolchains. Dynamical consistency is imposed via Green-Kubo constraints (Langevin/GLE), ensuring fluctuation-dissipation compliance. Validation on held-out homopolymers and random copolymers targets conservative, literature-aligned accuracy: $\leq 10\%$ error in $g(r)$, $S(q)$, and density; $\leq 20\%$ in self-diffusion and zero-shear viscosity after standard time rescaling; and correct trend-level glass-transition series against open experimental corpora (PoLyInfo, ThermoML). All inputs derive from public simulations and databases (e.g., RadonPy-automated AA/UA MD), keeping the study IRB-exempt. By coupling state-aware representation learning with physics-based optimization, PolyCG-Base provides a transferable CG prior for commodity thermoplastics that unifies polymer informatics with statistical-mechanics model reduction.

Submission Track: Paper Track (Full Paper)

Submission Category: AI-Guided Design + Automated Material Characterization

Institution Location: {Oakton, United States}

AI4Mat Journal Track: Yes

Submission Number: 143