SimPoly: Simulation of polymers with machine learning force fields derived from first principles

Abstract

Polymers are a versatile class of materials with widespread industrial applications. Advanced computational tools could revolutionize their design, but the complex, multi-scale nature of polymers poses significant modeling challenges. Classical force fields often lack the accuracy and transferability that polymers demand, while quantum-chemical methods are prohibitively expensive for the large systems and long timescales involved. Here we develop a fast, scalable machine learning force field and show that, trained on quantum-chemical calculations alone, without any experimental data, it predicts measured bulk properties across a broad range of polymers. It reproduces polymer densities as accurately as, or better than, established classical force fields and captures the glass transition signature in the density-temperature curve, enabling the prediction of glass transition temperatures. To accelerate progress in this domain, we release a benchmark of experimental bulk properties for 130 polymers and an accompanying quantum-chemical dataset, laying the foundation for a fully in silico design pipeline for next-generation polymeric materials. The authors develop a machine learning force field trained only on quantum-chemical data that can reproduce measured key polymer properties such as densities and glass transition temperatures.