Thomas O'Connor and his team at Carnegie Mellon University just cracked a puzzle that has stumped materials scientists for over 50 years: how to simulate polymers faithfully at scales that matter for real engineering. The breakthrough feels almost simple in hindsight—they taught artificial intelligence the laws of thermodynamics before asking it to learn anything else.
For decades, the problem has been one of scale and compromise. A single polymer chain contains tens of thousands of atoms; a practical material contains billions. Simulating all those atoms is computationally impossible, so researchers have always resorted to coarse-graining: replacing clusters of atoms with simpler particles to make the math tractable. But this shortcut came at a cost. Conventional coarse-grained models could capture either the structure of a polymer or its dynamics—rarely both—and they consistently failed to predict the entropic and viscous forces that actually govern how polymers flow, relax, and dissipate energy. Those are the forces that determine whether an adhesive will hold, how a biopolymer film stretches, whether a self-assembling material locks into the right nanostructure. Traditional machine-learning approaches, despite their flexibility, tended to break these fundamental physics rules too.
The solution came from marrying two different fields of expertise. O'Connor's team, collaborating with researchers at the University of Pennsylvania, built their AI framework around the metriplectic bracket—a mathematical structure developed over decades within the non-equilibrium thermodynamics and soft-materials communities. By translating this structure into a neural network skeleton, they created an architecture that is mathematically incapable of violating energy conservation or the Second Law of Thermodynamics. The physics wasn't bolted on as an afterthought or a constraint; it was the skeleton itself. "The long-standing problem with coarse-graining polymers is that the information you throw away is exactly the information that controls how the material flows during processing or fails in applications," O'Connor explained. "By building the laws of thermodynamics into the architecture itself, we get models that recover those missing entropic and viscous forces, not as an approximation but by construction."
The team added another clever innovation: a self-supervised learning strategy that lets the network discover hidden variables—like entropy and internal microstructure—on its own, simply by watching how particles move. This means the model can be trained directly from experimental video or trajectories, not just from expensive atomistic simulations.
When they tested the framework on notoriously difficult cases, it delivered. For star polymers—branched molecules whose dynamics have thwarted previous coarse-graining attempts—the framework recovered both the structural details and non-equilibrium dynamics at aggressive levels of simplification, where state-of-the-art graph neural networks failed. For a dense colloidal suspension under oscillatory shear, the model learned directly from high-speed video and captured the rare, localized rearrangement events that drive how the material actually flows.
Recognizing that tools mean nothing without access, O'Connor's team released open-source implementations in PyTorch and LAMMPS—the standard molecular-dynamics engine across academia, national labs, and industry. The LAMMPS version has already been tested at scales of millions of coarse-grained particles, ready to enable novel research in polymer adhesion, self-assembly, self-healing, and fracture mechanics. After more than half a century of workarounds, materials scientists finally have a principled path forward.