Energy-based models learn realistic molecular physics directly

Molecules in equilibrium follow a Boltzmann distribution defined by their energy landscape. EBMol, an energy-based model, learns this landscape directly using a Restoring Field Matching objective without expensive molecular simulation during training. The model learns an atom-additive scalar potential and samples via Mirror-Langevin dynamics combined with parallel tempering. It matches or exceeds diffusion and flow-matching approaches on standard benchmarks while providing a physically grounded quality metric for molecular configurations. The learned energy landscape enables controllable generation through potential composition and zero-shot linker design without model retraining.