Scaling Discretized Boltzmann Generators towards Solvation Free Energy Estimation

Reliable estimates of molecular thermodynamics require sampling the equilibrium conformational ensemble, but this remains a bottleneck for molecular simulations. Boltzmann Generators (BGs) address this by utilizing generative models with tractable likelihoods to enable importance sampling and free-energy estimation given a target energy. Recently, Rehman et al. proposed autoregressive BGs modeling discretized molecular coordinates as sequences, opening a path towards scalable sequence-model architectures, but introducing new tradeoffs between spatial resolution, model capacity, and sampling cost. In this work, we introduce Discretized Boltzmann Generators (DBG), a simple, flexible, and modular framework for extending transferable Boltzmann generation to larger and more chemically diverse molecular systems. We characterize how DBG performance depends on model size, training data, compute, molecular system size, and discretization resolution up to 1.4B parameters. We identify an empirical data, compute, and system-size scaling relationship for maintaining energetic accuracy. DBG achieves state-of-the-art performance on Many Peptides (up to 157 atoms) and demonstrates transferable generation for small molecules and explicit microsolvation tasks involving systems with over 280 atoms. These results provide practical guidance for scaling molecular samplers toward larger and more chemically diverse thermodynamic systems.