ConfDrift: One-Step Conformer Generation via Boltzmann-Weighted Drifting

Identifying ground-state molecular conformations is fundamental to chemistry but physics-based conformational search remains computationally expensive at scale. Recent advances in generative flow-based models improve both efficiency and accuracy but still rely on iterative sampling at inference and do not account for the Boltzmann distribution that governs the molecular energy landscape. We introduce ConfDrift, a distillation-free generative model that predicts a conformer from a molecular graph in one-step inference. On ensemble coverage of GEOM benchmarks, ConfDrift achieves comparable performance to the state-of-the-art models. Tilting the attractive component of the drift by per-conformer Boltzmann weights moves the equilibrium of the training objective, so calibration is learned rather than imposed afterwards: on GEOM-QM9 this raises the median ground-state basin population from 0.33 to 0.44 against a reference of 0.72, with no energy model and no per-sample energy evaluation at inference.