Predictive Feature Caching for Training-free Acceleration of Molecular Geometry Generation

The article presents a training-free caching strategy for accelerating molecular geometry generation using flow matching models, which typically face high inference costs due to extensive network evaluations. This method predicts intermediate hidden states during solver steps and is compatible with SE(3)-equivariant backbones and pretrained models. Experiments on the GEOM-Drugs dataset show that this caching approach can halve wall-clock inference time while maintaining sample quality, and when combined with other optimizations, can achieve up to a 7x speedup, making it significant for practitioners seeking efficient molecular sampling solutions.