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

Researchers propose a training-free caching strategy that accelerates molecular geometry generation in flow matching models by predicting intermediate hidden states, achieving 2-7x speedups without quality degradation. The method is compatible with pretrained models and compounds with existing optimizations, addressing a critical inference bottleneck in computational chemistry workflows.

Molecular geometry generation using flow matching models has become computationally prohibitive in practical applications, requiring hundreds of network evaluations per inference pass. This new caching approach tackles the core bottleneck by predicting and reusing intermediate hidden states across solver steps, eliminating redundant calculations without requiring model retraining. The technique operates on SE(3)-equivariant architectures, making it broadly compatible with existing molecular generation systems.

The work emerges from the broader trend of optimizing expensive generative models for production deployment. As molecular simulation becomes increasingly valuable in drug discovery and materials science, inference efficiency directly impacts research velocity and cost. Prior solutions have focused on training-based acceleration or system-level optimizations, leaving room for orthogonal improvements.

The experimental results on GEOM-Drugs demonstrate practical value: achieving 2x speedup at matched quality and up to 3x with minimal degradation represents substantial gains for computational chemistry pipelines. More significantly, the compounding effect with other optimizations reaching 7x speedup suggests the method addresses fundamental inefficiencies rather than marginal improvements. This matters for practitioners running high-throughput molecular candidate screening, where inference cost directly scales with research budgets.

Future adoption depends on whether the approach generalizes beyond drug-like molecules to larger chemical spaces and different model architectures. Integration into standard molecular generation pipelines could become standard practice, particularly as flow matching models see increased adoption in industry applications.

• →Training-free caching reduces molecular geometry inference time by 2x at matched quality and up to 7x combined with other optimizations.
• →The method predicts intermediate hidden states across solver steps, eliminating computational redundancy without model retraining.
• →Technique is compatible with pretrained SE(3)-equivariant models and orthogonal to existing accelerations, enabling widespread adoption.
• →Compounds effectively with other general optimizations, suggesting it addresses fundamental rather than marginal inefficiencies.
• →Reduces inference bottleneck critical for high-throughput molecular screening in drug discovery and materials science applications.