How Schrödinger sped up molecular discovery by 4x with AlphaEvolve

Computational chemistry has long faced a trade-off between fast classical force fields and accurate but slow quantum-mechanical methods. Machine-learned force fields (MLFFs) bridge this gap, but processing massive chemical libraries still demands more speed. Schrödinger identified two critical algorithms—neighbor list computation and Ewald summation—as performance bottlenecks in their PyTorch MLFF training pipeline. The Ewald sum, in particular, had no vectorized implementation and relied on simple for-loops, severely limiting throughput.

Schrödinger partnered with Google Cloud to deploy AlphaEvolve, an evolutionary AI coding agent from Google DeepMind. The agent iteratively generated and refined PyTorch code, replacing the slow for-loops in the Ewald summation with parallel batch matrix multiplication. Schrödinger used a rigorous evaluation framework measuring inverse time, functional correctness (passing regression tests on complex systems like disordered water), and success rate. The evolved code raised the program success rate from under 1% (40 of 5,000 evaluations) to over 60%, while the performance metric improved from a baseline of 7.9 to nearly 30.

The result was a 4x speedup in both MLFF training and inference, compressing molecular screening timelines for drug discovery, catalyst design, and materials development. Schrödinger plans to apply this evolutionary approach to custom GPU kernels, testing whether AI-generated code can outperform human-engineered implementations. For builders, the takeaway is clear: evolutionary AI agents can systematically optimize scientific algorithms, turning a previously intractable bottleneck into a source of competitive advantage.