AlphaEvolve agent delivers 4x speedup for Schrödinger molecular simulations

Full analysis

Schrödinger, a computational chemistry software company, partnered with Google Cloud to apply AlphaEvolve — Google DeepMind's evolutionary AI coding agent — to a key bottleneck in their machine-learned force field (MLFF) pipeline. MLFFs are neural networks trained on quantum data that let researchers simulate molecular interactions faster than pure quantum methods; a critical tool in drug discovery and materials design. AlphaEvolve iteratively generated and refined PyTorch code for the Ewald summation algorithm (a mathematically intensive calculation of long-range atomic interactions), replacing slow for-loops with parallel batch matrix multiplication. The result was a 4x speedup in both MLFF training and inference, and the share of evolved programs that were both correct and faster than baseline jumped from under 1% to over 60%.

The case study is notable for the agent stack: AlphaEvolve acts as an autonomous code-optimization agent that proposes, evaluates, and refines implementations against a multi-layered test suite — functional correctness, regression tests, and throughput — without human-written algorithmic solutions. Schrödinger's technical lead noted the acceleration compresses molecular screening from months to days. The company now plans to extend the approach to custom GPU kernels to see whether AI-generated code can beat hand-tuned human implementations.