Schrödinger's AlphaEvolve Shreds Ewald Summation for 4x Speedup

4x speedup in training and inference of machine-learned force fields — that’s what Schrödinger got by letting an evolutionary AI agent rewrite a couple of critical loops in their PyTorch code.

Computational chemists have long lived with a trade-off: fast classical force fields that are inaccurate, or slow quantum-mechanical methods that are precise. MLFFs aim for both by training neural nets on high-fidelity quantum data. But the algorithms that compute atomic neighbor lists and long-range potentials turned into bottlenecks. Schrödinger’s team, led by Gabriel Marques, took their three-decade-old codebase and pointed AlphaEvolve at the Ewald summation — the single worst performer.

For-Loops Were the Bottleneck

The Ewald sum had no established vectorized algorithm in PyTorch. Schrödinger’s implementation relied on simple for-loops that stumbled on large simulations. AlphaEvolve, a coding agent from Google DeepMind, iteratively generates and refines algorithms by searching the code path space. After targeting the Ewald sum, the system produced a batched implementation using parallel batch matrix multiplication — effectively replacing serial iterations with a fused tensor operation.

What the Numbers Actually Mean

Schrödinger measured three things: inverse time (throughput), functional correctness, and success rate (faster + correct). Baseline performance score was 7.9. After evolution, that number hit nearly 30. The success rate jumped from less than 1% (40 out of 5,000 evaluations) to over 60%. That’s not a fluke; the evolved code passed every regression test, including disordered water models. Marques says faster MLFF inference “carries real business impact, shortening R&D cycles in drug discovery, catalyst design, and materials development, and enabling companies to screen molecular candidates in days rather than months.”

From For-Loops to Batched Matrices

The 4x speedup isn’t just a benchmark number — it directly maps to screening larger chemical libraries in less time. Drug discovery teams can evaluate 100,000 candidates in a week instead of a month. Catalyst design pipelines can iterate on reaction simulations without waiting for weekend jobs. Materials researchers can run ensemble calculations that were previously infeasible.

Schrödinger plans to push this further by targeting custom GPU kernels, testing whether AI-generated code can beat hand-tuned implementations.

Source: How Schrödinger sped up molecular discovery by 4x with Alphaevolve
Domain: cloud.google.com