Discovered Materials benchmarks AI agents on inventing novel chip materials

Discovered Materials (YC P26) built a benchmark that tests whether frontier LLMs can act as autonomous computational materials scientists. Agents are tasked with proposing dynamically stable, genuinely novel, BEOL-compatible crystalline materials that hit specific targets for thermal conductivity, static dielectric constant, Young’s modulus, and shear modulus—and, critically, each candidate must include a synthesis recipe realistic enough that an expert reviewer would actually attempt it. Novelty is defined strictly: the material must never have been deposited as a thin film under BEOL-compatible conditions in the published literature.

The agents get a working scientist’s toolkit—Exa web search, a Python/bash sandbox with pymatgen, mp_api, and ASE, plus machine-learning surrogates for the hard physics. Rather than run expensive DFT, the harness leans on ML interatomic potentials (the PET-MAD foundation model) with Pheasy and Phonopy for phonon-based dynamic stability and lattice thermal conductivity, and a GMTNet tensor network trained on JARVIS DFPT data for the dielectric constant. Runs are open-ended, stopping only on an error or a 100-million-token budget, and the whole thing is wired through the AI Security Institute’s open-source Inspect framework. Synthesis recipes are scored against expert-calibrated rubrics with critical and fixable penalties, using a worst-of-three GPT-5.6 panel as the automated judge because it tracked human graders best.

The worked example shows how demanding the bar is: Claude Opus 5 proposed hexagonal diamond (lonsdaleite) grown by seeded microwave-plasma CVD, and the grader returned WOULD NOT ATTEMPT. A single critical penalty sank it—the recipe offered no credible route to the ordered P6₃/mmc phase, since growth would nucleate on predominantly cubic nanodiamond seeds rather than the buried buffer meant to template hexagonal stacking. The significance is less about any one material than the methodology: a physics-grounded, expert-anchored eval that measures whether AI agents can produce not just plausible-sounding candidates but ones a real lab would bother synthesizing.