Discovered Materials unveils AI benchmark for novel crystals — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Discovered Materials describes a benchmark tasking AI models with finding novel, BEOL-compatible crystalline materials meeting targets for thermal conductivity (κ), static dielectric constant (ε), Young's modulus, and shear modulus — with each candidate requiring a synthesis recipe an expert reviewer would actually attempt.
- Models are equipped with web search via Exa, a Python/bash coding sandbox with pymatgen, mp_api, and ASE, plus ML-based tools for stability and property prediction, and run with a 100-million-token budget through the UK AI Security Institute's Inspect framework.
- Property calculations rely on the PET-MAD universal machine learning interatomic potential (Ceriotti et al., Nature Communications 2025), with Phonopy/Pheasy handling phonon physics and GMTNet fitting dielectric tensors, in place of direct density functional theory.
- Synthesis grading uses a 'worst of three' GPT-5.6 Sol with OpenAI web search as the rubric grader — calibrated against human expert feedback — with critical penalties automatically disqualifying a recipe and fixable penalties left to a judge's discretion.
- In an example, Claude Opus 5's proposed microwave-plasma CVD recipe for hexagonal diamond (lonsdaleite) was graded 'WOULD NOT ATTEMPT' on a single critical penalty: detonation nanodiamond seeds are cubic, screening out the proposed hexagonal buffer, with no demonstrated ABAB stacking mechanism.
- The benchmark's authors flag future work to incorporate direct density functional theory calculations — either replacing MLIPs or as a hybrid approach.
Why it matters: Materials scientists gain a standardized harness for scoring AI-proposed crystals across thermal, dielectric, and mechanical targets, with a critical-penalty rubric that auto-rejects recipes like Claude Opus 5's lonsdaleite attempt — a harder, multi-tool test than typical LLM evals and a credibility filter for AI-generated synthesis proposals.
Ask SkimNews
