## TL;DR
AI is accelerating materials science from Edisonian trial-and-error to systematic discovery. DeepMind's GNoME used graph neural networks to discover 2.2 million new crystals — 45x the human-accumulated catalog — while autonomous labs synthesize AI-predicted compounds robotically.

## Core Explanation
Traditional materials discovery: hypothesize composition → synthesize manually (days to months) → characterize structure → test properties → iterate. Bottleneck: the space of possible inorganic crystals exceeds 10^12, only ~48,000 experimentally known structures in the ICSD database (as of 2023). AI approach: (1) Train GNNs on known crystal structures to predict formation energy (stability proxy) from composition and structure; (2) Active learning — model proposes candidate compositions, DFT (Density Functional Theory) validates most promising, validated results feed back into training; (3) GNoME pipeline generated 2.2M candidates, filtered to 381K stable below the convex hull; (4) Autonomous lab (A-Lab) synthesizes top candidates robotically with machine learning-guided process optimization.

## Detailed Analysis
GNoME architecture: message-passing GNN on crystal graphs — nodes = atoms, edges = bonds, messages aggregate 128-hop neighborhood information. Stability defined via energy above convex hull (Ehull): Ehull = 0 means thermodynamically stable, >0 means metastable or unstable. Active learning loop discovered structures near the convex hull boundary most efficiently. Materials applications: (1) Battery materials — Li-ion conductors, solid electrolytes; (2) Photovoltaics — novel perovskite compositions; (3) Superconductors — predicted hydride superconductors at high pressure; (4) Catalysis — CO2 reduction, ammonia synthesis. 2025 AI materials informatics survey (EPJ) highlights the shift from "discovery" to "inverse design" — specifying desired properties and having AI propose materials that satisfy them. Key challenge: bridging the gap from DFT prediction to experimentally realizable synthesis (many AI-predicted structures cannot yet be made).

## Further Reading
- Materials Project (DOE/UC Berkeley) — open materials database
- GNoME GitHub: google-deepmind/materials_discovery
- OQMD: Open Quantum Materials Database