Part VI · 2 — Physics and Chemistry

Where AI replaces costly simulators and discovers materials and molecules — accelerating cycles that used to take years.

🎨 Figure F-VI.2 — From equation to material. Brief: on the left, a weather map being forecast by a network (GraphCast); on the right, a crystalline/molecular structure emerging from diffusion. Compendium palette.

2.1 Physics — AI as a simulator

• Climate/weather: GraphCast (DeepMind), PanGu-Weather, FourCastNet — faster

  forecasts, competitive with numerical simulators.

• PINNs (Physics-Informed Neural Networks) and Fourier operators — solve

  PDEs by learning the physics.

• Materials: GNoME (DeepMind, massive discovery of stable crystals), MACE,

  MatterGen (material generation by diffusion).

2.2 Chemistry — molecules

• Representations/models: ChemBERTa, MolFormer.
• Molecular generation: EDM / GeoDiff (diffusion over 3D structures).
• Connects to GNNs (ch. 11) — molecules are graphs —

  and to diffusion (ch. 02).

Pattern: structures (graphs/3D) + physical verification (simulation) = reward to generate useful candidates. The same "data + verification" recipe from Part V.