SaddleFlow Al(100) adatom model
SaddleFlow generator for the Al(100) surface with an adatom (385 atoms, 128 fixed): given the atoms that should react, it starts from the minimum plus N(0, 0.3²) Å noise on those atoms, flags them, and integrates the learned velocity with 20 forward-Euler steps (one pass) to a structure close to a saddle point. Backbone UMA-S-1.2 with time-FiLM at all four blocks and a per-atom flag input, depth-3 equivariant velocity head; trained 100 epochs on the 490 saddle points of this surface (reactant → saddle). On those same 490 training saddles, with each saddle's own moving atoms (atoms moving more than 0.5 Å, plus the most-moving one), at least one of 10 draws reaches the saddle (every atom within 0.2 Å) for 434 of the 490. This measures recall of the training saddles, not discovery of new ones.
Files: model.safetensors (every tensor of the network, EMA weights), config.json (architecture, including UMA-S's
own model config, so no UMA checkpoint is needed; inference settings), al100_adatom_minimum.extxyz (the minimum).
Load and run it with backend.py of the demo Space.
The weights are derived from Meta's UMA and are released under the FAIR Chemistry License.
- Downloads last month
- 5