Garden's Almanac of Matter Models Models
Machine-learned interatomic potentials grouped by similarity of representation, architecture family, and publication year.
By Similarity of Representation
By Architecture Family
Fixed Descriptor
Computes a fixed descriptor of each atom's local environment in one shot and maps it directly to energy through a small per-element network, with no message passing. The first popular family of neural-network potentials.
Example paper Devereux et al., ANI-2x (2020)
Message Passing
The plain baseline tier of graph networks: atoms exchange messages with their neighbors over several rounds of aggregation. (Other families pass messages too, but layer on extra structure that sets them apart.)
Example paper Schütt et al., SchNet (2018)
Equivariant
Networks that build 3-D rotational symmetry directly into their layers, so predictions transform correctly under rotation by construction rather than by data augmentation.
Example paper Liao et al., EquiformerV2 (2023)
Atomic Cluster Expansion
Also equivariant. These models expand each atom's neighborhood into a body-ordered basis, with modern variants layering message passing on top.
Example paper Batatia et al., MACE (2024)
Learned Symmetry
Non-equivariant graph networks that learn rotational consistency from data rather than building symmetry into the architecture.
Example paper Neumann et al., Orb (2024)
Attention / Graph-Transformer
Self-attention transformers (like in LLMs) adapted to atoms. Each atom weighs other atoms by learned content rather than a fixed aggregation rule.
Example paper Qu et al., AllScAIP (2026)