Sovereign 8s.is mirror of standardgalactic/catalyst-exploration
  • Jupyter Notebook 50.9%
  • Python 48.8%
  • TeX 0.2%
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
2026-03-04 20:15:53 +09:00
assets Remove overview.png as part of asset cleanup to streamline project resources. 2026-03-03 23:22:26 +09:00
OC20 Refactor project structure by removing obsolete files and directories, updating .gitignore to include new output paths, and enhancing documentation references. This cleanup aims to streamline the repository and improve organization. 2026-03-03 23:34:40 +09:00
OMat24_and_HTS Update .gitignore to include 'logs/' and 'output/' directories, and add a new section in README for checkpoints related to OMat24 and HTS experiments with a link to Google Drive resources. 2026-03-04 00:14:01 +09:00
ORR_screening Add ORR catalyst screening pipeline with data and DFT structures 2026-03-04 20:15:53 +09:00
SPICE Update environment setup in README, rename configurations and scripts from 'crack' to 'ark', and adjust paths for scaling factors and model references to reflect the new naming convention. 2026-03-04 00:05:10 +09:00
.gitignore Update .gitignore to include 'logs/' and 'output/' directories, and add a new section in README for checkpoints related to OMat24 and HTS experiments with a link to Google Drive resources. 2026-03-04 00:14:01 +09:00
env.yml Refactoring the codebase 2026-03-03 23:15:20 +09:00
LICENSE Add MIT License to the project 2026-03-03 23:25:22 +09:00
README.md Update .gitignore to include .pkl files, revise OMat_and_HTS README title for clarity, and enhance SPICE README with detailed environment and training instructions for the EquiformerV2 model. 2026-03-03 23:43:05 +09:00

ARK: Angular Relational Knowledge Distillation

This repository contains the implementation used in:

Angular relational knowledge distillation of machine learning interatomic potentials for scalable catalyst exploration

ARK distills relational physics from a large teacher MLIP into a compact student by aligning edge-level relational vectors with a contrastive objective.

Model configurations and benchmark results

Model configurations

Compression ratio denotes the parameter reduction from teacher to student.

Domain Dataset Teacher Student Compression
Catalysis OC20 EquiformerV2 (153M) EquiformerV2 (22M) 7.0x
Materials OMat24 EquiformerV2 (153M) EquiformerV2 (22M) 7.0x
Molecules SPICE MACE-OFF Large (4.7M) GemNet-dT (0.67M) 7.0x

Distillation results

ARK uses n2n loss for OC20 and OMat24, and n2n + Hessian loss for SPICE.
Hessian distillation was not applied to OMat24.
Best student results are in bold.

Training strategy OC20 O* Energy OC20 O* Force OC20 200K Energy OC20 200K Force OMat24 Rattled-1000 Energy OMat24 Rattled-1000 Force OMat24 Rattled-1000 Stress SPICE Monomers Energy SPICE Monomers Force SPICE Solvated AA Energy SPICE Solvated AA Force SPICE Iodine Energy SPICE Iodine Force
Teacher 39.8 5.8 171.5 12.4 9.7 56.3 3.4 0.65 6.6 1.3 19.4 1.3 15.3
Undistilled 294.5 5.9 474.9 51.8 18.1 92.5 4.1 2.2 13.4 1.7 22.9 3.2 25.4
n2n 252.9 5.8 412.8 34.8 17.5 88.6 4.1 2.3 14.5 1.5 21.4 3.0 25.9
Hessian OOM OOM OOM OOM OOM OOM OOM 1.2 8.5 0.4 11.4 2.4 19.6
ARK 231.7 5.8 371.1 34.1 16.3 83.0 4.0 1.4 8.9 0.4 11.9 2.2 19.4

All entries are Mean Absolute Error (MAE) unless otherwise noted.
OC20 energy units: meV.
OMat24 and SPICE energy units: meV/atom.
Force units: meV/A.
Stress units: meV/A^3.

Experiments

Experimental details are described in the README.md files of each subfolder: OC20, OMat_and_HTS, and SPICE.

Acknowledgement

This repository builds on: