Generalization of graph-based active learning relaxation strategies across materials
Xiaoxiao Wang, Joseph Musielewicz, Richard Tran, Sudheesh Kumar Ethirajan, Xiaoyan Fu, Hilda Mera, John R. Kitchin, Rachel C. Kurchin and Zachary W. Ulissi, “Generalization of graph-based active learning relaxation strategies across materials”, Machine Learning: Science and Technology 5(2), 025018 (2024). https://doi.org/10.1088/2632-2153/ad37f0
This summary was written by a language model from the paper and has not been checked line by line. Rely on the published paper for details.
Question
Finetuna is an online active learning framework that relaxes structures with a graph neural network (GNN) potential pretrained on the Open Catalyst data and fine-tunes it on the fly with occasional DFT single points. On OC20-like adsorbate/slab systems it had cut DFT calls by about 90%. This paper asks how well the approach carries over to systems outside the pretraining domain: larger (C3 and up) alcohol adsorbates, spin-polarized metal oxides, and fully three-dimensional porous structures such as zeolites and metal-organic frameworks, and what querying strategies are needed when the pretrained model is less trustworthy (Wang et al. 2024).
Approach
In the workflow, a machine learning potential supplies forces to an ASE BFGS optimizer, and a VASP single point (a “parent call”) is triggered when a querying rule fires; its forces are used to retrain the model. The baseline rule queries at the initial structure, every 30 ML steps, and whenever the ML maximum force drops below the 0.05 eV/Å convergence criterion. A new, more conservative rule was added for difficult systems: if the retrained force error (the L2 norm of the difference between DFT and retrained ML forces) stays above 0.05 eV/Å, another parent call is made on the next step, repeating until the error falls below threshold. Pretrained models compared were GemNet trained on OC20 (GN-OC20), GemNet trained on OC22 (GN-OC22), and GemNet-OC trained on both (GNOC-OC20+22).
Test sets were 27 alcohol-to-ketone dehydrogenation systems (2-propanol, 2-butanol and cyclohexanol intermediates, computed with BEEF-vdW), 10 oxide systems from the OC22 validation set (spin-polarized PBE with Hubbard U, OC22 settings), and 81 3D structures (copper-doped zeolites, a MOF, and zeolites with adsorbates such as gold nanoparticles and aromatics, computed with RPBE), ranging from 50 to 250 atoms. Efficiency was measured as Finetuna DFT calls relative to a pure VASP BFGS relaxation, and accuracy as the relaxed energy difference per atom.
Key findings
For alcohols, GN-OC20 with the baseline rule converged 81% of runs using 18% of the DFT calls. Single points showed that RPBE and BEEF-vdW forces differ by less than 0.1 eV/Å, much less than the DFT versus ML discrepancy, and runs tended to fail when the initial DFT-ML force difference exceeded about 0.4 eV/Å. For oxides, no GN-OC20 run converged, which the authors attribute to the absence of oxides and spin information in OC20. GN-OC22 converged 90% of oxide runs with 58% of DFT calls, and GNOC-OC20+22 converged 70% with 36%. By comparison, VASP’s own conjugate gradient optimizer needed 154% of the BFGS calls and GPMin converged 70% with 70%. For the 3D structures, GN-OC20 with the added high-error rule converged 92% of runs with 19% of DFT calls.
Across all classes, most unconverged runs had ML-predicted forces above 5 eV/Å somewhere along the trajectory, which the authors propose as a heuristic for abandoning a Finetuna run in favor of plain DFT or a different initialization. A few alcohol runs were exceptions: they converged slowly, sometimes exceeding 100% of the DFT calls, to minima similar to the DFT result. A detailed AgIrO4 case showed two relaxations ending in nearly identical atomic positions but different total magnetization and energy, illustrating that spin is an extra degree of freedom that breaks the structure-to-energy mapping assumed by a local-cutoff GNN. A PCA map of per-atom GemNet embeddings suggested the alcohols are most in-domain and zeolites furthest out.
The text reports the headline savings inconsistently: the abstract cites 80% fewer DFT calls for alcohols and 3D structures and 42% for oxides, while the conclusion cites 85%, 49% and 82% for alcohols, oxides and MOFs. The table values above are the ones given in the results. This summary is based on the arXiv preprint text, and some supplementary figure references are unresolved in it.
Significance
The study maps the limits of fine-tuning a pretrained catalysis GNN during a relaxation: savings remain large when the new system resembles the pretraining data even if it is bigger, but shrink sharply for magnetic oxides, where long-range spin effects lie outside what the model represents. It argues that the querying strategy should become more conservative the further a system lies from the training domain, and points to out-of-domain detection, uncertainty-based querying, and magnetism-aware potentials such as CHGNet as next steps (Wang et al. 2024).