Transfer learning and conditioning reduced overfitting and forgetting in fine-tuning MACE potentials for catalysis

1/ AI utilization method: fine-tuning MACE-family machine-learning interatomic potentials, joint training, FiLM conditioning
2/ Applied task: compare transfer learning/fine-tuning for catalytic calculation potentials in terms of adaptation to a new distribution and generality
3/ Reported results: direct tuning showed catastrophic forgetting, joint training improved both accuracy on the new distribution and generality, and FiLM conditioning further improved accuracy
4/ This paper compared transfer learning strategies for fine-tuning MACE-family machine-learning interatomic potentials to adapt interatomic potentials for catalysis.
5/ According to the report, direct tuning on AQCat25 led to catastrophic forgetting that weakened prior knowledge, while joint training improved both accuracy on the new distribution and generality.
6/ In addition, the FiLM approach, which uses system metadata as conditioning, further improved accuracy.
7/ From a materials research perspective, these results suggest that when adapting MLIP for catalysis to a new dataset, learning designs that preserve both the original knowledge and the new distribution are more important than simple retraining.
8/ In particular, using system-specific metadata as input for conditioning may be useful when handling different catalytic families with a single model.
9/ However, the reported performance depends on a specific dataset, split, and catalytic benchmark, so the same improvement is not automatically guaranteed for different elemental compositions or reaction environments.
10/ Also, the core of this issue is a comparison of model adaptation strategies, not direct experimental proof of catalytic performance. [1]