Representations from Pretrained Machine-Learning Interatomic Potentials as Coarse Coordinates for Material Generation and Evaluation
Atom-averaged features from pretrained Machine-Learning Interatomic Potentials (MLIPs) like MACE can serve as powerful coarse coordinates for inorganic crystal structure generation and evaluation The Coarse-Fine Transport Distance (CFTD) is introduced as a novel distribution-based evaluation metric that simultaneously captures both quality and novelty of generated materials CFTD uses two featurizers — coarse MACE features for quality assessment and fine features for novelty detection — enabling
Analysis
TL;DR
- Atom-averaged features from pretrained Machine-Learning Interatomic Potentials (MLIPs) like MACE can serve as powerful coarse coordinates for inorganic crystal structure generation and evaluation
- The Coarse-Fine Transport Distance (CFTD) is introduced as a novel distribution-based evaluation metric that simultaneously captures both quality and novelty of generated materials
- CFTD uses two featurizers — coarse MACE features for quality assessment and fine features for novelty detection — enabling detection of model memorization
- The approach is validated against continuous SUN metrics, demonstrating CFTD's versatility in evaluating material generative models
- Coarse MACE features are shown to be effective as guidance signals for steering material generative models
Why It Matters
This work bridges the gap between machine learning interatomic potentials and generative materials science, offering a more physically meaningful representation than simple crystal structure encodings. For AI practitioners working in computational materials discovery, it provides both a better evaluation framework and a practical guidance mechanism that can improve the quality and novelty of generated crystal structures.
Technical Details
- MLIP-based featurization: Atom-averaged features extracted from pretrained MLIPs (e.g., MACE) are used as coarse representations of crystal structures, capturing physical interactions beyond simple geometric descriptors
- Coarse-Fine Transport Distance (CFTD): A dual-featurizer optimal transport-based distance metric where the coarse component (MACE features) assesses structural quality and the fine component evaluates novelty, all within a single distribution-based framework
- Memorization detection: CFTD is capable of identifying when generative models merely reproduce training data rather than generating genuinely novel structures
- Benchmarking: CFTD is compared against continuous SUN metrics, demonstrating its effectiveness in capturing crystal-structure quality while maintaining sensitivity to novelty
- Generative guidance: Coarse MACE features are integrated as conditioning signals in material generative models, improving generation quality
Industry Insight
- Pretrained MLIPs should be considered as a rich source of structural representations for materials generative models, potentially outperforming hand-crafted crystal descriptors
- The CFTD framework sets a new standard for evaluating generative models in materials science by unifying quality and novelty assessment, reducing the risk of overfitting going undetected
- As generative AI for materials discovery matures, distribution-based metrics like CFTD will become essential tools for benchmarking and ensuring that models produce genuinely novel, physically valid structures rather than memorized training examples
Disclaimer: The above content is generated by AI and is for reference only.