MLIP Features Enhance Material Generation and Evaluation

Paul Hagemann, Katharina Ueltzen, Simon M\"uller, Janine George, Philipp Benner· August 3, 2026 View original

Key takeaways

  • Pretrained MLIP features (e.g., MACE) are powerful for material generation and evaluation.
  • The Coarse-Fine Transport Distance (CFTD) metric assesses both quality and novelty of generated materials.
  • CFTD uses MACE features to capture crystal-structure quality and detect memorization.
  • Coarse MACE features can guide generative models for targeted material design.

Who benefits

Materials ScienceChemicalsPharmaceuticalsEnergyManufacturing

Summary

This paper demonstrates the utility of atom-averaged features from pretrained Machine-Learning Interatomic Potentials (MLIPs), like MACE, as coarse coordinates for generating and evaluating inorganic crystal structures. It introduces the Coarse-Fine Transport Distance (CFTD) metric, which uses MACE features to assess both the quality and novelty of generated materials.

New research highlights the significant potential of features derived from pretrained Machine-Learning Interatomic Potentials (MLIPs), such as MACE, for tasks involving the generation and evaluation of inorganic crystal structures. These atom-averaged features serve as effective "coarse coordinates" that capture essential structural information. The study introduces a novel distance measure called the Coarse-Fine Transport Distance (CFTD). This metric leverages the coarse MACE features to simultaneously quantify both the quality and novelty of crystal structures produced by generative models, offering a comprehensive distribution-based evaluation framework. The versatility of CFTD is showcased in its ability to accurately assess crystal-structure quality while also identifying instances of memorization within generative models, providing a robust alternative to existing metrics.

Why it matters

For materials scientists and engineers, this advancement provides more sophisticated tools to design and evaluate novel materials, potentially accelerating the discovery of materials with desired properties for various applications.

How to implement this in your domain

  1. 1Investigate the use of pretrained MLIPs (e.g., MACE) to extract atom-averaged features for material characterization.
  2. 2Integrate the Coarse-Fine Transport Distance (CFTD) metric into workflows for evaluating generative material models.
  3. 3Utilize MACE features as guidance or input for developing new material generative models to steer them towards desired structural properties.
  4. 4Collaborate with computational materials science teams to apply these techniques for discovering and optimizing novel inorganic crystal structures.

Original post by Paul Hagemann, Katharina Ueltzen, Simon M\"uller, Janine George, Philipp Benner

"arXiv:2607.28776v1 Announce Type: new Abstract: Generative machine learning is increasingly used for inorganic crystal structure generation. Most models and the corresponding evaluation approaches rely on simple forms of crystal structure representation. In this paper, we showcas…"

View on X

Originally posted by Paul Hagemann, Katharina Ueltzen, Simon M\"uller, Janine George, Philipp Benner on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses