[Paper Review] MatterGen: a generative model for inorganic materials design

Claudio Zeni, Robert Pinsler+19 more|arXiv (Cornell University)|Dec 6, 2023

Machine Learning in Materials Science69 citations

TL;DR

MatterGen is a diffusion-based generative model that creates stable, diverse inorganic materials and can be fine-tuned to meet target chemistry, symmetry, and scalar properties.

ABSTRACT

The design of functional materials with desired properties is essential in driving technological advances in areas like energy storage, catalysis, and carbon capture. Generative models provide a new paradigm for materials design by directly generating entirely novel materials given desired property constraints. Despite recent progress, current generative models have low success rate in proposing stable crystals, or can only satisfy a very limited set of property constraints. Here, we present MatterGen, a model that generates stable, diverse inorganic materials across the periodic table and can further be fine-tuned to steer the generation towards a broad range of property constraints. To enable this, we introduce a new diffusion-based generative process that produces crystalline structures by gradually refining atom types, coordinates, and the periodic lattice. We further introduce adapter modules to enable fine-tuning towards any given property constraints with a labeled dataset. Compared to prior generative models, structures produced by MatterGen are more than twice as likely to be novel and stable, and more than 15 times closer to the local energy minimum. After fine-tuning, MatterGen successfully generates stable, novel materials with desired chemistry, symmetry, as well as mechanical, electronic and magnetic properties. Finally, we demonstrate multi-property materials design capabilities by proposing structures that have both high magnetic density and a chemical composition with low supply-chain risk. We believe that the quality of generated materials and the breadth of MatterGen's capabilities represent a major advancement towards creating a universal generative model for materials design.

Motivation & Objective

• Address the need for fast, inverse-design of stable inorganic crystals across the periodic table.
• Develop a diffusion-based generative process tailored for crystalline materials by jointly denoising atom types, coordinates, and lattice.
• Enable fine-tuning via adapter modules to steer generation toward specific chemical, symmetry, and property constraints.
• Demonstrate that MatterGen generates more novel and stable structures and is closer to local energy minima than prior models.

Proposed method

• Propose a diffusion-based generative process that corrupts atom types, coordinates, and lattice with physically motivated noise distributions.
• Use an equivariant score network to jointly denoise atom types, coordinates, and lattice (base model).
• Introduce adapter modules for fine-tuning on labeled property datasets and apply classifier-free guidance to steer outputs.
• Pre-train on a large, diverse dataset (Alex-MP-ICSD) of 607,684 stable structures up to 20 atoms and define stability via 0.1 eV/atom threshold.
• Fine-tune to target chemistry, symmetry (space groups), and scalar properties (e.g., magnetic density, band gap, bulk modulus).
• Compare performance to baselines (e.g., CDVAE, P-G-SchNet, G-SchNet, FTCP) and show gains in stability, novelty, and proximity to relaxed structures.

Experimental results

Research questions

• RQ1 Can MatterGen generate stable inorganic materials across a wide range of elements and stoichiometries?
• RQ2 How well can the base model be steered toward target chemistry, symmetry, and scalar properties via adapters?
• RQ3 How does MatterGen perform in terms of stability, novelty, and distance to local energy minima compared with prior generative models?
• RQ4 Can MatterGen design multi-property materials such as low-supply-chain-risk magnets while satisfying multiple constraints?

Key findings

• 78% of generated structures lie below 0.1 eV/atom above hull (13% below 0.0 eV/atom) using the hull of Alex-MP-ICSD reference.
• 75% of generated structures lie below 0.1 eV/atom above hull using the Alex-MP-ICSD hull.
• 95% of generated structures have a relaxation RMSD to their relaxed structure below 0.076 Å.
• MatterGen shows 86% novelty and 100% uniqueness when generating 1,000 structures, with novelty remaining stable around 68% up to 1,000,000 structures.
• MatterGen-MP (trained on MP-20) achieves 1.8x higher structure percentage and 3.1x smaller average distance to relaxation than CDVAE; full MatterGen achieves further improvements (1.6x higher, 5.5x smaller).
• Fine-tuned MatterGen can discover materials in target chemical systems with higher hull-constrained stability and unique hull positions, especially in quinary systems where efficiency gains are substantial.