Skip to content

A generative model for inorganic materials design

Abstract

MatterGen is a diffusion-based generative model for inorganic crystals that jointly denoises atom types, fractional coordinates, and the periodic lattice, with optional adapter-based fine-tuning toward composition, symmetry, and scalar properties; the manuscript reports higher rates of stable, unique, novel structures and closer DFT-relaxed geometries than prior generative baselines, plus laboratory synthesis of one designed compound with measured property within about 20% of target.

Summary

The paper introduces MatterGen, framed as a diffusion process tailored to periodic crystals: corruption and denoising operate on atom types, coordinates, and the lattice. A pretrained equivariant score network is fine-tuned with adapters so generation can be steered toward chemistry, space-group symmetry, and properties such as magnetic density. Density functional theory (DFT) relaxation is used to assess stability and proximity to local energy minima relative to prior generative models. A laboratory synthesis example compares a measured physical property to a target value.

Methods

Generative modeling

  • Architecture: Diffusion model reversing a corruption of crystal structures; equivariant score network jointly denoises atom types, coordinates, and lattice (Fig. 1 in the paper).
  • Training / fine-tuning: Pretraining on a large stable-structure corpus; adapter modules encode property or constraint labels for fine-tuning (chemistry, symmetry, scalar targets).
  • Baselines: Comparisons include earlier generative crystal models and, for some tasks, substitution / random structure search style baselines (as presented in the figures).

Validation

  • DFT: Generated structures are evaluated with DFT relaxation to quantify stability and structural closeness to relaxed ground states (formation-energy and structure-matching metrics in the main text).
  • Experiment: At least one synthesized compound is reported with a measured property within ~20% of the target (Fig. 6 in the paper).

Findings

  • Performance: MatterGen is reported to more than double the fraction of stable, unique, novel (S.U.N.) structures versus selected prior models and to produce structures >10× closer to DFT-relaxed geometries in the stated benchmark.
  • Conditioning: After fine-tuning, the model can target multiple constraints (illustrative figure shows high magnetic density with composition chosen for supply-chain considerations).
  • Experiment: The proof-of-concept synthesis aligns a measured property with the target within about 20%. The PDF in pdf_path is a Nature accelerated article preview; use the VOR for final table and supplement numbers.

Limitations

The corpus PDF is a Nature accelerated article preview (pre-copyedit disclaimer). DFT settings, full training data, and experimental protocols appear in the main text / Supplement rather than fully in the first pages alone.

Relevance to group

Not a ReaxFF paper; useful as inverse materials design and generative-model context adjacent to ML atomistic workflows.

Citations and evidence anchors

  • DOI: 10.1038/s41586-025-08628-5 — Nature (accepted manuscript PDF in corpus).