MiAD: Mirage Atom Diffusion for De Novo Crystal Generation

TL;DR

MiAD addresses a key limitation in de novo crystal generation: fixed atom counts during diffusion. By introducing mirage infusion, it expands the generative space with mirage atoms that can materialize into real atoms or vanish, enabling dynamic changes in atom numbers during generation while preserving crystal symmetries. The method integrates with a joint diffusion model over the lattice, fractional coordinates, and atom types, and demonstrates substantial gains in stability, uniqueness, and novelty, achieving competitive or state-of-the-art S.U.N. scores on MP-20. The approach broadens the practical impact of diffusion-based materials discovery, albeit at increased computational cost, and suggests broad applicability to other diffusion-based frameworks in crystallography and related domains.

Abstract

In recent years, diffusion-based models have demonstrated exceptional performance in searching for simultaneously stable, unique, and novel (S.U.N.) crystalline materials. However, most of these models don't have the ability to change the number of atoms in the crystal during the generation process, which limits the variability of model sampling trajectories. In this paper, we demonstrate the severity of this restriction and introduce a simple yet powerful technique, mirage infusion, which enables diffusion models to change the state of the atoms that make up the crystal from existent to non-existent (mirage) and vice versa. We show that this technique improves model quality by up to $\times2.5$ compared to the same model without this modification. The resulting model, Mirage Atom Diffusion (MiAD), is an equivariant joint diffusion model for de novo crystal generation that is capable of altering the number of atoms during the generation process. MiAD achieves an $8.2\%$ S.U.N. rate on the MP-20 dataset, which substantially exceeds existing state-of-the-art approaches. The source code can be found at \href{https://github.com/andrey-okhotin/miad.git}{\texttt{github.com/andrey-okhotin/miad}}.

Figures (6)

• Figure 1: Overview of the proposed mirage infusion technique.
• Figure 2: Comparison of MiAD (DiffCSP with mirage infusion) and DiffCSP in terms of stability, uniqueness, novelty, and S.U.N. Stability is estimated via eq-V2.
• Figure 3: Comparison of MiAD (DiffCSP with mirage infusion) in its final version and DiffCSP in terms of stability, uniqueness, novelty, and S.U.N. Stability is estimated via (left) eq-V2 and (right) CHGNet. MiAD outperforms DiffCSP across all metrics, especially in terms of stability rate and S.U.N., achieving the highest quality after $8000$ epochs.
• Figure 4: Number of atoms in S.U.N. crystals generated by MiAD. We consider only S.U.N. crystals among the 10000.0 crystals generated by MiAD. After generation, crystals are prerelaxed via CHGNet. Stability is also estimated via CHGNet.
• Figure 5: Comparison of models for de novo crystal generation in terms of numbers of S.U.N. crystals belonging to main spacegroup families. We categorize only S.U.N. crystals produced by generative models with a fixed budget of 10000.0 generated crystals. After generation, crystals are prerelaxed via CHGNet. Stability is also estimated via CHGNet. Spacegroup families are identified via SpacegroupAnalyzer from the pymatgen package with tolerance: $0.1$.