DMFlow: Disordered Materials Generation by Flow Matching
Liming Wu, Rui Jiao, Qi Li, Mingze Li, Songyou Li, Shifeng Jin, Wenbing Huang
TL;DR
DMFlow is a novel generative framework that models disordered crystals using flow matching and graph neural networks, enabling the creation of physically valid disordered structures and advancing materials discovery.
Contribution
Introduces DMFlow, a flow matching-based generative model for disordered crystals with a unified representation and a Riemannian flow framework ensuring valid disorder weights.
Findings
Outperforms state-of-the-art methods in crystal structure prediction.
Provides a new benchmark dataset for disordered crystal structures.
Successfully generates diverse disordered materials with physical validity.
Abstract
The design of materials with tailored properties is crucial for technological progress. However, most deep generative models focus exclusively on perfectly ordered crystals, neglecting the important class of disordered materials. To address this gap, we introduce DMFlow, a generative framework specifically designed for disordered crystals. Our approach introduces a unified representation for ordered, Substitutionally Disordered (SD), and Positionally Disordered (PD) crystals, and employs a flow matching model to jointly generate all structural components. A key innovation is a Riemannian flow matching framework with spherical reparameterization, which ensures physically valid disorder weights on the probability simplex. The vector field is learned by a novel Graph Neural Network (GNN) that incorporates physical symmetries and a specialized message-passing scheme. Finally, a two-stage…
Peer Reviews
Decision·Submitted to ICLR 2026
1. The paper proposes a unified representation for both ordered and disordered crystals. 2. On the CSP tasks, the proposed method outperforms the baseline methods on a SPD dataset - a mixed set combining both SD and PD structures. 3. The paper contributes a new dataset for benchmarking disordered crystal generation.
1. The scope of the paper is limited. The paper only considers binary PD while leaving generality to >2 positions per site untested. This may limit applicability to broader disordered classes. 2. In the CSP experiments, the proposed DMFlow only marginally outperforms the baseline FlowMM-Prob on binary PD cases. 3. In the DNG experiments, the authors do not compare against existing generative baselines. The authors argue that the existing methods can only generate one-hot element assignments.
The paper is quite strong in terms of novelty, as it is the first work I am aware of that applies diffusion models to disordered materials. It is clearly written, and the planned release of the datasets represents an important step toward enabling more generative models for this type of problem.
I believe the paper would benefit from a more detailed discussion of the design of the CSP experiment (see questions below). There are also some design choices that are not fully ablated, such as the use of five different heuristics for the majority vote during sampling. It would be interesting to see how each heuristic performs individually and whether all of them are necessary. The framework also relies heavily on manually set thresholds during sampling, and it is not entirely clear how the va
* The paper tackles an **underexplored but relevant problem** in generative modeling—extending flow-based methods to **disordered crystal structures**. * The work contributes a **new benchmark dataset** for disordered materials, which could be a useful resource for future research.
* **Relatively marginal improvement over baselines:** The experimental results show only small performance gains compared to **FlowMM**, suggesting that the proposed method offers limited novelty or practical advancement from a machine learning perspective. I wonder the performance compared to more recent crystal generation methods e.g. CrysBFN [1] and MatterGen [2] * **Relatively minor machine learning contribution** The paper's contribution to the core machine learning field is relatively
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsMachine Learning in Materials Science · Block Copolymer Self-Assembly · Catalysis and Oxidation Reactions