[Paper Review] CatFlow: Co-generation of Slab-Adsorbate Systems via Flow Matching

Minkyu Kim, Nayoung Kim+2 more|arXiv (Cornell University)|Feb 5, 2026

Machine Learning in Materials Science0 citations

TL;DR

CatFlow jointly co-generates slab structures and adsorbate coordinates using flow matching with a factorized slab-adsorbate representation, improving structural fidelity and adsorption-energy alignment on OC20 compared to baselines in both de novo generation and structure prediction.

ABSTRACT

Discovering heterogeneous catalysts tailored for specific reaction intermediates remains a fundamental bottleneck in materials science. While traditional trial-and-error methods and recent generative models have shown promise, they struggle to capture the intrinsic coupling between surface geometry and adsorbate interactions. To address this limitation, we propose CatFlow, a flow matching-based framework for de novo design and structure prediction of heterogeneous catalysts. Our model operates on a primitive cell-based factorized representation of the slab-adsorbate complex, reducing the number of learnable variables by an average of 9.2x while explicitly encoding the surface orientation of the slab-adsorbate interface. Experiments on the Open Catalyst 2020 dataset demonstrate that CatFlow significantly improves the structural fidelity of generated catalysts compared to autoregressive and sequential baselines. Further experiments show that the generated structures accurately capture the adsorption energy distributions of physically plausible interfaces and lie closer to thermodynamic local minima.

Motivation & Objective

• Address the bottleneck of discovering heterogeneous catalysts by coupling surface geometry with adsorbate interactions.
• Propose a unified framework that co-generates slab structures and adsorbate coordinates.
• Introduce a factorized representation to reduce model dimensionality while preserving surface orientation.
• Demonstrate end-to-end generation and structure prediction on the OC20 benchmark with superior fidelity and energetics.

Proposed method

• Define a conditional joint distribution p(S_prim, M, k_vac, x_ads | a_ads) and train a single model with continuous and discrete flow matching to co-generate slab-adsorbate systems.
• Introduce a factorized representation of slab-adsorbate systems into primitive cell, transformation matrix, vacuum scaling factor, and adsorbate components to reduce learnable variables while preserving surface orientation information.
• Employ discrete flow matching with masking for the atomic species and continuous flow matching for geometric variables, including reparameterizations and a relaxed M during training.
• Use a transformer-based neural network (DiT-inspired encoder/decoder) to process joint atom-level representations and predict both continuous coordinates and discrete compositions.
• In inference, solve an ODE for continuous variables and perform iterative unmasking for discrete tokens to generate de novo structures, or fix composition for structure prediction.

Figure 1 : Visualization of the co-generation trajectory conditioned on the adsorbate. We illustrate the synchronized evolution of the slab-adsorbate system from the initial noise distribution ( $t=0$ ) to the final structure ( $t=1$ ) for de novo generation (top) and structure prediction (bottom).

Experimental results

Research questions

• RQ1 Can CatFlow jointly generate slab structures and adsorbate placements while capturing surface-adsorbate interactions?
• RQ2 Does the factorized representation reduce dimensionality without sacrificing surface orientation and physical validity?
• RQ3 How does end-to-end co-generation compare to modular pipelines (e.g., DiffCSP + AdsorbDiff) in de novo generation and structure prediction on OC20?
• RQ4 What is the quality of generated adsorption energetics relative to reference minima across diverse adsorbates?

Key findings

• CatFlow outperforms CatGPT on de novo generation across validity, uniqueness, and convergence efficiency, achieving 97.33% validity vs 92.67% and 94.69% uniqueness vs 79.91%, with lower system energy change and faster convergence.
• In structure prediction, CatFlow greatly surpasses the two-step DC+AD baseline in validity (98.16% vs 64.95%), match rate (11.09% vs 0.01%), RMSD (0.0973 vs 0.1833), and adsorption-energy success rate (9.72% vs 1.85%).
• CatFlow generates adsorption-energy distributions that more closely align with reference energies across many adsorbates, indicating generation of configurations near physically plausible local minima.
• The factorized representation reduces dimensionality (average ~9.2x, up to 96x) while preserving surface orientation, enabling efficient co-generation of slab-adsorbate systems.
• The end-to-end framework explicitly models surface-adsorbate interactions and conditions on adsorbate identity, enabling generalization beyond predefined adsorbate tokens.

Figure 2 : Histogram of atom counts in catalyst structures. We compare the histograms of atom counts for slab structures (blue) and their corresponding primitive cells (green). The primitive cells require fewer atoms than the slab structures, reducing the number of learnable variables for the genera