Title: A Benchmark Dataset for Catalysis Research on 2D MXenes

Abstract:

By integrating first-principles calculations with machine learning (ML), this study seeks to expedite the discovery of catalytic properties in emerging materials. We concentrate on two-dimensional (2D) Ti$_2$CT$_y$ MXenes, which are highly attractive for catalytic applications due to their adaptable surface chemistry. However, determining their precise composition and structure under realistic operating conditions remains beyond the capabilities of standard density functional theory (DFT) because of the prohibitive computational expenses involved.

To overcome this bottleneck, we have constructed a robust dataset comprising 50,000 DFT calculations for training purposes and 10,000 for testing. This collection includes various Ti$_2$CT$_y$ MXene configurations as well as molecular systems. Furthermore, we introduce a separate test set containing 1,000 genuinely novel, larger systems to assess the generalization capabilities of the models. We train and evaluate several prominent machine learning interatomic potential (MLIP) models, such as EquiformerV2, MACE, MatRIS, and UPET. These models demonstrate the ability to accurately predict atomic forces and formation energies—critical metrics that DFT must frequently calculate for structural and catalytic studies—within these 2D materials.

This hybrid DFT-ML approach delivers a computational speedup of roughly $1-4 \cdot 10^3$ on CPU hardware while preserving high accuracy levels (approximately +/- $10$ meV/A for forces and +/- $1$ meV for per-atom energies). This efficiency opens new avenues for more streamlined investigations into the catalytic behavior of MXenes. Additionally, we conduct a thorough qualitative assessment of the trained models, highlighting the value of comprehensive simulation-based comparisons that go beyond standard benchmark metrics. The dataset, trained models, and associated code are publicly accessible at https://huggingface.co/datasets/CatalystAnonymous/catalyst_mxenes.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC