Accurate Screening of Functional Materials with Machine-Learning Potential and Transfer-Learned Regressions: Heusler Alloy Benchmark

TL;DR

The paper presents an ML-accelerated high-throughput workflow (ML-HTP) that integrates fast structure optimization with predictive physics-aware property models to screen vast Heusler spaces for stable, high magnetocrystalline anisotropy materials. By combining eSEN-30M-OAM for structure optimization and hull analysis with eSEM-based predictions trained on the DxMag Heusler database, and employing frozen transfer learning from a universal interatomic potential, the approach achieves high precision validated by DFT and demonstrates strong generalization to unseen elements. The study identifies 334 conventional quaternary and 924 all-$d$ Heusler candidates with large $E_{ m{aniso}}$ that meet stability criteria, and shows substantial speed-ups over full DFT screening. The framework is modular and transferable to other material classes with available training data, and active-learning strategies could further enhance coverage and accuracy for future discovery.

Abstract

A machine learning-accelerated high-throughput (HTP) workflow for the discovery of magnetic materials is presented. As a test case, we screened quaternary and all-$d$ Heusler compounds for stable compounds with large magnetocrystalline anisotropy energy ($E_{\mathrm{aniso}}$). Structure optimization and evaluation of formation energy and distance to hull convex were performed using the eSEN-30M-OAM interatomic potential, while local magnetic moments, phonon stability, magnetic stability, and $E_{\mathrm{aniso}}$ were predicted by eSEM models trained on our DxMag Heusler database. A frozen transfer learning strategy was employed to improve accuracy. Candidate compounds identified by the ML-HTP workflow were validated with density functional theory, confirming high predictive precision. We also benchmark the performance of different uMLIPs, discuss the fidelity of local magnetic moment prediction, and demonstrate generalization to unseen elements via transfer learning from a universal interatomic potential.

Figures (6)

• Figure 1: Frozen transfer learning overview and ML-HTP workflow. a) Schematic of the development of the via frozen transfer learning, using as the base model. The is used to perform structure optimization, formation energy calculation, and convex hull distance evaluation. The predicts properties from structures. b) Workflow of the case study which identified stable conventional quaternary and all-$d$ Heusler compounds exhibiting strong . Counts of quaternary and all-$d$ compounds at each stage are reported as quaternary/all-$d$.
• Figure 2: DFT validation summary of ML-HTP selected compounds. For ML-selected candidate lists of conventional quaternary (334) and all-$d$ (924) Heusler compounds, the percentages that DFT results satisfy the screening criteria (i.e., the ML‑HTP precision) are shown as blue and yellow bars. For comparison, the precision of the ML models measured on the test set of conventional ternary compounds is also shown as green bars. The test set size for $c/a$ ratio, , and is 10,000 and for , , , and is 10% of the dataset size shown in Table \ref{['tab:ml_metrics']}.
• Figure 3: Distribution of ML-selected compounds on elements contained. Distribution of ML-selected candidate compounds based on whether 4d or 5d elements are present and distribution over 4d and 5d elements contained. The distribution of DFT validated strong candidates is also shown.
• Figure 4: Benchmark of performance. Lattice constants $a$ and $c$, $c/a$ ratio, total energy ($E$), formation energy (), and convex hull distance () predicted by various are benchmarked against DFT references. For each property, the fraction of compounds with predictions falling within specified relative error (RE) or absolute error (AE) thresholds is reported. Energetic quantities ($E$, , and ) are expressed in eV/atom. The test set consists of 10,000 ground-state compounds randomly sampled from .
• Figure 5: Local magnetic moment prediction performance. a) Scatter plot comparing ML-predicted with DFT values for the test set. b) Learning curves for prediction. The top panel shows $R^2$ scores for both local moments and their magnitudes. The bottom panel reports the magnetic/nonmagnetic classification accuracy, and the fraction of compounds with absolute prediction error below 0.1 $\mu_{\mathrm{B}}$ for all compounds and for the magnetic subset.