Researchers from the Tokyo University of Science, Kyoto Institute of Technology, University of Tsukuba and National Institute for Materials Science (NIMS) have developed an interpretable machine-learning framework that automatically detects anomalies in Fermi surface maps of the spintronic Heusler alloy Co₂MnGaₓGe₁₋ₓ (CMGG). The approach uses principal component analysis (PCA) on simulated Fermi-surface images to pinpoint compositions where the electronic structure changes sharply, and links these anomalies directly to nodal-line formation and variations in spin polarization.
In this work, the team focuses on CMGG, a Heusler alloy with half-metallicity, nodal-line features and high spin polarization, known for its anomalous Nernst effect arising from nodal lines on the Fermi surface. Using density functional theory (DFT), they first generate a composition-dependent band-structure dataset and extract kₓ–kᵧ Fermi-surface cuts through the Γ point. These images are blurred to roughly approximate ARPES data, then converted into one-dimensional vectors and analyzed via PCA to obtain a low-dimensional representation where each point corresponds to a specific Ga content x.
The core result is that pronounced “jumps” in PCA space correlate with characteristic changes in physical properties. Compositions near Ga = 0.94–0.95 appear as clear outliers in the PCA trajectory, and these outliers coincide with extrema and inflection points in the spin-polarization curve. Band-structure analysis attributes these anomalies to the emergence of a nodal line in the vicinity of the Fermi level, and the position of this nodal line in momentum space is automatically detected by a differential analysis of the Fermi-surface images.
The authors quantify this behavior by evaluating jump distances between neighboring compositions in PCA space. They find that the top 1% of compositions ranked by jump magnitude correspond to the appearance of the nodal line, while the top 10% have a particularly strong impact on spin polarization. This shows that a simple PCA‑based ranking can efficiently extract compositions associated with key changes in electronic and spin-dependent properties, without requiring manual inspection of every Fermi-surface map.
To test robustness, the team intentionally broadens the images and adds strong noise to mimic less‑than‑ideal ARPES conditions. Even under these adverse settings, the method continues to flag Ga ≈ 0.94–0.95 as anomalous compositions and correctly associates them with nodal-line formation and changes in spin polarization. Because the framework is built around unsupervised PCA and distance-based anomaly detection, and is formulated as a human‑in‑the‑loop ranking tool, it is particularly suited for organizing and screening large-scale ARPES datasets on composition-gradient (combinatorial) films, where the main goal is to detect non‑systematic anomalies superimposed on an otherwise smooth composition trend.
Looking ahead, the authors note that incorporating energy dispersion into the current framework could extend the analysis from Fermi-level cuts to full band-structure modulations, potentially revealing deeper correlations between electronic structure and physical properties. They also highlight that the same simple scheme - PCA plus anomaly detection on image data—could be applied to other material systems where characteristic Fermi-surface features, such as nodal lines or flat bands, play a central role in determining spintronic or topological functionality.