Machine-Learning Classification of Topological Materials Using Crystallographic and Compositional Descriptors for Pre-DFT Screening

Abstract

Identifying topological insulators and semimetals among inorganic crystalline materials traditionally requires density functional theory (DFT) calculations with spin–orbit coupling, a computationally expensive process that limits high-throughput screening. This study examined whether machine learning models trained on ground-state crystallographic and compositional descriptors, obtainable without DFT or space-group information, could reliably classify materials as topologically non-trivial, trivial, or semimetallic. A dataset of 1,391 labelled crystalline entries (230 semimetal, 781 topological, 380 trivial), curated from the Inorganic Crystal Structure Database and the Topological Quantum Chemistry database, was cleaned, including manual correction, via direct ICSD lookup, of a systematic lattice-parameter data-entry artefact (an Excel auto-fill error that had propagated a single genuine c-parameter value across 1,208 rows, identified by cross-validation against unit-cell volume), prior to a further, now pipeline-integrated audit that removed 47 true-duplicate ICSD entries and attempted live TQC API resolution of 10 conflicting-label ICSD groups, none of which returned a match, so the first-encountered row was conservatively retained in each case (net reduction from 1,450 to 1,391 entries), and used to engineer 31 structural and compositional features motivated by the physics of spin-orbit-coupling-driven band inversion. Six classifiers, Random Forest, XGBoost, LightGBM, Support Vector Machine, an Artificial Neural Network, and a soft-voting Ensemble, were trained and optimised using nested stratified cross-validation, Bayesian hyperparameter search, and Synthetic Minority Oversampling Technique (SMOTE)-based class rebalancing applied strictly within each training fold. XGBoost produced the highest observed test-set performance (macro-F1 of 0.744, accuracy of 78.1%, macro-AUC-ROC of 0.896), narrowly ahead of LightGBM (macro-F1 0.721, accuracy 74.9%, macro-AUC-ROC 0.887), though the difference was not statistically significant (McNemar's test, p = 1.000, Bonferroni-corrected); LightGBM was retained as the primary model for interpretability and screening on the basis of its lower cross-validation variance and better-behaved post-hoc probability calibration. Shapley Additive exPlanations (SHAP), permutation, and impurity-based feature importance analyses converged on a consistent predictor hierarchy: number of constituent elements, mean ionisation potential, electronegativity range, maximum atomic number, and mean atomic radius, corroborated by polymorph-level diagnostics providing evidence that the model exploits structural, not merely compositional, information in most cases, although structural signal proved insufficient in at least one hard case. Applied to the full dataset, LightGBM ranked 424 entries as Tier 1 (P(Topological) ≥ 0.80) high-confidence topological predictions within the labelled dataset, of which 419 corresponded to already-known topological labels — a precision of 98.8% against known labels (retrospective validation precision, not prospective discovery precision). Interpretability analysis was also used as a data-quality diagnostic, motivating a closer investigation that, together with cross-validation against unit-cell volume and a subsequent ICSD/TQC provenance audit, uncovering and correcting a latent artefact in an earlier version of the model's most important feature. These findings demonstrate that pre-DFT structural and compositional descriptors carry physically plausible and interpretable predictive signal for topological classification, offering an inexpensive pre-screening tool to prioritise candidate materials for further investigation.