Physics-Informed Feature Engineering for Unsupervised Discovery of Metal-Insulator Transition Materials

Yüksel G. D., Sargın I.

E-MRS 2026 Fall Meeting, Warszawa, Polonya, 14 - 17 Eylül 2026, ss.1, (Özet Bildiri)

  • Yayın Türü: Bildiri / Özet Bildiri
  • Basıldığı Şehir: Warszawa
  • Basıldığı Ülke: Polonya
  • Sayfa Sayıları: ss.1
  • Orta Doğu Teknik Üniversitesi Adresli: Evet

Özet

The discovery of materials exhibiting metal to insulator transitions (MIT) underpins next-generation switching, sensing, and optoelectronic devices, but is limited by the cost of first principles screening and the absence of compact descriptors capturing correlated-electron physics. We present an unsupervised, physics informed feature pipeline for MIT compound discovery that operates directly on crystallographic input (CIF files), bypassing DFT calculations. The pipeline constructs its feature space from descriptors encoding the structural and electronic environment of the cation site: Goldschmidt tolerance factors, fractional ionicity, Madelung potentials, octahedral distortion and breathing metrics, estimated Hubbard U and charge transfer energies, Goodenough-Kanamori superexchange indicators, Harrison tight binding hopping integrals, crystal-field stabilization energies, and spin only magnetic moments. A Principal Component Analysis (PCA) projection with strict training projection separation resolves three interpretable clusters corresponding to metals, insulators, and MIT compounds for a curated dataset of approximately 340 correlated oxides. Cross validation with kNN, random forest, and gradient boosting classifiers confirms three-class separability with per class AUC between 0.8 and 0.9, serving as diagnostic validation of the unsupervised representation. Discovery is performed by unsupervised proximity analysis in this PCA space: unseen Materials Project candidates are scored via distance weighted k-nearest-neighbour voting against the labelled clusters, with bootstrap uncertainty bands, Mahalanobis and Isolation Forest domain checks, and nearest twin retrieval providing per candidate confidence. The analysis additionally surfaces training set compounds whose feature space position suggests likely mislabelling, together with borderline cases whose latent MIT potential could be activated by small chemical or structural perturbations. Posthoc SHAP analysis on the dominant PC directions traces predictions back to physical descriptors such as Hubbard U, charge transfer distance, bandwidth distortion coupling, and t2g filling, yielding chemistry interpretable insight into the metal-insulator boundary. Top candidates are consistent with the existing literature on MIT relevant compositions and stand ready for DFT verification. The framework demonstrates that domain aware unsupervised feature engineering on widely available crystallographic data can recover the correlated electron signatures of MIT behaviour without ab initio cost or supervised classifier training, offering a scalable, interpretable route for materials discovery in correlated oxides. The same strategy of building a compact set of physically informed dimensions, each descriptor chosen for its mechanistic role rather than added indiscriminately, transfers to other discovery problems where domain knowledge points to a meaningful, low dimensional representation.