Machine Learning Approaches for Predicting Structural Properties of Oxide Thin Films
Article Sidebar
Main Article Content
Abstract
thin films exhibit structural characteristics—such as lattice parameters, strain states, defect concentrations, and film thickness—that critically influence their functional behavior in microelectronics, optoelectronics, catalysis, and spintronics. However, conventional experimental techniques and first-principles computations remain resource-intensive for mapping wide structural parameter spaces. This study develops a machine-learning (ML) framework for accurate prediction of structural properties of oxide thin films using multimodal datasets sourced from experimental literature, materials repositories, and first-principles simulations. Feature engineering incorporated domain-specific descriptors, including ionic radii mismatch, electronegativity differences, Goldschmidt tolerance factors, strain tensors, and thermodynamic deposition conditions. Multiple ML algorithms Random Forests (RF), XGBoost, Support Vector Regression (SVR), Artificial Neural Networks (ANN), and Gaussian Process Regression (GPR) were trained and evaluated. XGBoost achieved the highest prediction accuracy for lattice constants (MAE = 0.014 Å, R² = 0.985), while GPR provided the most reliable strain prediction under limited data regimes (RMSE = 0.18%). Defect concentration prediction showed strong performance using RF (R² = 0.912), and film thickness prediction yielded RMSE = 1.25 nm across varied deposition techniques. SHAP-based feature importance revealed that ionic radii mismatch, oxygen partial pressure, substrate lattice mismatch, and deposition temperature had the strongest influence on predicted properties. Model predictions exhibited strong agreement with independent experimental reports and DFT benchmarks, confirming generalizability across perovskite, spinel, and binary oxide families. The study demonstrates that ML-driven structural prediction can significantly accelerate materials design and thin-film optimization, enabling inverse design of deposition conditions and reducing reliance on exhaustive experimentation.
Article Details
References
Jain, A. et al., Materials Project Database.
Curtarolo, S. et al., AFLOW Materials Repository.
King, D. et al., “Machine Learning for Materials Structure Prediction,” npj Computational Materials.
Butler, K. T. et al., “Machine Learning for Molecular and Materials Science,” Nature, 2018.
Ward, L. et al., “A General-Purpose ML Framework for Predicting Material Properties,” npj Computational Materials.
Choudhary, K. et al., “High-Throughput DFT and ML for Materials,” Scientific Data.
Hwang, H. Y., Iwasa, Y., Kawasaki, M., Keimer, B., Nagaosa, N., & Tokura, Y. (2012). Emergent phenomena at oxide interfaces. Nature Materials, 11, 103–113. https://doi.org/10.1038/nmat3223. Nature
Ohtomo, A., & Hwang, H. Y. (2004). A high-mobility electron gas at the LaAlO₃/SrTiO₃ heterointerface. Nature, 427, 423–426. https://doi.org/10.1038/nature02308. Nature
Schlom, D. G., Chen, L.-Q., Pan, X., Schmehl, A., & Zurbuchen, M. A. (2007). Strain tuning of ferroelectric thin films. Annual Review of Materials Research, 37, 589–626. https://doi.org/10.1146/annurev.matsci.37.061206.113016. mmm.psu.edu
Butler, K. T., Davies, D. W., Cartwright, H., Isayev, O., & Walsh, A. (2018). Machine learning for molecular and materials science. Nature, 559, 547–555. https://doi.org/10.1038/s41586-018-0337-2.
Bibes, M., & Barthélémy, A. (2007). Oxide spintronics. IEEE Transactions on Electron Devices, 54(5), 1003–1023.
Zubko, P., Gariglio, S., Gabay, M., Ghosez, P., & Triscone, J. M. (2011). Interface physics in complex oxide heterostructures. Annual Review of Condensed Matter Physics, 2, 141–165.
Schlom, D. G., Chen, L., Pan, X., Schmehl, A., & Zurbuchen, M. A. (2008). A thin film approach to engineering functionality into oxides. Journal of the American Ceramic Society, 91(8), 2429–2454.