Machine Learning-Driven R&D of Perovskites and Spinels: From Traditional Models to Deep Learning

Area of Science:

• Materials Science
• Data Science
• Artificial Intelligence

Background:

• Traditional trial-and-error methods for developing strategic materials (e.g., spinels, perovskites) are inefficient and time-consuming.
• Machine learning (ML) presents a data-driven alternative to accelerate materials discovery and design.

Purpose of the Study:

• To review the evolution of ML in materials science, from traditional ML (TML) to deep learning (DL).
• To assess DL's advantages in autonomous feature extraction and its application in forward screening and inverse design.
• To explore strategies for overcoming data scarcity and enhancing ML model reliability in materials discovery.

Main Methods:

• Analysis of the progression from TML with manual feature engineering to DL with end-to-end feature extraction.
• Evaluation of DL models for mapping atomic structures to material properties.
• Discussion of advanced techniques for data augmentation and uncertainty quantification.

Main Results:

• DL models demonstrate superior precision in predicting material properties compared to TML.
• Addressing data scarcity and enhancing model robustness are critical for practical applications.
• Bayesian learning and confidence-aware modeling improve the reliability of AI-guided materials design.

Conclusions:

• Future materials discovery requires a shift towards reliable and robust ML models, not just predictive accuracy.
• Developing interpretable AI models and standardized databases is crucial.
• Integrating AI with automated experimentation will create a closed-loop research ecosystem for accelerated scientific discovery.