Assessing the potential of machine learning for predicting the nonlinear refractive index of glasses

Abstract

Glass is a fundamental material in various applications, from telecommunication to healthcare and energy. Understanding the relationship between the glass matrix composition and the optical properties is crucial for developing advanced photonic materials. Machine learning (ML) provides a powerful tool to model complex, nonlinear relationships in materials science, enabling predictive insights that surpass conventional approaches. In this work, we explore the feasibility of five ML algorithms to predict the nonlinear refractive index (n2) of glasses based on their chemical composition. Data were compiled from the literature and analyzed separately as oxide and chalcogenide datasets. The predictor variables were preprocessed, and the ML models were tuned and evaluated using stratified train-test splits with leave-one-out cross-validation and nested cross-validation. Ensemble-based ML models, such as Random Forest Regressor (RF) and Gradient Boosting Regressor (GB), consistently outperformed others, with GB achieving the lowest Mean Absolute Percentage Error (MAPE) on oxides (28% ± 8%) and RF showing stable performance on chalcogenides (13% ± 3%). MLPRegressor and Support Vector Regression showed high variability, while Linear Regression performed poorly. These findings highlight the potential of ML-driven approaches for accelerating the discovery and optimization of nonlinear optical materials. However, the small dataset used in this study reflects a common limitation in the field, which restricts the models’ reliability and generalizability for predicting new data, potentially leading to overly optimistic performance metrics. The results should be considered preliminary, serving as proof-of-concept for the application of ML to nonlinear optical materials.