Compositional design of oxide glass with multi-target performances

To address the challenges of multi-objective property optimization in oxide glass design, here we developed an interpretable machine learning framework leveraging the SciGlass database to predict glass compositions with tailored combinations of refractive index (n), softening temperature (T_littletons), and thermal expansion coefficient (CTE). Three distinct algorithms—Random Forest (RF), Linear Regression (LR), and Multilayer Perceptron (MLP)—were systematically trained to capture composition-property relationships. By defining oxide component constraints and permissible concentration ranges, the framework generated a combinatorial space containing ∼500 million potential formulations. Due to its superior predictive accuracy and robustness, the optimized RF model was selected to conduct property predictions across this extensive sample space. Through iterative filtering (n ≤ 1.5, CTE<8.7 × 10⁻⁶/ °C, and T_littletons between 700–790 °C), 516,325 candidate formulations meeting stringent multi-property criteria were identified. Experimental validation of four compositionally distinct glasses confirmed the model's predictive reliability, with the maximum deviation < 5 % from theoretical property values. This methodology enables concurrent optimization of optical, thermal, and mechanical properties, significantly accelerating the development cycle of functional glasses while reducing traditional trial-and-error costs.