Predicting oxide glass properties with low complexity neural network and physical and chemical descriptors

Due to disordered structure, glasses present a unique challenge in predicting the composition–property relationships. Here, we low-complexity neural network (LCNN) that provides improved performance properties of oxide glasses. In addition, combine LCNN with physical and chemical descriptors allow development universal models can provide predictions for components beyond training set. By on large dataset (∼50,000) glass components, show outperforms state-of-the-art algorithms such as XGBoost. interpret using Shapely additive explanations gain insights into role played by governing property. Finally, demonstrate universality new were not original Altogether, approach promising direction towards accelerated discovery novel compositions.