Inverse design of high-performance glasses through an encoder-decoder machine learning approach toward materials discovery: Application to oxynitride glasses

Machine learning approaches have shown remarkable potential in elucidating the complex, nonlinear relationships between compositional effects, structural configurations, and macroscopic properties in materials science. Nevertheless, their application within glass science for guiding the inverse design of formulations with properties tailored for high-performance applications remains limited. In the present work, we report the implementation of an autoencoder within glass science. To address scenarios where the dimensionality of the input exceeds that of the output, a case study was selected wherein 34-dimensional formulations were projected into a 4-dimensional latent space through an autoencoder’s encoder. This strategy thereby enabled the prediction of formulations (with up to 34 different elements) from four material parameters (Young’s modulus, hardness, glass transition temperature, and specific mass), whereby an artificial neural network ( ANN ) links these parameters to the latent representation, and the decoder subsequently reconstructs the formulation in its original 34-dimensional form. Two promising candidate compositions generated through this framework, upon synthesis and experimental characterization, yield glasses that rank among the stiffest, hardest yet light large-scale inorganic, non-metallic systems amongst reported literature, exhibiting a unique combination of high Young’s modulus (E = 155 GPa) and hardness (H = 12.2 GPa) at a comparatively low specific mass (ρ = 3.62 g·cm −3 ). These findings emphasize the potential of machine learning methodologies for the systematic quantification and subsequent rationalization of composition-property trends based on physicochemical descriptors, offering novel insights both within and across chemical systems – demonstrated here for oxynitride glasses, where the processing- and content-dependent incidence of nitrogen is revealed.