Property Prediction and Resilience Analysis in Polyionic Glasses Using a Machine‐Learning Algorithm with Multivariate Output

Abstract Designing the chemical composition of multi‐component glasses toward a set of target properties requires information on a complex range of trade‐off correlations. Consistent reference datasets are typically not available for this size, which will allow for the training of multi‐task neural network models. As an alternative, multi‐output Gaussian process regression (mGPR) is employed for multi‐task machine learning (ML) of combinations of glass properties from small, but highly consistent datasets, such as those typically generated in coherent laboratory campaigns. To this end, a dataset on polyionic glasses from the system of Na 2 O‐AlF 3 ‐Al 2 O 3 ‐P 2 O 5 ‐SO 3 is used, in which property resilience to variations in chemical composition is a particular challenge given the presence of multiple volatile species. Using the example of simultaneous predictions of refractive index and Young's modulus, it is demonstrated how mGPR outperforms single‐task modeling even when one of the physical properties is used as an additional output feature. A cloud of ≈6 . 10 7 samples is generated by random‐walk sampling in order to identify the compositional flexibility when aiming for a given set of target properties. In this way, trade‐off correlations between physical properties and chemical boundaries of compositional design are revealed.