Multi-objective Bayesian optimization of optical glass compositions

Optical glass must satisfy multiple specifications for a given optical instrument. Machine learning has recently been applied to the development of materials. The optimization of compositions with multiple target properties has been investigated. Here, we focus on the trade-off between target properties, namely the refractive index nd and the Abbe number ?d, for optical glass and apply a multi-objective Bayesian optimization algorithm called ParEGO to find glass compositions with low target property values using INTERGRAD data. In addition, we examine the effects of normalizing the target properties and descriptors on the search performance of ParEGO. The results reveal that ParEGO can effectively find compositions with low target property values and that normalization is necessary for sufficient ParEGO search performance. The normalization of target properties and descriptors may be effective for other multi-objective Bayesian optimization algorithms and materials. We also discuss how to apply the method to the actual development of optical glass.