Proposing Optimum Composition of CAD/CAM Resin Composites via Bayesian Optimization

Objectives: We developed an AI model of the gradient boosting decision tree (GBDT) to predict the flexural strength of CAD/CAM resin composite blocks (RCBs), and conducted an exhaustive search to explore effective composition of CAD/CAM RCBs resulting in a higher predicted flexural strength. However, exhaustive search requires a significant computational cost. Bayesian optimization (BO) is an approach to optimize time-consuming objective functions. Instead of predicting the flexural strength for each composition, the BO method could propose candidate compositions with high likelihood of improving flexural strength. Therefore, the purpose of this study was to develop a Bayesian optimizer to reduce the time required for exhaustive search in finding effective composition of CAD/CAM RCBs.
Methods: The components and filler contents of CAD/CAM RCBs were defined using either one-hot encoding or numeric representations to instantiate the compositional space. BO was conducted, utilizing either Gaussian process (GP) or random forest as the surrogate model, and employing expected improvement (EI), probability of improvement (PI), upper confidence bound (UCB), and Thompson sampling as the acquisition functions. The time to identify the highest predicted flexural strength among compositions was compared between exhaustive search optimized by BO and without BO.
Results: With five different initializations, optimization with GP as surrogate model and EI, PI or UCB as acquisition function was able to identify the composition that resulted in a flexural strength of 269.5 MPa within 10 iterations. GP with EI exhibited the best performance, with an average time of 4.67 hours, which was significantly lower than 47.83 hours required for exhaustive search (Student t-test, p<0.05).
Conclusions: A Bayesian optimizer utilizing GP as the surrogate model and EI as the acquisition function was developed. This optimizer reduced the time for GBDT model to identify effective composition of CAD/CAM RCBs compared to exhaustive search, regardless of the initialization data.