Finding the optimal solution for production conditions is not easy with regression model construction due to numerous explanatory variables, interactions, and experimental confounding, and it can be difficult to find improvement conditions. This report presents the results of applying Bayes optimization as an alternative method and verifying its effectiveness and limitations. Using Bayes optimization to search for welding conditions, we were able to find improvement conditions without constructing a regression model. In the example of metal content formulation to minimize casting defect rate, Bayes optimization was able to find conditions that achieved a lower defect rate than the generalized regression model with a similar number of runs. In cement particle size optimization, we were able to reach the target particle size of 20 in 2 to 8 experiments, even though the initial data was far from the target particle size of 20. In the example of Reactor yield optimization, the best yield rate obtained with Bayes optimization was nearly 20% lower than that of the regression model, suggesting that regression models including interactions may be more advantageous than Bayes optimization. In the context of improving production conditions, a large number of experiments like those in experimental design are unacceptable, but Bayes optimization presents improvement conditions sequentially, making it easier to implement. Moreover, each experiment may offer the potential to discover improvement conditions that could not be obtained with conventional methods.
