In laser drilling, it is necessary to set appropriate laser parameters to obtain the desired hole. Therefore, it is necessary to study the effects of process variables on the machined hole parameters and establish a method to construct the optimal irradiation condition search. In this study, we established a method to predict the machined hole parameters by AI based on actual machining after investigating the distribution of machined hole parameters through laser drilling. Furthermore, by applying this research to the determination of irradiation conditions, it is possible to efficiently determine process variables when considering the desired irradiation conditions.

First, the effects of pulse width and pulse spacing of multiple pulses on the machined hole parameters were investigated. In this study, the parameter distribution of the machined hole diameter was evaluated by the response surface method by varying the pulse width of the CO2 laser. The results showed that the parameter distribution of the machined hole diameter changed significantly by changing the combination of several pulse widths.

Next, we proposed an AI-based method for predicting machined hole parameters with known process variables on a small number of data sets. Training a prediction model on a small data set is likely to lead to over-training, which increases the prediction error. In this study, we propose a method to reduce the prediction error of machined hole parameters by ensembling AI models in a prediction model using a stacking method. During training, the prediction model is trained using measured hole parameters, and during validation, the prediction of hole parameters is predicted by feeding back the prediction results in the model.

To verify the validity of this method, two models were compared. First, we checked how the prediction error changes with and without stacking. The proposed method with stacking had lower prediction accuracy, indicating that stacking improves the performance of the prediction model. Second, by inputting measured parameters into the forecasting model, we checked whether feeding back the values in the first place would lead to an improvement in the model. The results showed that models with measured parameters had lower prediction errors, and that including values that were closer to the actual measurements improved model performance. Therefore, the validity of stacking was demonstrated.

In addition, by using the response surface method in the prediction model, we proposed a method to visually search for process variables to obtain good machined hole parameters. By visually searching for irradiation conditions that obtain the desired machined hole parameters, the experimental conditions can be selected efficiently.