Adaptive process control or online optimization in laser cutting requires a model that computes the through-cut probability from the current process parameters in real time, in order to protect the system from failed cuts and to ensure complete material separation.

In this work, data-driven models are developed that predict the through-cut probability from the process parameters laser power, feed rate, focus distance, nozzle distance, gas pressure, and sheet thickness. The database comprises several thousand laser cutting experiments across multiple sheet thicknesses, systematically covering the parameter space using statistical design of experiments.

Several classification models are compared in a stratified cross-validation, including logistic regression, random forest, gradient boosting, and a neural network. The neural network achieves the highest prediction accuracy across all metrics. For more than ninety percent of the predictions, the confidence exceeds eighty percent, with the error rate in this range remaining below three percent. In a leave-one-thickness-out validation, where one sheet thickness is entirely excluded from training and used as the test set, the model demonstrates reliable generalization to unseen material thicknesses.

Furthermore, parameter regions are identified in which the process exhibits stochastic behavior and identical settings lead to different outcomes. These zones mark a physical limit of predictability regardless of the chosen modeling approach. Due to the short inference time in the millisecond range, the models can be integrated into a closed-loop control system, deployed as a surrogate model in Bayesian online optimization, or used for offline parameter selection prior to the cutting process.