One of the most prevalent additive manufacturing processes, the powder bed fusion process, is driven by a moving heat source that melts metals to build a part. This moving heat source, and the subsequent formation and moving of a melt pool, plays an important role in determining both the geometric and mechanical properties of the printed components. The ability to control the melt pool during the build process is a sought after mechanism for improving quality control and optimizing manufacturing parameters. For this reason, efficient models that can predict melt pool size based on the process input (i.e., laser power, scan speed, spot size and scan path) offer a path to improved process control.
Towards improved process control, a data-driven melt pool prediction model is built with a neighborhood-based neural network and trained using experimental data from the National Institute of Standards and Technology (NIST). The model considers the influence of both manufacturing parameters and laser scan paths. The scan path information is encoded using two novel neighborhood features of the neural network through locality. The neural network is used to generate a surrogate model, and we demonstrate how the performance of the resulting surrogate model can be further improved by using several ensemble methods. We then demonstrate how the trained surrogate model can be used as a forward solver for developing novel laser power design algorithms. The resulting laser power plan is designed to keep melt pool size as constant as possible for any given scan pattern. The algorithm is implemented and validated with numerical experiments.