Evaluating the performance of active learning models on selective laser melting data sampling

Data acquisition in additive manufacturing, specifically selective laser melting, is always expensive and worsens when the material under study is also costly. To address this challenge, researchers use available design of experiment (DoE) tools. This marks a shift away from trial-and-error and one-factor-at-a-time approaches, which are ineffective and cause the number of required experiments to grow exponentially as the number of parameters increases. However, the traditional design of experiments struggles in analysing complex, multi-parameter, and noisy systems, inherent characteristics of selective laser melting data. Active machine learning can excel in this limited data and sophisticated field. This study evaluates the performance of active machine learning models based on neural networks and Gaussian process regression (GPR) with a D-optimal design for predicting Ti-5-5-5-3 and Beta 21S samples. This study demonstrates the ability of both active learning methods to reduce data required with improved predictability, with GPR outperforming the others. These results demonstrate the potential of GPR for effective SLM experimentation and emphasize the necessity of improving active learning based on neural networks to increase experimental accuracy with less data.