Simulation-based training of multi-material machine learning models for the prediction of melt pool dimensions and as-printed densities in laser powder fed fusion parts

Laser powder bed fusion (LPBF) is a powerful additive manufacturing technique, with numerous applications, yet the process optimization remains challenging due to the complexity of laser-matter interactions governing the melt pool dynamics and part densification phenomena. In this study, a commercial LPBF process simulation software, ANSYS ^® package, is used to build a multi-material (Al357, 17-4PH, IN718, Ti64, AlSi10Mg, IN625, 316 L and CoCr) training dataset comprising 7340 melt pool geometry and 2336 as-printed density values. This dataset is used to train three machine learning models, namely, Artificial Neural Networks, Support Vector Machines and Gaussian Process Regression. The models’ performances are assessed using experimental data obtained for five of eight simulated materials (Ti64, AlSi10Mg, IN625, 316 L, CoCr), and explored for two new, non-simulated, materials (Ti-Ni and W). The validation results demonstrated that while all the models could reproduce the results of the simulations used to generate the datasets, the Artificial Neural Networks manifest superior generalization and robustness, especially for extrapolating to new materials. The validation results also demonstrated that the performance of the machine learning models is logically limited by the quality of the numerically-simulated data. Furthermore, a novel methodology combining the Artificial Neural Network model predictions and the known dimensionless melt pool geometric ratios is proposed to accelerate the LPBF process optimization. This study highlights the potential of training LPBF machine learning models using simulation-based datasets to compensate for the limited availability of training data and therefore, facilitate the optimization of LPBF technology.