A machine learning approach for enhancing process screening and qualification in metal additive manufacturing

The prevailing screening and qualification methodologies heavily depend on conventional manufacturing processes, which incur significant costs and prolonged lead times due to extensive physical testing. These challenges are also present in the growing field of additive manufacturing (AM), where numerous process parameters must be considered. However, the net-shape forming advantage of AM renders conventional screening and qualification methods inadequate. In the context of ongoing industrial digital transformation, a promising approach to enhancing process screening and qualification for metal AM is the adoption of a digital methodology tailored to the unique characteristics of this manufacturing technique. In this study, a convolutional neural network model is employed to extract features from images to predict material properties in laser-directed energy deposition (L-DED) processes. The model achieved a mean absolute percentage error of 2.3% and a root mean square error of 15.0 MPa for predicting ultimate tensile strength, with a prediction residual within 1% for density. Unlike conventional approaches that rely on bulk or multilayer builds, this study uniquely demonstrates the feasibility of using early-stage single-track print features to predict final part properties with limited view and material involvement. This established model and workflow pave the way for highly efficient and low-cost property prediction in L-DED processes.