Quantitative prediction of microstructure and mechanical properties of additively manufactured titanium alloys using machine learning

The variety of thermal cycles occurring during metal Additive Manufacturing (AM) and the complexity of the process programming complicate the prediction of the final microstructure and mechanical properties of 3D-printed components. Therefore, the selection of process parameters in practical applications is very challenging, seriously hindering the widespread adoption of AM technology. In this work, a Machine Learning (ML) approach based on thermo-physical parameters and thermal history is proposed for the prediction of the final microstructures and mechanical properties of Direct energy deposited (DED) titanium alloys. The training space consists of 104 experimental data sets acquired under different process parameters and the correspondent thermal histories. A semi-automated measurement method is utilized to extract the microstructure features in the samples and a high-fidelity numerical simulation scheme is adopted to complement the data generation for the DED process. A Back-Propagation Neural Network (BP-NN) model is designed and trained with the obtained data sets. The model is also validated by comparing its outputs with the results documented in prior literature, revealing a mean deviation of 32.8 % for the width of α. It has been demonstrated that quantitative prediction of both microstructure features and mechanical properties for Ti6Al4V can be accomplished with the proposed model, providing possibility for its extension to other titanium alloys typically utilized in metal AM.