Machine Learning Enabled Exploration of Process-Structure-Property Relationships in Metal Additive Manufacturing

In this work we present machine learning (ML) guided optimization of metal additive manufacturing (AM) processes including directed energy deposition (DED) using both wire and powder feedstocks, and laser powder-bed fusion (L-PBF). Symbolic regression was applied to predict the fraction of ferrite in steel manufactured with L-PBF across a range of compositions and processing conditions which were later verified experimentally. Our feature importance analysis shows that the nickel to chromium ratio in steel plays a dominant factor in determining the ferrite content in AM steel, while laser power and scan speed plays relatively minor role. In DED, a combination of simulated and experimental data was used to train ML tools to predict bead geometry, including height, width, and depth, for a given set of processing conditions of laser power and wire feed rate. Our results enable improved control, part quality, and the development of closed-loop feedback systems for metal AM.