An image-based transfer learning approach for using in situ processing data to predict laser powder bed fusion additively manufactured Ti-6Al-4V mechanical properties

The mitigation of material defects from additive manufacturing (AM) processes is critical to reliability in their fabricated parts and is enabled by modeling the complex relations between available build monitoring signals and final mechanical performance. To this end, the present study investigates a machine learning approach for predicting mechanical properties for Ti-6Al-4V fabricated through laser powder bed fusion (PBF-LB) AM using in situ photodiode processing signals. Samples were fabricated under different processing parameters, varying laser powers and scan speeds for the purpose of probing a wide range of microstructure and property variations. Photodiode data were collected during fabrication, later to be arranged in image format and extracted to information-dense vectors by the transferal of deep convolutional neural network (DCNN) structures and weights pre-trained on a large computer vision benchmark image database. The extracted features were then used to train and test a newly designed regression model for mechanical properties. Average cross-validation accuracies were found to be 98.7% (r2 value of 0.89) for the prediction of ultimate tensile strength, which ranged from 900 to 1150 MPa in the samples studied, and 93.1% (r2 value of 0.96) for the prediction of elongation to fracture, which ranged from 0 to 17%. Thus, with high accuracy and hardware accelerated inference speeds, we demonstrate that a transfer learning framework can be used to predict strength and ductility of metal AM components based on processing signals in PBF-LB, illustrating a potential route toward real-time closed-loop control and process optimization of PBF-LB in industrial applications.