Harnessing of in situ processing data to predict mechanical properties of laser powder bed fusion AlSi10Mg

Although many studies have used machine learning models with in situ processing data to predict the properties and performance of laser powder bed fusion (PBF-LB) additively manufactured metallic parts, the ability to apply a single model to multiple materials or different processing parameter sets using transfer learning (TL) has not yet been established. In this paper, we evaluate the use of an image-based deep convolutional neural network (DCNN) TL model to predict the mechanical properties of AlSi10Mg and Ti-6Al-4V, which were sampled across two different sets of processing parameters, both of which were anticipated to produce large differences in porosity and mechanical properties. The predictions obtained from using a single sensor input were compared to those obtained after using multi-sensor data fusion techniques at the data-, feature-, and decision-level. Prediction accuracies as high as 97.4% for ultimate tensile strength and 95.8% for elongation to fracture were achieved after feature-level fusion with the DCNN-TL approach. The ability to reduce the training data sets during TL was also investigated, where selecting training data based on properties achieved an 94/89% prediction accuracy (for strength and elongation) with only 30% of the processing parameter sets compared to 92/84% prediction accuracy with random sampling. These findings not only confirm the ability to predict accurate properties for different PBF-LB materials and processing parameters with TL approaches, but also provide insight into the reduced amount of data, time, and computational cost, required to achieve accurate predictions when fusing multiple sources of in situ processing data.