Developing a machine-learning-assisted inherent strain method for fast and reliable residual stress and distortion prediction in LPBF components

This project aims to develop a machine-learning-assisted (MLA) inherent strain (IS) method for fast and accurate residual stress and
distortion prediction for real-size Hastelloy X components manufactured through laser powder bed fusion (LPBF). Currently employed
IS methods for LPBF deviate from the original IS framework and come with limitations as they assume a constant thermomechanical
history all over the part. Instead, we propose to employ artificial neural networks (ANNs) to learn the relationship between site-specific
IS values and geometrical features of the part first based on simulations. The ANN aims to approximate the non-uniform IS field of the
entire component as unique source of residual stresses. This proposed method is in line with the original framework behind the IS theory.
Once established, the proposed framework brings a drastic reduction in computational cost. The approval of this beam time request
allows to additionally train the ANN based on experimental data, and thus, it is expected to improve the reliability and accuracy of
the proposed method.