Deep-learning-based surrogate modeling of laser-ultrasound testing for additive manufacturing

Context: on-destructive testing using a Laser ultrasound (LU) source is a promising technique for characterizing microstructures throughout the additive manufacturing of metallic components at the bead scale. The COLUMBO research project focuses on LU inspection applied to wire-laser additive manufacturing (WLAM) and the modeling of these processes. Objective: Modeling and numerical simulation can be used to predict microstructures and estimate surface displacements induced by a LU source, but not fast enough for near-real-time control and optimization of the WLAM process. For this purpose, however, they can serve as a basis for the conception of a surrogate model (aka, metamodel), i.e., a statistical model that is computationally cheaper than the original model, while providing a good approximation of the output quantities of interest. Approach: (i) Database: ensemble of 2D FE simulations via CIVA. Input: variable anisotropy and microstructure, the latter showing large elongated grains (typical of additive manufacturing). Output: B-scan based on a receiver line on the same surface as the LU source. (iii) High-dimensional data (~105), not easily reducible: major hurdle in surrogate modeling. (ii) Surrogate modeling using a deep neural network (DNN) with an innovative architecture combining convolutional encoder-decoder with additional Factorized Fourier layers. The latter are key building blocks of F-FNO architectures, which have shown good performance in similar applications: surrogate modeling of surface displacements due to seismic wave propagation in heterogeneous media.