Reconfigurable Computing in Neural Network‐Enhanced Finite Element Analysis in Structural Dynamics

ABSTRACT In the present study, artificial neural networks (ANNs) trigger a better physically plausible initial guess in the Newton–Raphson scheme of an implicit material integration scheme of rate‐dependent viscoplastic dynamic response of plate structures in an in‐house finite element (FE) solver. The trained neural network acts as an offline solver that predicts a better initial guess of the Newton–Raphson scheme, thereby reducing the iterations and the computational complexity of the classical root‐finding algorithm. The reduction of the Newton iterations mitigates the uneven branching in online iterative loops by replacing them with more fixed forward‐pass computations, enabling a more efficient vectorized implementation of the implicit material integration and a more effective exploitation of CPU‐level single instruction multiple data (SIMD) hardware. Furthermore, energy‐efficient field programmable gate array (FPGA) hardware is programmed to accelerate the neural network architecture of the material integration scheme, and inference time is compared with CPU and GPU implementations. Essentially, the FPGA is configured and tailored specifically for the trained ANN that accounts for the material nonlinearities. The motivation for employing ANNs in FPGAs lies in the possibility of tailoring and reconfiguring the hardware specifically for the desired neural network architecture, leading to a more sustainable neural network inference in terms of energy consumption and computational effort. We propose this framework for impulsively loaded plates exhibiting strain‐rate‐dependent plasticity. Results are compared with the classical FE solver, and the neural network inference on FPGA is compared to CPUs and GPUs, showing that FPGA outperforms CPU by and GPU by .