2024 Volume 12 Issue 2 Article ID: 23-15051
A neural network is required to be coordinate system independent when it is applied to solve physical problems. Instead of training with augmented data, this paper develops a post-processing which is applicable for any arbitrary neural network. The essence of the post-processing is synthesized rotation that generates a set of rotated input data, applies the neural network to each of the set, re-rotates the output, and extracts a coordinate system independent solution from the output. Numerical experiments of the post-processing are carried for structural response analysis. It is shown that the post-processing works as designed. Discussions are made for further use of the post-processing of neural networks to make them more physically admissible.
