Latent-Energy-Based NNs: An interpretable Neural Network architecture for model-order reduction of nonlinear statics in solid mechanics

Nonlinear mechanical systems can exhibit non-uniqueness of the displacement field in response to a force field, which is related to the non-convexity of strain energy. This work proposes a Neural Network-based surrogate model capable of capturing this phenomenon while introducing an energy in a latent space of small dimension, that preserves the topology of the strain energy; this feature is a novelty with respect to the state of the art. It is exemplified on two mechanical systems of simple geometry, but challenging strong nonlinearities. The proposed architecture offers an additional advantage over existing ones: it can be used to infer both displacements from forces, or forces from displacements, without being trained in both ways.