Thermodynamics-based Artificial Neural Networks for constitutive modeling

Machine Learning methods and, in particular, Artificial Neural Networks (ANNs) have demonstrated promising capabilities material constitutive modeling. One of the main drawbacks such approaches is lack a rigorous frame based on laws physics. This may render physically inconsistent predictions trained network, which can be even dangerous for real applications. Here we propose new class data-driven, physics-based, neural networks modeling strain rate independent processes at point level, define as Thermodynamics-based (TANNs). The two basic principles thermodynamics are encoded network's architecture by taking advantage automatic differentiation to compute numerical derivatives network with respect its inputs. In this way, free-energy, dissipation and their relation stress internal state variables hardwired network. Consequently, our does not identify underlying pattern thermodynamic during training, reducing need large data-sets. Moreover training more efficient robust, accurate. Finally important, remain thermodynamically consistent, unseen data. Based these features, TANNs starting physics-based networks. We demonstrate wide applicability elasto-plastic materials, hardening softening. Detailed comparisons show that outperform those standard ANNs. ' general, enabling applications materials different or complex behavior, without any modification.