Thermodynamically consistent viscoelastic constitutive artificial neural networks: Automating the pipeline from experimental data to finite element simulations

Viscoelastic constitutive artificial neural networks (vCANNs) leverage neural networks for data-driven modeling of the viscoelastic behavior of materials. Here, we propose a thermodynamically consistent extension of vCANNs that captures anisotropic, nonlinear, and time-dependent material behavior. A key strength of this approach is its ability to incorporate arbitrary auxiliary features—such as temperature, microstructural descriptors, or processing parameters—directly into neural constitutive laws. We propose an automated computational pipeline for the generation and implementation of such constitutive laws within the proposed framework into finite element (FE) simulations without manual model design. The proposed framework is validated across a broad range of representative material tests, including the nonlinear, thermo-viscoelastic response of soft polymers and arterial tissue with fiber dispersion. In addition, we demonstrate the accuracy and robustness in FE simulations using benchmark problems such as Cook’s membrane. The results underscore the flexibility, physical plausibility, and numerical stability of vCANNs as a powerful class of constitutive models for modern FE simulations enhanced by machine learning.