Deep Material Network (DMN) has emerged as a data-driven surrogate model for heterogeneous materials. Compared to other neural networks, DMN distinguishes itself by its capability to directly encode the morphology of a particular microstructure through its physically interpretable parameters. Instead of fitting a particular physical (e.g., stress–strain) response, it learns the microstructure per se. We have recently presented a parametric DMN model for fiber composites with a varying parameterized microstructure [1]. A single-layer feedforward neural network is used to account for the dependence of DMN fitting parameters on the microstructural characteristics such as fiber volume fraction, aspect ratio and orientation. Micromechanical constraints are prescribed both on the parametric architecture and the outputs of this new neural network. Offline training is performed by minimizing a loss function that aggregates solely the linear elastic homogenization data generated across various morphologies. During online inference, the trained model accurately extrapolates to history-dependent complex thermomechanical behaviors such as (visco)plasticity, with different microstructural parameters. The proposed parametric DMN model has been implemented as a built-in user material available in Abaqus/Standard [2]. The rotation-free formulation is adopted thanks to its significant efficiency improvement. The offline training data as well as the reference online prediction results are obtained from Abaqus finite element simulations using the micromechanics plugin. Numerical simulations demonstrate that our parametric DMN is able to accurately predict the effective linear and nonlinear thermomechanical responses, even when physical properties are temperature-dependent. It successfully captures the structure-property relationships, demonstrating satisfactory generalization capabilities when the microstructural morphology varies. Energetic consistency, strain increment insensitivity and numerical robustness are verified under highly nonlinear loading conditions. Moreover, significant speed-ups are achieved compared to full-scale finite element analysis. This opens possibilities for its application in component level simulations of composite structures.