Composite structures often operate under complex conditions characterized by highly dissipative phenomena such as viscoplasticity, viscoelasticity or damage. This is especially true for those made of thermoplastic materials. These processes generate significant heat due to energy dissipation, which in turn affects the mechanical response of the composite. Consequently, the interaction between thermal and mechanical fields becomes critical, particularly under complex and cyclic loading conditions. Although traditional multiscale frameworks effectively capture these coupled thermomechanical behaviors [1], their high computational cost presents a significant barrier to their use in large-scale simulations. To overcome this limitation while preserving thermodynamic consistency, the Multiscale Thermodynamics Informed Neural Networks (MuTINN) framework was introduced in prior research. Originally designed to model elastoplastic behavior with damage in 2D woven composites, this study extends MuTINN to capture fully coupled nonlinear thermomechanical behavior in 3D composite structures. Grounded in thermodynamic principles, MuTINN integrates two specialized neural networks: one describing the evolution law and the other capturing the state law. Building on its core approach of representing history- and time-dependent behavior through macroscopic variables, the enhanced MuTINN model introduces additional parameters, including temperature, time, and entropy, to account for thermomechanical and viscoplastic effects. Validation of the enhanced model against mean-field Mori–Tanaka scheme combined with TFA theory [3] demonstrates its robustness and predictive accuracy across multiple material scales. Indeed, simulations of a composite tube subjected to complex cyclic loading reveal a significant reduction in computational cost compared to traditional FE² methods, without compromising precision.