A new physics-based neural network framework has been developed to autonomously discover constitutive models for fully coupled thermomechanical systems. This innovative artificial intelligence approach diverges significantly from traditional methodologies that typically hinge on the Helmholtz energy function. Instead, the framework leverages internal energy and a dissipation potential as its primary constitutive functions, defining them in terms of material deformation and entropy. This deliberate architectural choice circumvents the intricate need to enforce mixed convexity-concavity, a pervasive challenge in classical thermomechanics formulations1. By simplifying this fundamental constraint, the system offers a more direct and computationally efficient pathway to accurately model how materials respond under combined thermal and mechanical loads. Such a capability is crucial for advanced engineering and materials science. This methodological improvement has implications for designing robust systems, enabling more precise simulations of material integrity, and ultimately influencing the development of resilient technologies foundational to critical infrastructure and national security applications.