Predicting the long-term performance of cementitious composites requires an integrated understanding of mechanics, thermodynamics, chemistry, and transport phenomena, whose coupled and cumulative contributions govern material behavior. Conventional constitutive models rely on predefined energy potentials, assumed dissipation mechanisms, and prescribed rheological representations, which restrict their applicability for investigating the coupled, time-dependent response of concrete. These assumptions are particularly limiting for phenomena such as creep–fatigue, where material response emerges from strongly coupled multiphysics interactions that may originate from mechanisms with negligible short-term influence. In such systems, effects that appear secondary at early times—such as minor chemical reactions, weak moisture gradients, or small thermal fluctuations—can accumulate and ultimately dominate macroscopic behavior. Capturing these effects requires a unified constitutive framework in which mechanical, thermal, chemical, and transport processes are modeled in a fully coupled manner rather than imposed through prescribed laws. This work presents a by-construction, thermodynamically informed machine learning framework for constitutive modeling within a Lemaitre-type thermodynamic setting [1]. Admissible energy storage and dissipation are embedded directly into the neural architecture without assuming specific rheological laws or predefined mechanical analogs. Physical consistency is enforced at the architectural level rather than imposed a posteriori: constrained output representations ensure thermodynamic admissibility, and the neural network structure is designed such that the Clausius–Duhem inequality is satisfied by construction through non-negative dissipation. The framework is demonstrated using one-dimensional viscoelastic behavior associated with concrete creep as a representative example. The objective is not to propose a universal constitutive theory, but to illustrate a direction for data-driven constitutive modeling in which physically admissible time-dependent behavior—such as creep, relaxation, and hysteresis—emerges directly from data without prescribing classical rheological models. Within this scope, the framework yields physically consistent predictions and meaningful extrapolation beyond the training domain. By transforming neural networks from black-box predictors into interpretable, physics-aware constitutive operators, this approach