The construction of accurate and computationally affordable computer models for simulating non-equilibrium fluid flows is a challenging task, particularly when dealing with hypersonic applications. The popular Navier-Stokes (NS) model yields unreliable predictions whenever the continuum assumption fails (approx. for a Knudsen number Kn > 0.1). On the other hand, the large computational cost entailed in solving the more suited Boltzmann equation makes its numerical solution impossible for non-trivial cases. In such cases, even the most efficient Direct-Simulation Monte Carlo (DSMC) solution methods may result prohibitively expensive, in particular if approx. Kn < 10 and if thermochemical nonequilibrium is considered.This paper focuses on improving the computational capabilities for predicting nonequilibrium hypersonic flows in the transitional regime 0.1 < Kn < 10. Namely, we build physics-constrained Deep Learning (DL) augmentation terms that take advantage of the known and resolvable physics e.g., first principles, thermodynamics laws, and conservation laws, and exploit them to improve the accuracy of the NS solution. The aim is to achieve a prediction accuracy comparable to that of the DSMC solution while maintaining a computational cost comparable to solving the standard NS equations.Previous studies employed Deep Neural Networks (DNNs) for enhancing the modeling of incompressible fluid flows, with particular reference to turbulent phenomena, but their potential has not been fully deployed for hypersonic applications, for which literature concerning DL-DNNs augmentation models is still scarce. Recent research by Sirignano et al. [1] proposed an innovative online training strategy that leverages the adjoint method to optimize the parameters of a DL model by solving a constrained PDE. This approach allows consistent mathematical-physical training and was successfully applied to hypersonic problems by Nair et al. [2] for a 1D shock flow in the transitional regime. However, this model does not embed the first principle of physics, making it case-dependent and preventing its extension to general 2D/3D applications.The present work aims to overcome these issues and extend the state-of-the-art by proposing a flexible framework for building physics-constrained neural network closures of general application. To overcome these drawbacks, the pursued approach is based on the premise that physical laws and constraints are inherently embedded in real applications and they can be exploited to guide the development of more robust and accurate closures. By incorporating these constraints into the training process, we aim to develop a model that can capture both the universal physics, which is general, and the physics encoded in the data, which is case-dependent. It is important to note that the main advantage of the approach proposed in this work, compared to other alternative approaches, is its ability to generalize: the structure of the developed model is case-independent and permits the use of the same model on very differ...