Nowadays, Computational Fluid Dynamics (CFD) is a fundamental tool for industrial design. However, the computational cost of doing such simulations is expensive and can be detrimental for real-world use cases where many simulations are necessary, such as the task of shape optimization. Recently, Deep Learning (DL) has achieved a significant leap in a wide spectrum of applications and became a good candidate for physical systems, opening perspectives to CFD. To circumvent the computational bottleneck of CFD, DL models have been used to learn on Euclidean data, and more recently, on non- Euclidean data such as graphs and manifolds, allowing much faster and more efficient surrogate models. Nevertheless, DL presents the intrinsic limitation of extrapolating out of training data distribution. In this study, we present a pioneer work to increase the generalization capabilities of Deep Learning by incorporating the physical gradients (derivatives of the outputs w.r.t. the inputs) to the models. Our strategy has shown good results towards a better generalization of DL networks and our methodological/ theoretical study is corroborated with empirical validation.