This work introduces a new data-based approach to Reynolds-Averaged Navier-Stokes (RANS) wall model for aerodynamic simulations at low Mach numbers. This new approach aims to replace the Dirichlet condition at the wall by a Neumann condition at the interface between the wall-modeled area and the fully resolved flow. Inspired by the Dirichlet-to-Neumann map concept, this method enforces the flux across the interface of the wall-modeled region by using a data-driven model consisting of two interconnected neural networks. The first network directly estimates the local skin friction, the second approximates the Dirichlet-to-Neumann mapping, evaluating the wall normal derivative of the velocity field at the RANS interface. Contrary to the data-based approach from, this new strategy is significantly cheaper and more robust since it does not require to solve a nonlinear equation through a Newton-like algorithm. Both neural networks are trained using data from fully-resolved RANS simulations of turbulent flows over various two-dimensional bump geometries. Different flow conditions are considered by varying the Reynolds number and the bump height (which changes the wall pressure gradients). Following classical dimensional analysis concepts, the training strategy aims at learning relations between relevant non-dimensional quantities, thus ensuring a better generality of the learned model. After training, the model capabilities are assessed across seen and unseen flow conditions over different bump configurations. Additionally, the NACA 4412 airfoil is considered as a final test case to demonstrate the robustness of our new wall model on a geometry significantly different from training configurations. Overall, this data-driven model demonstrates good robustness and accuracy, allowing the reference skin friction coefficient to be reproduced within an error of a few percentage points for all test cases considered, which significantly outperforms classical models such as Musker's wall law.