Urban air pollution is a major concern that has a significant impact on both the environment and public health. In this context, machine learning has enabled the development of surrogate models, which have proven to be valuable tools for simulating atmospheric pollution. These models are a cost-effective and efficient way to replicate complex and computationally expensive atmospheric systems. They are particularly important in critical situations, such as accidental or malicious releases, where the ability to swiftly predict pollution consequences can make a difference in managing the emergency. In our previous work, we introduced the MCxM, a new approach to learning air pollution dispersion in urban environments. This framework allows for the prediction of 2D concentration fields by applying a series of masking and correction operations that gradually incorporate the influence of obstacles into the physics of pollutant transport and dispersion. In this paper, we propose to enhance the MCxM architecture by extending its capabilities to predict three-dimensional concentration fields and addressing the challenge of mass conservation of the model inferences. These enhancements involve a series of complementary preprocessing steps and the integration of new neural networks to scale and approximate the underlying physics. To validate the effectiveness of our approach, we used synthetic integrated concentration data generated by the PMSS modelling system, considering extensive twin experiments in the French cities of Grenoble and Paris. The results indicate that the accuracy and mass consistency of the predicted integrated concentration field have improved significantly compared to our previous work. This highlights the potential of the enhanced MCxM architecture to bring our predictions closer to real-world conditions, thus facilitating more effective pollution management strategies.