FAST-DREAM aims to enhance atmospheric composition (AC) modeling crucial for climate, weather, and air quality forecasting by overcoming computational challenges through machine learning (ML). Traditional AC modeling faces limitations in accuracy and spatial resolution due to computational demands, particularly from the advection operator. FAST-DREAM proposes ML emulation models as a solution to accelerate atmospheric advection simulations, offering speed, flexibility, and advanced hardware support. The project will address physical consistency, long-term accuracy, and scalability challenges using deep learning to develop a scalable, efficient Transformer-based model for predicting tracer concentration distribution. This model will incorporate dimensionality reduction for handling multiple tracers and ensure physical consistency. The ultimate goal is to extend the model to 3D inputs and multi-time step predictions, enhancing long-term forecasting accuracy while managing computational and memory demands through parallelization strategies.