The presenting author Audrey Gonzalo joined the Data Assimilation Research Team in the course of her French Master 1 degree at IMT-Atlantique, Brest, France, for 6 months under the RIKEN International Program Associates (IPA). This presentation provides a summary of the short- term IPA internship on the acceleration of Numerical Weather Prediction (NWP) models with Machine Learning (ML). NWP is based on a complex knowledge-based model to compute the high-resolution (HR) states of the atmosphere. However, the extensive computational cost for real-time prediction using the NWP model makes it paramount to both alleviate and accelerate the computation. In this regard, recent studies have shown the interest of ML, more specifically, Neural Networks (NN), to emulate non-linear processes at a lower computational cost. We propose to achieve model acceleration by degrading the NWP model resolution and to downscale the low-resolution (LR) outputs, resorting to a Convolution Neural Network (CNN). This method is commonly called Super-Resolution (SR) and finds a wide range of applications in various fields, but most recently geosciences. However, most of the methods are designed for RGB static images to address issues close to image decompression. In the case of NWP models, the data comprises underlying dynamics of physical processes, and the approximation of the LR-to-HR mapping function is fundamentally different from a deblurring task. In fact, our data encompasses multi-scale phenomena evolving at various scales. Due to the chaotic nature of the atmosphere, the HR features start diverging in different scales depending on phenomena, making the use of LR alone insufficient. Considering these specific features, we aim to adopt a cross-disciplinary approach bridging physics to these cutting-edge SR methods, by building a hybrid method merging both process-driven approaches of NWP and data-driven approaches of ML. In this study, we explore accelerating a HR NWP model known as the SCALE model, by applying a ML-based SR method to map the LR to the HR SCALE outputs. As a first step, we narrow the problem to only the precipitation variable at the 5000-m level from idealized SCALE simulations. CNNs require uncorrelated examples to train efficiently, so that we only select training data with independent HR small-scale features, since the main challenge of SR is to map accurately these fine-scale residuals. The model’s architecture embedding two input branches was designed using this key concept of scale-separation. Overall, our model shows a major improvement compared to the traditional bicubic interpolation. Further analysis like real-topography experiments may also raise discussions on its potential future integration within a real prediction system.