Building a high-resolution machine learning weather model

After numerous successful applications of machine-learning-based global weather models, a new interesting direction of application is to seek high-resolution regional ML-based models that could complement high resolution numerical models serving day-to-day purposes. Development of such a model would combine speed and resource efficiency of ML models with high-resolution capabilities available so far only in the numerical models. Most ML-based models created so far are restricted to the resolution of underlying ERA5 data, often further downsampled due to various constraints, leaving substantial room for further research. With the objective of building a high-resolution ML model for Iceland and equipped with 30 years of 2-km reanalysis data covering Iceland and the surrounding ocean, we are exploring possibilities of the applications of existing ML architectures to our domain. The model we are currently building is based on ClimaX architecture from Microsoft, which we are modifying to best serve our objectives. Understanding the unique needs of regional models during training is one of the key factors in generating a successful regional model. While some of the architectures of the available global models can be applied directly to build a local model, many questions arise: do we need to adjust the cost function during training to handle domain boundaries? Which model levels should we prioritize during training — would it be better to focus on lower levels if the resolution is high and the timescale is short? To what extent can we use transfer learning (leveraging pre-trained weights from the global experiment) and how much will it guide the model toward the optimum? In this talk, we will discuss some of the above considerations for successfully running a regional model and present our high-resolution model for Iceland. The successful development of large machine-learning-based weather models has given weather and climate scientists confidence that models and reanalysis data built over decades are capable of capturing enough variability for ML-based inference. This now opens a new world of possibilities for model improvements and scientific advancements.