Discovering convection biases in global km-scale climate models using computer vision

Convective clouds are a key component of the climate system, impacting the hydrological cycle, and leading to the redistribution of heat, moisture, and momentum. Traditional low-resolution climate models rely on parameterisations to represent convection and thus struggle to realistically capture convective processes. In contrast, km-scale models can directly simulate deep convection, improving the accuracy of cloud and precipitation fields. However, significant uncertainties remain, due to parameterisations of remaining unresolved subgrid-scale processes, which must be addressed.Traditional model evaluation methods rely on aggregated spatial and temporal statistics, which overlook the fine-grained details critical to understanding the physical processes underlying convection. In addition, conventional dimensionality reduction techniques (e.g., principal component analysis) cannot capture the non-linear relationships of small-scale physical processes.To address these limitations, we use computer vision models to learn meaningful low-dimensional embeddings of outgoing longwave radiation (OLR) fields and evaluate km-scale models in this new embedding space. More specifically, we use contrastive learning, a self-supervised technique that trains machine learning models to distinguish between similar and dissimilar data points, to train a deep neural network to generate compact representations of OLR fields.We present results from a case study evaluation of two km-scale models, the Integrated Forecasting System (IFS) and the Icosahedral Nonhydrostatic Model (ICON), developed as part of the nextGEMS project. The simulations are compared to observations from the Geostationary Operational Environmental Satellites (GOES-16). We quantitatively assess the realism of km-scale models by comparing the embedding distributions of models and observations using vector quantisation. Finally, we use explainability methods to identify key factors influencing the accuracy of simulated convection. Our results highlight the value of our approach in understanding and improving the performance of high-resolution climate models, contributing to more reliable climate projections at finer spatial scales.