arXiv subjects

George C. Craig

Publications and source records attributed to George C. Craig.

3 recordsLinked to original sources

arXiv2026

Using Distributional Regression Networks to Retrieve Cloud Properties from Solar Satellite Channels for Data Assimilation

Satellite observations in the solar spectrum (including visible and near-infrared channels) offer high-resolution information on clouds and atmospheric properties valuable for data assimilation. While forward operators for a direct assimilation of solar images have become available recently and a first visible channel is already used operationally, their assimilation remains challenging due to strong non-linearities, ambiguities and high inter-channel correlations. This study addresses two central questions: what is the potential impact of assimilating multiple solar channels jointly, and can observed reflectances be transformed into physically meaningful, uncertainty-quantified variables better suited to assimilation than the raw reflectances themselves? As a proof of concept, we assess the joint information content of six solar channels from the Flexible Combined Imager (FCI) onboard Meteosat Third Generation and introduce a novel "Backward Operator" (BO) for probabilistic retrievals of cloud-related variables. The BO is implemented in a machine learning approach as a distributional regression network that is trained on synthetic images from a NWP regional model run and produces multivariate Gaussian estimates of total optical thickness, column cloud fraction, ice fraction, and effective radii of water and ice. The BO predictions are unbiased and well-calibrated, with realistic, situation-dependent and non-trivial covariance structures. The retrieved variables can be overall usefully constrained. Despite strong inter-channel correlations, combining multiple channels yields substantial performance improvements. As the BO does not require prior information, is consistent with an existing forward operator, and yields cloud variables more linearly related to the NWP model state, assimilating these variables could be a viable alternative to direct reflectance assimilation.

physics.ao-ph↗

arXiv2026

Mechanistic Interpretability Tool for AI Weather Models

Artificial Intelligence (AI) weather models are improving rapidly, and their forecasts are already competitive with long-established traditional Numerical Weather Prediction (NWP). To build confidence in this new methodology, it is critical that we understand how these predictions are generated. This is a huge challenge as these AI weather models remain largely black boxes. In other areas of Machine Learning (ML), mechanistic interpretability has emerged as a framework for understanding ML predictions by analysing the building blocks responsible for them. Here we present an open-source, highly adaptable tool which incorporates concepts from mechanistic interpretability. The tool organises internal latent representations from the model processor and allows for initial analyses, including cosine similarity and Principal Component Analysis (PCA), enabling the user to identify directions in latent space potentially associated with meteorological features. Applying our tool to the graph neural network GraphCast, we present preliminary case studies for mid-latitude synoptic-scale waves and specific humidity. These demonstrate the tool's ability to identify linear combinations of latent channels that appear to correspond to interpretable features.

physics.ao-ph↗

arXiv2024

Improving forecasts of precipitation extremes over Northern and Central Italy using machine learning

The accurate prediction of intense precipitation events is one of the main objectives of operational weather services. This task is even more relevant nowadays, with the rapid progression of global warming which intensifies these events. Numerical weather prediction models have improved continuously over time, providing uncertainty estimation with dynamical ensembles. However, direct precipitation forecasting is still challenging. Greater availability of machine learning tools paves the way to a hybrid forecasting approach, with the optimal combination of physical models, event statistics, and user-oriented post-processing. Here we describe a specific chain, based on a random forest pipeline, specialised in recognizing favourable synoptic conditions leading to precipitation extremes and subsequently classifying extremes into predefined types. The application focuses on Northern and Central Italy, taken as a testbed region, but is seamlessly extensible to other regions and timescales. The system is called MaLCoX (Machine Learning model predicting Conditions for eXtreme precipitation) and is running daily at the Italian regional weather service of ARPAE Emilia-Romagna. MalCoX has been trained with the ARCIS gridded high-resolution precipitation dataset as the target truth, using the last 20 years of the ECMWF re-forecast dataset as input predictors. We show that, with a long enough training period, the optimal blend of larger-scale information with direct model output improves the probabilistic forecast accuracy of extremes in the medium range. In addition, with specific methods, we provide a useful diagnostic to convey to forecasters the underlying physical storyline which makes a meteorological event extreme.

physics.ao-ph↗