Machine Learning & Atmospheric Processes Group

Machine learning is an increasingly popular tool in the geosciences, offering new approaches to numerical weather prediction and complex dataset analysis. However, as reliance on these techniques grows, pressing questions about model transparency, internal biases, and trust emerge. While post-hoc explainability analyses can provide insights on how neural network (NN) outputs are generated, a robust framework for interpreting internal decision-making remains underdeveloped. We address this challenge by exploring a framework to better understand the inner structure of NNs using sparse autoencoders (SAEs). With simplified multilayer perceptrons (MLPs), we demonstrate that hidden layer neurons often exhibit polysemantic behavior where each feature is mapped to a linear combination of neurons, creating an overcomplete representation. This phenomenon, known as superposition, arises when networks encode more features than available neurons, causing neurons to respond to multiple, seemingly unrelated inputs. By introducing a regularized SAE that learns from the original MLP’s activations, we can disentangle these representations, resulting in a 33% reduction in the average number of sensitive inputs per neuron. Applied to a precipitation classification model, this framework reveals evidence of monosemantic behavior in which neurons respond to a single meaningful concept tied to specific physical phenomena such as temperature and fallspeed thresholds for precipitation phase partitioning. We observe similar monosemantic behavior in SAE activations from a snowfall rate regressor related to particle concentration intensity, and vertical radar structures. This framework supports the development of more physically consistent interpretations of hidden neuron activations and improved trust in operational ML models across the geosciences.

High-quality hydrometeor microphysical observations are essential for accurate precipitation estimates and for evaluating weather and climate models. However, analyzing these properties is challenging due to their high variability, complex interactions, and large data volumes. In this study, we examine more than 1.5 million minute-scale rain and snow particle attributes and collocated meteorological variables from seven global measurement sites over 9 years. Applying Uniform Manifold Approximation and Projection (UMAP) for nonlinear dimensionality reduction, we reduce the dataset’s dimensionality by 75%, identifying nine distinct precipitation groups and associated particle evolution pathways. UMAP effectively captures the global structure of precipitation phases such as rain, snow, and mixed-phase types, revealing clear patterns that linear methods struggle to resolve. The resulting UMAP manifold offers a unique perspective on precipitation phase and intensity, advancing our understanding of particle evolutionary processes and offering valuable insights for improving weather and climate models and remote sensing precipitation estimates. Analyzing rain and snow particles reveals hidden patterns, improving precipitation classification for better model predictions.

CloudSat’s Cloud Profiling Radar is a valuable tool for remotely monitoring high-latitude snowfall, but its ability to observe hydrometeor activity near the Earth’s surface is limited by a radar blind zone caused by ground clutter contamination. This study presents the development of a deeply supervised U-Net-style convolutional neural network to predict cold season reflectivity profiles within the blind zone at two Arctic locations. The network learns to predict the presence and intensity of near-surface hydrometeors by coupling latent features encoded in blind zone-aloft clouds with additional context from collocated atmospheric state variables (i.e., temperature, specific humidity, and wind speed). Results show that the U-Net predictions outperform traditional linear extrapolation methods, with low mean absolute error, a 38% higher Sørensen–Dice coefficient, and vertical reflectivity distributions 60% closer to observed values. The U-Net is also able to detect the presence of near-surface cloud with a critical success index (CSI) of 72% and cases of shallow cumuliform snowfall and virga with 18% higher CSI values compared to linear methods. An explainability analysis shows that reflectivity information throughout the scene, especially at cloud edges and at the 1.2-km blind zone threshold, along with atmospheric state variables near the tropopause, are the most significant contributors to model skill. This surface-trained generative inpainting technique has the potential to enhance current and future remote sensing precipitation missions by providing a better understanding of the nonlinear relationship between blind zone reflectivity values and the surrounding atmospheric state. Significance Statement Snowfall is a critical contributor to the global water–energy budget, with important connections to water resource management, flood mitigation, and ecosystem sustainability. However, traditional spaceborne remote monitoring of snowfall faces challenges due to a near-surface radar blind zone, which masks a portion of the atmosphere. In this study, a deep learning model was developed to fill in missing data across these regions using surface radar and atmospheric state variables. The model accurately predicts reflectivity, with significant improvements over conventional methods. This innovative approach enhances our understanding of reflectivity patterns and atmospheric interactions, bolstering advances in remote snowfall prediction.