Technical Summary: ClimateSOM

Problem Statement

Ensemble datasets are increasingly critical in climate science for capturing variability in projections under future conditions (e.g., varying greenhouse gas emissions). However, analyzing these datasets remains challenging due to their sheer size and complexity. Existing visual analytics methods often rely on summary statistics (e.g., spatial means) or topology-based methods that may overlook key spatial and temporal structures or fail to capture the distributional nature of ensemble members. Furthermore, current tools often struggle to provide both interpretability and robustness when dealing with highly complex, chaotic model behavior, making it difficult to compare or cluster multiple model runs effectively.

The authors identify three essential tasks for climate scientists working with ensemble data:

1. Exploration: Understanding the general behavior of a single set of model runs.
2. Comparison: Analyzing differences between two sets of model runs.
3. Clustering: Identifying meaningful groupings of model runs or Global Climate Models (GCMs) based on their behavior.

Methodology: The ClimateSOM Workflow

ClimateSOM is a visual analysis workflow that abstracts spatiotemporal climate ensemble model runs into a distribution over a user-steerable 2D space. The workflow consists of a pre-step and three main interactive steps:

1. SOM Training (S0)

The system trains a Self-Organizing Map (SOM) on the flattened ensemble dataset (across all spatial realizations). The SOM projects high-dimensional climate data (spatiotemporal time series) onto a 2D grid while preserving local topological relationships.

• Abstraction: Instead of treating a model run as a sequence, ClimateSOM treats it as a distribution. For each time step in a model run, the system identifies the Best Matching Unit (BMU) in the SOM. The collection of these BMU locations forms a 2D distribution representing the model run.
• Design Choice: Temporal ordering is intentionally discarded during this abstraction to emphasize broad spatial patterns and reduce complexity, aligning with how GCM outputs are often studied for general differences.

2. Anchor (S1)

While SOM grids are locally smooth, they can suffer from global distortions. To enhance interpretability, users can "anchor" specific SOM nodes to fixed positions.

• Technique: The system uses Minimum-Distortion Embedding (MDE) to project the SOM nodes into a 2D layout. Users drag nodes to desired locations (e.g., placing "wet" nodes on the left and "dry" nodes on the right). The MDE algorithm adjusts the remaining nodes to minimize distortion while respecting these fixed anchors, creating an Adjusted SOM Node Space.

3. Annotate (S2)

Users define regions of interest within the Adjusted SOM Node Space and assign text labels to them, creating an Annotated SOM Node Space. This step is augmented by Large Language Models (LLMs) to assist in sensemaking via two directions:

• Forward Direction (Query-to-Region): Users input natural language queries (e.g., "Show me nodes with above-average precipitation in Southern California"). The LLM parses the query, resolves region names to specific county boundaries, and generates a PostGIS query to filter and highlight the corresponding SOM nodes.
• Backward Direction (Region-to-Summary): Users select a region in the SOM space. The system samples nodes within that region, aggregates per-county statistics, and uses the LLM to generate a ~50-word textual summary describing the precipitation patterns in that region.

4. Analyze (S3)

The annotated space serves as the basis for the three core analysis tasks:

• Exploration: A single model run is visualized as a Kernel Density Estimate (KDE) over the Annotated SOM Node Space. This "fingerprint" highlights hotspots where the model run's realizations are most concentrated.
• Comparison:
  • Side-by-Side: Direct visual comparison of KDE distributions for two runs.
  • Vector Field: For ordered comparisons (e.g., historical vs. future), the system calculates optimal transport between bootstrap samples of two runs to generate a vector field. This visualizes the spatial transition from one distribution to another.
• Clustering:
  • Model Runs: Clustering is performed using Earth Mover's Distance (EMD) to quantify similarity between distributions, followed by UMAP and HDBSCAN.
  • GCMs by Forcing: Clustering is performed on the vector fields representing the transition from historical to future scenarios for different GCMs, identifying models with similar forcing responses.

Key Contributions

1. Novel Workflow: A visual analysis workflow that abstracts spatiotemporal ensemble members as distributions over a steerable 2D space framed by a SOM, moving beyond mean-field aggregation.
2. LLM Integration: The integration of LLMs to facilitate bidirectional interaction (natural language queries to locate regions and automated summarization of regions) within the SOM-defined latent space.
3. ClimateSOM Interface: A complete visualization interface supporting end-to-end exploration, comparison (via side-by-side and vector fields), and clustering of ensemble model runs.

Results and Evaluation

The authors evaluated ClimateSOM using a dataset of LOCA-downscaled CMIP6 precipitation projections for California and the Northwestern United States under three Shared Socioeconomic Pathways (SSPs).

• Case Studies:
  • California: An expert identified that model runs predicting a wetter January were associated with a drier November, a trend not immediately obvious. The system also revealed that the MIROC6 GCM exhibited distinct precipitation patterns compared to other GCMs under the SSP370 scenario.
  • Northwestern US: An expert observed that the transition to a high-emission future (SSP585) resulted in drastically different spatial patterns between April (wetter Northeast) and November (wetter Northwest). The system also helped identify outlier GCMs with anomalous forcing behaviors.
• Expert Review: Six domain experts (including the case study participants) reviewed the system. They reported that the workflow could uncover novel, climatologically significant insights that would otherwise require substantial manual effort (e.g., generating and sifting through hundreds of static plots). Experts particularly valued the distributional view, the vector field comparison, and the ability to track the evolution of clustering results over months.
• LLM Evaluation: A small-scale evaluation with 8 participants tested the LLM's ability to map region names to counties and summarize regions. Participants found the generated lists and summaries generally reasonable (average Likert scores of 2.11 and 2.5, respectively, where lower is better/agreement), though challenges remained with ambiguous region definitions and disconnected county lists.

Significance and Claims

The paper claims that ClimateSOM addresses a gap in existing climate ensemble visual analytics by providing integrated support for exploration, comparison, and clustering. By abstracting model runs into distributions over an interpretable 2D space, the system allows researchers to detect patterns and relationships that are difficult to uncover with traditional statistical summaries or isolated subtasks.

The authors emphasize that the system's ability to combine steerable dimensionality reduction with LLM-assisted sensemaking enables users to quickly identify "hotspots" and meaningful trends. The feedback from domain experts suggests that the system is capable of generating insights of climatological significance, particularly in understanding the variability and evolution of ensemble projections. However, the authors remain modest, acknowledging limitations such as the lack of native statistical significance metrics, visual complexity for new users, and the current abstraction's disregard for temporal ordering.