Strategies to Accelerate US Coal Power Phaseout Using Contextual Retirement Vulnerabilities

Paper — Quick Context

• Published in Nature Energy: a data‑driven framework for accelerating U.S. coal retirements.
• Classifies every coal plant into 8 retirement archetypes using 68 technical, economic, environmental, and political variables.
• Reveals why plants persist and where interventions are most effective (financial, regulatory, public health, grid context).
• Introduces the Thema algorithm: a robust unsupervised method for learning optimal graph representations from high dimensional datasets.

Our Graph Models

• The graph shows all U.S. coal plants; connections indicate similar features across environmental, sociopolitical, and financial factors.
• This is a Mapper graph (Singh et al.), a tool from topological data analysis. Nodes represent clusters of coal plants, and edges connect clusters that share membership (i.e., overlap in plants).

Retirement Groups — Quick Overview

• Fuel Blend Plants: Regulatory non-compliance and clean energy targets.
• Retrofitted but Vulnerable Plants: Only partial retirements planned; economic unviability and renewable competition.
• Democratic Majority Plants: Political and regulatory pressure for clean energy transitions.
• High Health Impact Plants: Air-quality related public health concerns and environmental regulations.
• Expensive Plants: High operating costs and environmental retrofit requirements.
• Young Plants: No planned retirements.
• Plants in Anti-Coal Regions: Political opposition and economic difficulties.
• Air Quality Offenders: No planned retirements.

Thema — Learning Relevant Representations

When analyzing the US coal fleet, we faced a common challenge: deriving robust, actionable insights from a complex, high-dimensional dataset. We were overwhelmed by the noise and variability inherent in simply extracting a usable representation of the data. Traditional dimensionality reduction techniques often create unstable "insights"—a small parameter change can dissolve the entire embedding. Thema manages this sensitivity as a feature, not a bug. Our design embraces the variation from different modeling choices, allowing us to produce and reason about a distribution of data structures (graphs). This approach ensures we capture the consistently real facets of the underlying data, which we then condense and optimize for downstream tasks, like accelerating coal retirement.

How it behaves

• Build graphs that approximate a manifold structure over your data.
• Simplify the "multiverse" of parameter settings, optimizing model selection for the downstream task of your choice.
• Leverage graph algorithms to scalably extract insights from your data.

Why it helps

• No more “pick the pretty embedding.” Confidently select a usable representation for your analysis.
• Reproducible by design; the result isn’t one lucky hyperparameter.
• Plays nicely with clustering, visualization, and decision‑making.