Fusion of Probabilistic Projections of Sea-Level Rise

Benjamin S. Grandey, Justin Dauwels, Zhi Yang Koh, Benjamin P. Horton, Lock Yue ChewResearch Square 2024 · DOI

TL;DR

Researchers have developed a "fusion" approach to sea-level rise projections that combines the reliability of standard climate models with the cautionary insights of expert elicitations. By blending these different "workflows," the paper provides a single, continuous probabilistic distribution that offers a "very likely" range for sea-level rise—something previously missing from major international assessments.

What Problem Does This Paper Address?

Predicting how much the oceans will rise by 2100 is notoriously difficult, largely due to "method ambiguity." Scientists use different "workflows" to create projections: some rely on well-understood physical models (medium confidence), while others incorporate poorly understood but potentially catastrophic processes like marine ice cliff instability (low confidence).

The Intergovernmental Panel on Climate Change (IPCC) Sixth Assessment Report (AR6) presented these as separate options. This leaves decision-makers—like city planners and coastal engineers—in a bind. If the models don't agree on the "tails" (the extreme high-end risks), how do you design a sea wall? Because of this ambiguity, the IPCC was unable to define a "very likely" (5th–95th percentile) range for sea-level rise, which is the gold standard for risk management.

Methods

The authors propose a fusion approach that applies an additional layer of expert judgment to combine alternative projections into one "best estimate" distribution.

  1. Quantile Functions: Instead of looking at simple averages, they treat each projection as a "quantile function" (mapping probability to sea-level rise values).
  2. The P-Box: They use a "probability box" to envelope all four major IPCC AR6 workflows, identifying the absolute upper and lower bounds of current scientific thought.
  3. Trapezoidal Weighting: This is the core innovation. The authors use a quantile-dependent weighting function. They give 100% weight to the "medium confidence" models for the central part of the distribution (the 17th to 83rd percentiles). However, as they move toward the "tails" (the extremes), they smoothly transition the weight toward the "low confidence" outer bounds.
  4. Mathematical Fusion: This creates a single, continuous, and monotonic probability distribution that captures both the most likely outcomes and the high-risk "storylines."

Datasets

The study utilizes the IPCC AR6 sea-level projections data. Specifically, it focuses on four workflows:

  • Workflow 1e & 2e (Medium Confidence): Based on multi-model ensembles (like ISMIP6 and GlacierMIP) and statistical emulators.
  • Workflow 3e & 4 (Low Confidence): Incorporates high-impact processes like Marine Ice Cliff Instability (MICI) and Structured Expert Judgment (SEJ).
  • Scenarios: Projections are evaluated under SSP1-2.6 (low-emissions) and SSP5-8.5 (high-emissions) scenarios for the year 2100.

Key Insights and Ideas

  • Reliability vs. Possibility: The "aha moment" here is the realization that different models are "best" at different things. Physical models are great at predicting the most likely central outcomes, but expert judgment is better at capturing the "known unknowns" in the extreme tails.
  • Continuous over Discontinuous: Previous attempts to create "outer bounds" resulted in mathematical discontinuities at the median. The fusion approach fixes this, creating a "well-behaved" distribution that is easier for statistical downscaling and economic modeling.
  • Catering to Diverse Users: A "one-size-fits-all" number doesn't work for everyone. A farmer might care about the median rise, while a nuclear power plant operator needs to know the 95th or 99th percentile. This fusion provides a single framework that serves both.

Key Results

The fusion provides concrete probabilistic answers for Global Mean Sea-Level (GMSL) rise by 2100:

  • Very Likely Range (5th–95th percentile):
    • Low-emissions (SSP1-2.6): 0.3–1.0 meters.
  • High-emissions (SSP5-8.5): 0.5–1.9 meters.
  • High-End Risk: Under high emissions, the 95th percentile projection is 1.9 meters. This is significantly higher than what standard physical models suggest (1.1m) but more grounded than the most extreme "low confidence" upper bounds (2.3m).
  • Probability of Exceeding 1.0m: The fusion estimates a 16% chance of exceeding 1 meter under high emissions, compared to just 4% under low emissions.

Limitations and Future Work

  • Subjectivity: The choice of the trapezoidal weighting function and the specific percentiles (17th/83rd) is an act of expert judgment. While transparent, it is still subjective.
  • Consensus Challenges: Implementing this at an institutional level (like the IPCC) would require a panel of experts to agree on these weights, which could be politically and scientifically difficult.
  • Scope: The current paper focuses on global mean sea level. The authors suggest that future research should apply this fusion method to local sea-level rise, vertical land movement, and other climate variables like surface temperature.
  • Static Nature: As ice sheet science evolves, the "strengths" of different workflows will change, meaning the fusion will need periodic updates.

Summary updated

Abstract

<title>Abstract</title> Alternative projections of sea-level rise from ice sheet mass loss differ markedly. The differences drive ambiguity in probabilistic projections of sea level. Can the alternative projections be combined to produce a single probabilistic projection that quantifies a best estimate of the scientific uncertainty? To produce this best estimate, we propose a <italic>fusion</italic> of alternative projections. The fusion combines the complementary strengths of the ‘medium confidence’ and the ‘low confidence’ projections. Two assumptions undergird the fusion. First, we assume that the ‘medium confidence’ projections reliably sample the most-likely central possibilities. Second, we assume that the p-box outer bound – which includes the ‘low confidence’ projections – provides a more reliable estimate of unlikely tail possibilities driven by poorly-understood processes. Accordingly, the fusion is weighted towards the ‘medium confidence’ projections near the median and towards the p-box outer bound near the tails. We demonstrate our proposed fusion by combining alternative projections of sea-level change at the end of the 21 <sup>st</sup> century. The fusion provides probabilistic answers: according to the fusion, the probability of global mean sea-level change exceeding 1.0 m by 2100 is 6% under a low-emissions scenario and 31% under a high-emissions scenario. The fusion also informs summary statements based on credible intervals: under a high-emissions scenario, for example, we are 90% certain that global mean sea level will rise by 0.5–2.2 m during the 21 <sup>st</sup> century. By quantifying a best estimate of scientific uncertainty, the fusion caters to diverse users.
BibTeX
@inproceedings{grandey2024fusion,
  author = {Benjamin S. Grandey and Justin Dauwels and Zhi Yang Koh and Benjamin P. Horton and Lock Yue Chew},
  title = {{Fusion of Probabilistic Projections of Sea-Level Rise}},
  booktitle = {Research Square},
  year = {2024},
  doi = {10.21203/rs.3.rs-2922142/v2},
}
Figure