Robustness Space of AMOC Early Warning Signals: A Comprehensive Sensitivity Analysis of Window Sizes, Proxy Datasets, and Bootstrap Baselines

This dataset provides a systematic parameter-space mapping of Critical Slowing Down (CSD) indicators (rolling variance and lag-1 autocorrelation) for the Atlantic Meridional Overturning Circulation (AMOC).

Scientific Value

Recent predictions of an AMOC collapse rely heavily on Early Warning Signals (EWS). However, these metrics are highly sensitive to subjective choices, such as sliding window widths and baseline selection. This dataset maps the entire EWS significance surface (p-values) across a comprehensive parameter space to eliminate parameter cherry-picking and enable direct robustness testing.

Files Included

• master_dataset.parquet (Single file, Zstd-compressed, ~72,000 rows): Calculated rolling variance, AR1, and empirical p-values mapped across 10 window sizes (120–1200 months), 2 proxy datasets (HadISST and NOAA ERSSTv5), and 2 bootstrap null models (full_series vs baseline_only 1870–1950).
• run_metadata.parquet: Model calibration coefficients and exact temporal baseline ranges.
• kernel_validation_report.parquet: Quantitative benchmark showing mathematical correlation (corr = 1.0, bias = 0.0) against synthetic reference signals.
• pipeline_run.log: Complete execution log of the data generation run.

Methodology

The AMOC proxy is defined as the sea surface temperature (SST) fingerprint of the subpolar gyre (SST of 50°N–65°N, 60°W–10°W minus global average SST; Caesar et al., 2018). Seasonal anomalies are computed relative to the monthly climatology. Significance is evaluated using 1000 AR(1) bootstrap surrogates. The rolling engine is fully vectorized via NumPy memory striding for exact numerical reproducibility.

Usage (Python / Polars)

import polars as pl
df = pl.read_parquet("master_dataset.parquet")
sub = df.filter(
    (pl.col("dataset") == "HadISST") &
    (pl.col("bootstrap_variant") == "baseline_only") &
    (pl.col("window_size_months") == 360)
)