← All work

Climate Warming Pattern Recognition from CMIP6 Simulations

A scalable analysis of CMIP6 climate simulations that measures how the annual distribution of temperatures is changing across the globe, identifies regions with similar warming patterns, and examines how these patterns shift over the long term.

Climate data · Clustering · Scalable data processing
World map colored by six clusters of warming patterns identified from historical climate simulations for 1948–2014.
Six warming patterns identified by K-means clustering for 1948–2014. Grid cells of the same color share a similar warming profile across their annual temperature distribution.

The problem

Climate change is often summarized by a single number, such as the rise in global mean temperature. However, warming varies from region to region, and it also varies within each location’s temperature distribution: the coldest periods of the year may warm at a different rate than the warmest ones. These differences matter because many impacts, such as heat stress, frost, and ice loss, depend on the extremes rather than on the average.

The goal: measure how the annual distribution of temperatures is changing at every location on Earth, and identify regions that share similar warming patterns.

Data

CMIP6 · NOAA-GFDL GFDL-ESM4 climate model

51,840grid cells covering the globe (1° × 1.25°)
3-hourlynear-surface air temperature
153years of historical and projected temperatures
~93 GBof temperature data analyzed

The data come from simulations of the GFDL-ESM4 model in the Coupled Model Intercomparison Project Phase 6 (CMIP6). The analysis uses the historical simulation for 1948–2014 and the future projection for 2015–2100 under SSP1-2.6, a low-emissions scenario. Both are read directly from the Pangeo cloud catalog of CMIP6 data.

Approach

A three-step pipeline was built to turn high-frequency temperature data into a warming profile for each location and then group locations with similar profiles. The same pipeline is applied separately to the historical and future periods to identify long-term climate shifts across the globe.

Schematic of the analysis pipeline. Temperature T (year × time × lat × lon) is reduced to annual quantiles Q (quantile × year × lat × lon) in step 1; a least-squares trend over years gives warming rates β (quantile × lat × lon) in step 2; these are reshaped into warming profiles X (51,840 grid cells × 101 quantiles), which K-means groups into 6 clusters in step 3.
Analysis pipeline, applied separately to each period. Boxes show the arrays and their dimensions; numbered steps correspond to the sections below. Y is the number of years (67 for 1948–2014, 86 for 2015–2100).

1. Annual temperature distributions

For each grid cell and year, the empirical distribution of 3-hourly temperatures is characterized by 101 quantiles, spanning the coldest to the warmest conditions of the year.

2. Warming profiles

The warming rate of each quantile is estimated as the slope of a least-squares linear trend across years, in kelvin per year. The resulting 101 rates constitute the warming profile of each grid cell.

Two line charts of warming rate against quantile of annual temperature for 1948–2014. Near the North Pole, rates are highest for the lower and middle quantiles and fall close to zero for the warmest. In the central United States, rates rise steadily toward the warmest quantiles.
Warming profiles at two locations, 1948–2014. Near the North Pole, the colder and middle parts of the distribution warmed fastest, while the warmest periods changed little. In the central United States, the warmest periods warmed fastest.

Mapping the warming rate of a single quantile shows the spatial pattern of change in one part of the distribution; the warming profile integrates all 101 quantiles at each location.

World map of the warming rate of annual median air temperature from 1948 to 2014, with the strongest warming over the Arctic Ocean and along parts of the Antarctic coast.
Warming rate of the annual median temperature (quantile 0.5), 1948–2014, in kelvin per year.

3. Warming pattern recognition

Grid cells are partitioned into k clusters based on their warming profiles, independent of geographic location, to identify regions that share similar warming patterns. The value of k is determined using the elbow method.

Scalable processing

Given the volume of the data, the pipeline relies on xarray and Dask for lazy, chunked computation, reading data directly from cloud storage without loading the full dataset into memory.

Results

  • Fastest historical warming near the poles. From 1948 to 2014, the median temperature warmed fastest over the Arctic Ocean and along parts of the Antarctic coast.
  • Two dominant patterns. In the historical period, two patterns cover about 83% of the Earth’s surface. The other four are found mostly at high latitudes.
  • Rapid warming of cold extremes. In the smallest historical pattern, covering about 1% of the surface east of Greenland and along parts of the Antarctic coast, the coldest periods of the year warmed several times faster than the warmest.
  • Slower projected warming, fastest in the Arctic. Projected median warming for 2015–2100 is slower than in the historical period across most of the globe and fastest over the Arctic Ocean north of Alaska and eastern Siberia.
  • Regional instead of polar patterns. The ring-like polar patterns of the historical period break up into regional ones. One pattern covers about half of the Earth’s surface, and areas with little or no warming, south of Greenland and off West Antarctica, form their own pattern.

The top image shows the historical patterns; the map below shows the projected patterns for 2015–2100.

World map colored by six clusters of warming patterns identified from projected climate simulations for 2015–2100 under SSP1-2.6.
Six warming patterns for 2015–2100. Clusters are identified separately for each period.

Methods & tools

Climate data analysis · statistical trend analysis · pattern recognition · geospatial analysis · large-scale data processing · Python · xarray · Dask · scikit-learn