Forecasting the extremes with Climate and Weather models

Climate scientists have developed around 120 models to study natural processes, interactions, and feedbacks within the Earth's system, such as melting glaciers, rising sea levels, and how atmospheric circulation in one hemisphere affects extreme floods and droughts in another.

Both climate and weather models are national projects supported by billions of dollars in investment. Today, there are approximately 120 institutional global climate models, each involving several thousand climate experts and modelers.

Both weather models and climate models act as digital twins of the Earth. Weather models produce short-term predictions, while climate models focus on natural variability and long-term trends. Weather models run for limited areas (just one region) and for short periods (only a few days), with very high spatial resolution of 1 km. In contrast, climate models operate at a resolution of 50-100 km but cover the entire planet, simulating weather anomalies and extremes over weeks, months, years, decades, and centuries.

Unlike weather models, climate models do not assimilate real-time observations. This is simply because there were no satellite weather data 300 years ago, and no observations are available for 300 years into the future!

From time to time, climate models from different research centers participate in a global “Olympic Games” of sorts, which spans 2-3 years. The most recent took place from 2018-2021: the Coupled Model Intercomparison Project - Phase 6, also known as CMIP6.

To ensure that the final results are comparable, all models follow the same rules: same land use, same population, same carbon emissions. Climate change scenarios for all models are defined by the IPCC group of climate experts. All models run under three main scenarios: RCP2.6, RCP4.5, and RCP8.5. Certain models also run with RCP1.9, RCP6.0, and RCP7.0 scenarios. In the finance world, when using climate data for stress testing, human-readable names were given to these technical scenarios: ‘orderly’ for RCP2.6, ‘disorderly’ for RCP4.5, and ‘hot house world’ for RCP8.5. In the previous climate model run, back in 2005-2007, there was a scenario called RCP8.5, which corresponds to RCP4.5 in the modern scenario definitions. Back then, this “worst case scenario” was referred to as ‘business as usual.’ Today, RCP4.5 is no longer considered the worst-case scenario.

Combining different models under various climate scenarios helps us better quantify uncertainty and to map and capture extreme events that occur once every 50, 100, or 200 years - those not yet observed but likely to happen.

And this is where the concept of the direct validation should be reconsidered. If the idea of flood modeling is to map floods in those areas where floods were not observed, yet likely to happen once every 50, or 100 years, then there is no way you can validate these “at risk” maps, because there are no observations of floods.

By studying past and future climate conditions over long time periods, scientists can better distinguish between natural climate variability and human influences on climate, such as deforestation, land use, and pollution of soil, water, and air.

Validation

Prior to the physical climate risk assessment, the climate models should be validated. Similar to the operational weather forecasts, the success score ("hit rate") is a measurable quantity.

The validation process can be resumed as followed:
How many floods, droughts and wildfires were successfully predicted by each individual model, and by the ensembles of climate models ("majority vote")?

Successful alert: the extreme event happened and was predicted by the model.
False alert: the extreme event was predicted, while nothing happened.
Missed alert: the extreme event happened but was not predicted.
Wrong alert: the extreme event happened but was predicted with the wrong sign.

One example of a "wrong alert": extreme cold winter predicted instead of extreme warm winter.

Thanks to the IPCC community effort and global scientific advances, today there is a good understanding of how the ensembles of global climate models performed over the historical period 1950-2022.

NOTES:

• Ensemble modeling of extreme events with a "60% majority vote" performs better than each individual model separately;
• Model performance depends on the geography: coastal areas vs. continental areas / tropical vs. temperate climate;
• Model performance is much better for temperature extremes rather than precipitation and wind extremes;
• Model performance is higher for CMIP6 High Resolution models, compared to a 15-year-old CMIP5 data in 200-300km resolution.

CMIP6 stands for the Coupled Model Intercomparison Project Phase 6.