Biases in Initialized Climate Models

Biases in Initialized Climate Models

ISCCP Mean Annual Frequency of Cloud Occurrence with Location of Cross Section

Another way to examine climate model bias is to run the model as a forecast model. This has been done by initializing the climate model, not from observations, but rather from re-analysis products. After running the model for several days, researchers can examine how biases develop as the simulation drifts away from the observed climatology toward the model's steady-state climate.

In these simulations, researchers were looking at the bias in tropospheric moisture and temperature over a transect from San Francisco out to the Equatorial Pacific. Along this transect, the cloud regime goes from low marine stratus near the California Coast to deep convection near the Equator. Thus, this transect provides a way to examine biases in a variety of cloud processes.

Note: ISCCP = International Satellite Cloud Climatology Project

Forecast Error of Temperature and Specific Humidity Showing Rapid Drift Toward Model Climatology

The results are startling. These plots are vertical cross sections of model bias relative to observed climatology for 1 day (left), 5 days (center), and the long-term climate model mean (right). The top row of plots shows temperature, while the bottom row shows moisture.

What you see, is that bias starts to build immediately and within 5 days the forecast bias in temperature looks very similar to the long-term climate bias. Similarly, the dry bias seen in the climatology near the coast in the lower troposphere shows up in the forecast after 5 days.

This experiment provides researchers with a very powerful tool for research. Why?

It takes a lot of computer time to run climate models for 10, 20, or 100s of years to look at biases. If the same bias shows up in a 5-day initialized forecast, it enables us to very quickly analyze biases due to physical parameterizations. We no longer have to run the model for decades or centuries to look at some of the biases related to different physical parameterizations. Rather, we track down the source of the bias using a series of 5-day forecasts!

This experiment also very clearly shows how weather forecast models will drift if they aren't re-initialized frequently to keep pulling them back to observations. Model simulations drift very quickly into their own biased state. Weather forecasters know this from looking at longer runs of weather forecast models.