Why an imperfect climate model is more helpful than you think

Our understanding of future climate change depends largely on complex computer simulation models. These models are huge – they are made up of over one million lines of computer code, representing the knowledge of dozens of scientific subfields – without any individual scientist fully understands all the internal work of the model.

More importantly, climate scientists should focus not only on one model, but develop and compare dozens of In various climate model comparison projects, or more such models. Opportunity reproduction, but so is complexity. As a philosopher of science, I am fascinated by the way model comparisons can help scientists draw true insights from this entanglement of complexity and uncertainty.

Inspired by “Storm Outposts” written by climate modeler Ricky Rood, here I reflect on how models designed for one purpose ultimately produce insights in unexpected ways.

Climate scientists often announce statistician George Box that “all models are wrong, but some are useful.” But what exactly does “error” mean? On the one hand, it may be related to the idealization and abstraction contained in all scientific models, including climate models:

• Ideal Related approximate truths – for example, climate models parametric cloud microphysics, meaning they represent larger scale effects of cloud formation and other processes without explicitly expressing these processes.
• abstract Related to missing a process or complete feature – for example, many climate models today omit the representation of brown carbon aerosols (but they do represent other types of aerosols).

Therefore, “all models are wrong” may mean All models include some idealization or abstraction. As the box says, it is still advantageous to adopt such idealized and abstract models. Indeed, a highly idealized model will eliminate a large amount of real-world processes than a model that tries to capture the wind and every drop of rain.

Climate models, on the other hand, may be wrong because they make us inaccurate in the output of past, present or potential future climates. That is, climate models can give us wrong answers.

This is easiest to see for past and present climate situations, and the model output can be compared with observations. These comparisons can be made for a variety of climate variables involving observations such as mean, variability, and temperature, precipitation, pressure, and other climatic phenomena of interest.

A major challenge is that performance analysis of climate models produces a mixed bag: some models perform better relative to certain variables, while others perform well elsewhere. For example, if we rank the best effect of simulating global average surface temperatures in the 2010s and which has best simulated the rate of Arctic sea ice melting in the past 30 years, different models are likely to emerge.

To capture the second idea, “all models are wrong” may mean All models produce at least some inaccurate output. How would such an inaccurate model be useful? Here, the story is not that simple.

One answer is to emphasize that the model is Suitable for purpose. As Rood said, it is necessary to “design, build and evaluate models to solve specific applications.” If your goal is to teach students the basics of Earth’s energy balance, a simple model that you can solve by hand may be the most appropriate. If your goal is to project the global average climate at the end of the century with reduced emission reduction, a state-of-the-art general cycle model would be more appropriate.

One challenge in making the model “fits purpose” is that while certain priorities are obvious, many more finer decisions may not be clear. For example, it seems reasonable to update the model's cloud representation to better capture cloud feedback, thereby improving the model's ability to simulate temperature changes Fewer Accurately simulate such temperature changes.

They were able to diagnose why the model became inaccurate, but only after facts and iterative testing and analysis performed within a few months. More realistic models are often more complex, and more complex models involve more interactions, feedback, and emerging behaviors than scientists track in their minds. Therefore, it is quite a difficult task to determine which purpose the model will ultimately be suitable for before developing and testing it. (Not to mention that modelers and model users are often different groups of people, and model builders cannot predict how others will use their own models on the road.)

The second challenge is that models deemed unsuitable for certain purposes may be for another (sometimes related) purpose. In this case, simply judging that the model is “not suitable for purpose” would be too fast. Instead, it is often possible to reuse incorrect, inappropriate or “wrong” models and ultimately play a crucial role in generating new knowledge. Let's see how this works.

In a 2020 study, Katarzyna Tokarska and colleagues generated a model-based estimate called transient climate response. Transient climate response, or TCR, is a measure of climate sensitivity, which describes the global temperature increase when carbon dioxide concentrations gradually increase over a period of time. The team’s analysis involves looking at recent model simulations of recent warming and model simulations for potential future warming, developed using the latest comprehensive model suite, called CMIP6, to support the development of the Intergovernmental Group on Climate Change.

Some CMIP6 models have done relatively poorly in simulating recent warming. Let's call it a model that's too hot. The model that is too hot obviously encountered important mistakes and was rightly criticized on these reasons. However, scientists don’t want to rely solely on a few models in case those models are biased or not represented in ways that are not yet understood. It can be said that quantity has strength.

So instead of throwing out the model that was too hot, the team repurposed them to play a key role. These models are drawn with other models and a clear pattern emerges: the warmer the model has shown in recent decades, the higher its estimates of TCR.

This pattern tells scientists two things. First, the ensemble of the entire model does a great job of capturing the critical relationship between past and potential future warming. Second, scientists can focus on models that match the observed historical warming and follow patterns to update their estimates of TCR. As a result, they generated a narrower, more robust estimate rather than a wide TCR range-shaved at the top of the model-based range near 1°C. Models that are too hot are not used directly in the final estimate, but they play a crucial role: they help establish the pattern and show that the climate model performs well. (See Deeper Diving here).

This situation reflects general patterns in climate modeling. In 2005, Ben Santer and colleagues published an analysis in which climate models do not do well alone in simulating two variables (surface temperature variability and lower troposphere temperature variability). relation Between these two variables. These “error” models provide enough evidence to ensure that some controversial satellite data is discarded. Recently, models that may perform poorly for certain purposes of interest can be used to provide test bed data for “better” models in the model weighting framework.

In this way, climate modeling studies exemplify what former options trader Nassim Taleb called “anti-difference”. The idea of ​​Taleb is that setting things up (whether it's your career, investment, home, research project or anything else) is a good thing so that you can benefit riots out of error, failure or unforeseen. Due to some distraction, fragile things will break. The above riots will prosper.

“You want to be a fire, hope the wind,” Talib said.

A gust of wind can put out the candle, but the air (containing oxygen) increases the intensity of the wildfire. It's best to be a wildfire.

Therefore, in a significant sense, climate modeling may be like wildfires. Discovering that your model performs poorly is a failure, but the modeling community is structured as Learn from learning. These failures – whether simulated Arctic sea ice trends or transient climate responses – do not mark the end of usefulness.

Instead, as we see, despite the imperfect model, scientists can repurpose their models to generate new knowledge. We may be tempted to consider this type of activity as a pragmatic solution. However, I think the added values ​​of flawed models reveal more in-depth information about the growth of knowledge in climate modeling. Even when in error, the model itself captures many robust relationships. Through comparison, iterative and critical explanation, scientists can make progress becausenot just althoughimperfect.

Ryan O'Loughlin is an assistant professor who studies and teaches philosophy and climate change at Queens College and CUNY Graduate Center.