To adapt to climate change, we need better models » Yale Climate Connections

For decades, scientists have simulated Earth’s climate, and they have done a good job representing the rise in the planet’s average temperature. Now we must rise to the need to adapt to a warming climate, and much work remains to make modern-day simulations usable to the people doing the on-the-ground work of updating municipal infrastructure, designing coastal defenses, and more. 

Early models set the stage for advanced simulations 

Researchers first began using computers to simulate Earth’s atmosphere in the 1950s. The global climate models developed in the 1970s, though not developed specifically for predictions, showed remarkable accuracy in their projections of global warming. 

“A lot of those early models ended up proving quite prescient in terms of predicting what would actually happen in the real world in the years after they were published,” Zeke Hausfather, a research scientist at Berkeley Earth, said in an interview with Yale Climate Connections.

Hausfather led a 2020 study that evaluated those early models. “Of the 17 we looked at, 14 of them were pretty much spot-on,” he said. 

Over time, global climate modeling capacity evolved to focus largely on the year 2100 and the investigation of policies to reduce the emissions of heat-trapping greenhouse gases. Most of these simulations are part of the highly successful Coupled Model Intercomparison Project, or CMIP, the primary database of climate simulations. Since the mid-1990s, CMIP has coordinated a series of model-based experiments that are used in the assessments of the Intergovernmental Panel on Climate Change.

When addressing a real-world problem, scientists, like me, often start with the CMIP models and their archived simulations. The CMIP database is a rich source to interrogate, and with sufficient expertise and knowledge of the models, usable knowledge for many applications can be extracted. Such interrogation also reveals the limitations of the design of the CMIP models and simulations to address problems such as adaptation, geoengineering, and the details of climate tipping points.

This series of articles focuses on framing new strategies to provide models and simulations for these critical applications. This requires us to consider not only the traditional approaches to climate modeling, but to recognize that we “feel” climate through weather and, especially, changing weather extremes. Therefore, strategies that better integrate weather and climate modeling approaches are warranted.

The first article focuses on adapting to a more variable, warming climate. Adaptation provides a use case to expose limitations of current modeling capabilities and motivate the need for new approaches. Future articles highlight applications other than adaptation, describe barriers to the use of model-based knowledge, and describe changes to our approaches to climate modeling that are needed to address the increasingly obvious threats that come from a warming climate.

What models can and can’t do 

A foundational principle of model-based science is that models must be fit for purpose – in other words, designed, built, and evaluated to address a specific application. In many instances, we have lost touch with this principle.

Models are ingrained in the practice of science. Scientists become used to them as essential tools, and we lose our grounding that models are part of a process. It is easy to ascribe more reality to their numbers and patterns than is justified. In public communications, models are reported, frequently, with the impression that they are factual expectations.

The primary roles of models are to facilitate investigation, describe uncertainty, and contribute to reasoning. They organize observations, help isolate and describe processes, and provide guidance for the future.

Why we need good models for adaptation work

When it comes to adaptation, we are already making decisions about building and rebuilding, agriculture, and ecosystems. The decisions we make today will be implemented and expected to function in the climate 30 to 50 years from now. That future climate will be warmer and, short of emission-eliminating interventions to reduce carbon dioxide and methane emissions, still warming.

Since the 1970s, climate models have evolved to contain enormous complexity and require the use of the largest computers. The models were used, first, to probe processes important to the essential physics of Earth’s climate, and a primary goal was understanding how these systems work. As model complexity has increased, development decisions are often guided by the design of CMIP to support assessments of our past and current climate and what to anticipate in the next 100 years. The 100-year outlooks explore scenarios of different policy interventions and economic development.

The use of CMIP simulations in adaptation planning is a repurposing of the original goals. Therefore, an evaluation of fit for purpose is the correct first step in adaptation applications.

By focusing on climate adaptation in the North American Great Lakes region as a use case, we reveal three classes of shortcomings in the available simulations. The first is logistical. Though widely and freely available, the sheer complexity of dealing with the archive can be overwhelming.

The second comes from an evaluation of the simulations by comparison with observations. Substantial errors are present, often represented by biases. Though bias adjustments to fit the simulations to observations are possible, such adjustments are likely to contribute to overstatements about how certain we are about particular outcomes.

From a feature-based perspective, model evaluation shows that weather phenomena are represented with widely varying fidelity. For example, winter cyclones are better represented than summer convective systems. Since weather is one of the ways we feel the climate, representation of weather features is important to adaptation. Local officials are often motivated to think about adaptation due to weather-related catastrophes and vulnerabilities.

The third class is structural. This is easy to see if you zoom in on the Great Lakes region. Do the models represent the lakes and if they do, how do they represent the lake’s behaviors? For example, does the model include the ability to form and melt lake ice?

In the paper, “Large lakes in climate models: A Great Lakes case study on the usability of CMIP5,” a multiyear process to determine the presence or absence of the Great Lakes is detailed. This paper found that most of the models do not include lakes in a way that captures their impact on regional climate. At a time of rapid warming and urgency for adaptation decisions, a multiyear exercise that reveals such fundamental gaps compels us to search for new strategies to provide model guidance for adaptation.

I bring a couple of these issues together. The representation of lakes in current climate models might be adequate to describe their small effects on the world’s climate system. However, if lake effect precipitation and how it changes is not represented, it does little good for a professional working on an adaptation problem in Buffalo, New York. Models that can’t account for lake effect snow won’t win any credibility contests with the city’s mayor.

Similarly, we have a good sense of where temperature is headed in the Great Lakes region. We have defensible guidance on increasing variability of precipitation. The simulations tell us that storm sewer pipes will need to be to accommodate increasingly intense downpours. But the size of those pipes and when they are needed remains highly uncertain.

For problems such as the rise and fall of lake levels, which depend on the accumulated effects of precipitation, evaporation, and runoff. When precipitation, evaporation, and runoff are combined over the seasonal and annual time spans necessary to calculate lake water budgets, the bias errors also accumulate and exceed 100% in each part of the budget calculation. Such biases are indefensible for decision-making and require expert guidance to interpret and describe the uncertainties.

There’s a similar story playing out in the Southeast, where coastal cities are working to adapt to rising seas and more severe hurricanes. Detailed representation of our coastline and coastal processes is not essential in a global climate model. However, they are at the core of making decisions about coastal defenses, and how long the coasts can be defended. 

High-resolution regional models offer a tool to address some of the issues of locality that must be addressed. However, regional models rely on global models to be of high quality to define the boundaries of the region to be simulated. Regional models inherit not only the global model’s description of climate parameters but also the global model’s systematic errors.

We can identify many cases of flood, drought, fire, and water resources for which we need to be making decisions to improve our environmental security. They require knowledge about both the instantaneous consequences of severe weather and the accumulated effects of heat, water scarcity, and water abundance. Given how fast the climate is warming and the severity of problems we face, taking years to establish that models are or are not fit for purpose is unsatisfying and unacceptable. 

The need for a better way

We can, with the present-day knowledge and resources, develop answers that describe the current observed warming and that address adaptation concerns. However, the barriers that we need to cross in answering these problems are large. Addressing the logistical barriers is conceptually straightforward. However, the structural barriers and the results of evaluation for fit for purpose are likely to leave us with results that are difficult to reconcile.

This is especially true given the real-world consequences of our decisions. We need to assure our audiences that the scientific evidence is well grounded in evaluation and uncertainty descriptions. We have the difficult ethical requirement to not overstate our presumed knowledge about the future climate, while at the same time not understate the vulnerabilities and risks that we face.

As a community, we know the current situation is inadequate. We know we need better ways to use models, as well as fundamental improvements to the models.

A natural response to this need is to continue to do what we have done but to do a better job to accommodate applications in the CMIP protocols. An intuitive approach is the development of more complex, more comprehensive, more “Earthlike” models. Often our modeling strategies are anchored around the acquisition and use of the biggest computers.

Though such proposals might be elements of a more robust modeling strategy, they would be only part of the strategy and, perhaps, not the most urgent and important parts. The expense and complexity alone are so daunting as to challenge the credibility of such approaches. Two decades of calling for immense centers organized around the largest, most comprehensive models have proved not viable organizationally and politically.

More important, however, are these questions: Would such an approach, in fact, address our applications more effectively and convincingly? Would the resulting models be fit for purpose?

The applications essential for environmental security need to be at the front of a large portion of our model development paths. These applications include determining how fast (and why) the climate is changing, adaptation, geoengineering, and the risks and consequences of tipping points. Without purpose, we do not have robust criteria for evaluation. Such criteria are essential to constrain the seemingly infinite choices we could make of what to put into models, how to put it in, and then how to use and interpret them. 

Though adaptation was used in this article as a use case, the barriers described above are universal. Some have been known for years, and there have been substantive efforts to address them. However, we see those efforts countered, if not overwhelmed, by the growth, complexity, and cost of computational environments and data systems, simulation design, model comprehensiveness, and the structural requirements for the uncountable number of applications that a warming climate brings us. 

Many attributes of a more strategic approach to climate modeling are easily listed. We need to reduce the time it takes to determine the usability and to be confident of the relevance of simulations to our applications. We require the flexibility to address a wide range of problems with models and simulations tailored to purpose and to describe uncertainty. We need to consider the roles of potentially revolutionary technologies such as machine learning. The list needs to be developed fully. 

Rather than focusing on computers and scaling models with additional complexity over silicon, we need to accelerate scaling over the innovating power of human intelligence. Standardized, evaluated models and simulations need to be available to a wide range of application specialists to engage the expertise, creativity, and the capacity of the extended community.

Richard B. (Ricky) Rood is a professor emeritus at the University of Michigan, Ann Arbor. The author thanks Dr. Monica Morrison of the National Center for Atmospheric Research and Dr. David Stainforth of the London School of Economics for their comments on an earlier version of the article.