SciencePediaWhy is the future climate so uncertain when it is governed by the deterministic laws of physics? This question lies at the heart of modern climate science. Far from being a sign of scientific failure, uncertainty is an intrinsic feature of complex systems, and understanding its structure is crucial for making reliable projections and sound decisions. This article addresses the challenge of moving beyond a vague acknowledgment of uncertainty to a rigorous, structured understanding. In the following chapters, we will first deconstruct the core principles of climate uncertainty, exploring its origins in chaotic dynamics, model imperfections, and human choices. Subsequently, we will see how this quantified uncertainty becomes an indispensable tool, enabling more robust and resilient strategies in fields ranging from ecology and public health to global economic policy.
If the laws of physics that govern our climate are deterministic, written in the precise language of mathematics, why is the future so uncertain? One might imagine a climate model as a giant, intricate clockwork, set in motion from a known starting point. If we know the rules and the initial state, shouldn't the future unfold with perfect predictability? The fascinating answer is no, and understanding why is our first step into the beautiful and complex world of climate uncertainty.
The climate system, from the swirling of the winds to the slow churning of the oceans, is a textbook example of a chaotic system. This doesn't mean it's random in the sense of a coin toss; it means the system exhibits a profound sensitive dependence on initial conditions. This is the famous "butterfly effect": the notion that the tiny flutter of a butterfly's wings in Brazil could, through a long chain of cause and effect, eventually set off a tornado in Texas. A minuscule, immeasurable difference in the starting state of the system can lead to wildly different outcomes a short time later.
This chaotic nature draws a fundamental line between weather and climate. Weather is the specific state of the atmosphere at a given time and place—the "mood" of the system. Because of chaos, our ability to predict the weather is fundamentally limited. Small errors in our measurement of the current state of the atmosphere grow exponentially, making specific forecasts beyond a couple of weeks essentially impossible.
Climate, on the other hand, is the statistics of weather over long periods—the "personality" of the system. While we can't predict the exact mood on a given day far in the future, we can describe the personality with great confidence. More formally, we can think of "climate" as a stable statistical distribution, a concept known in physics as an invariant measure. The weather at any moment is a single draw from this distribution. The natural wandering of the system within this distribution, driven by its own internal chaotic dynamics, is what we call internal variability. This is the climate system’s own unforced dance—phenomena like El Niño are a perfect example. Climate change, in this powerful framework, is not just a change in the average temperature; it is a fundamental shift in the entire statistical distribution itself, caused by an external push.
How do we separate the chaotic noise of weather from the underlying signal of climate? The key is time averaging. If we take daily temperature data, which fluctuates wildly, and average it over a month, or a year, or a decade, the variance of that average gets smaller and smaller. The fast, chaotic wiggles tend to cancel each other out, allowing the slower, more stable climate signal to emerge. This is like listening to a symphony: in any single millisecond of sound (the weather), you might hear a chaotic clash of instruments, but over the course of minutes (the climate), a coherent and beautiful structure is revealed.
The inherent chaos of internal variability is just one piece of the puzzle. To navigate the landscape of uncertainty, it's helpful to have a map. Scientists divide uncertainty into two broad families, a distinction that clarifies not just what we don't know, but how we don't know it.
Aleatoric Uncertainty: This is the inherent, irreducible randomness in a system. Think of it as rolling a fair die. You can know everything there is to know about the die and the laws of physics, but you can never predict the outcome of a single roll. It's an uncertainty about the world itself. Internal variability is a prime example. We can't eliminate it, but we can describe its statistical properties. Some modern climate models even include stochastic parameterizations, which intentionally inject randomness to represent the unpredictable nature of small-scale processes like individual clouds, making the model's overall variability more realistic.
Epistemic Uncertainty: This is uncertainty due to a lack of knowledge. Think of it as being handed a die and not knowing if it's fair or loaded. This is an uncertainty in our minds, about our model of the world. Crucially, this type of uncertainty is, in principle, reducible. By gathering more data—by studying the die, rolling it many times—we can reduce our ignorance and learn more about its properties.
Most of the uncertainty in climate projections, beyond internal variability, is epistemic. We can organize this "knowledge-based" uncertainty into a few key categories, providing a powerful framework for analyzing climate projections.
Our climate models are monumental achievements of science, but they are imperfect approximations of reality. This imperfection is a major source of epistemic uncertainty. We can break it down further:
Parameter Uncertainty: Models contain numerous parameters—coefficients and numbers that represent physical processes. For example, a coefficient might determine how quickly water droplets in a cloud merge to form rain. We don't know the exact, correct value for many of these parameters, only a plausible range. Exploring this uncertainty by running the same model with many different plausible parameter values is the goal of a Perturbed Parameter Ensemble (PPE).
Structural Uncertainty: This is a deeper form of uncertainty. It's not just about the values of the knobs on our model; it's about the fundamental wiring and design of the model itself. Different scientific teams may have different theories, and thus different equations, for representing complex processes like cloud formation or ocean turbulence. A Multi-Model Ensemble (MME), like the famous Coupled Model Intercomparison Project (CMIP), brings together models from centers around the world to explore this structural uncertainty.
The distinction is critical. Tweaking a parameter in a model is like changing the seasoning in a recipe. Choosing a different set of equations entirely is like choosing a different recipe altogether. A stark example comes from ecology: a model built to predict a plant's habitat based on current climate data might work perfectly for today's world (interpolation). But when we use it to predict its habitat in a future, warmer world (extrapolation), the model might fail completely. Why? Because under these novel conditions, a new physiological limit—a heat tolerance we never observed before—might suddenly become the dominant factor, rendering the old statistical relationships invalid. This is a profound structural uncertainty about whether our model's "rules" will hold true in a world we've never seen.
This is perhaps the most unique and challenging source of uncertainty. It has little to do with the physics of the climate and everything to do with us. Scenario uncertainty is our lack of knowledge about the future path of human civilization. Will we transition rapidly to renewable energy? Will global population continue to grow? Will we engage in large-scale deforestation or reforestation? These choices generate vastly different trajectories of greenhouse gas emissions and land use change. Since we cannot predict these socioeconomic and political choices, scientists explore a range of plausible futures, from optimistic, low-emissions scenarios to pessimistic, high-emissions ones.
These different sources of uncertainty don't contribute equally at all times. Their relative importance changes dramatically depending on the timescale of the prediction. This evolution tells a powerful story.
Near-Term (The Next Decade or Two): On shorter timescales, the largest source of uncertainty is internal variability. The climate system has enormous thermal inertia, especially the oceans. This means it responds very slowly to external forcing. The difference in warming between a high-emissions and low-emissions path is tiny over just a decade or two. Instead, the chaotic, natural fluctuations of the climate system—the random draw of a strong El Niño or La Niña, for instance—dominate the picture. The signal of forced change is drowned out by the noise of the system's own variability.
Mid-Century (2040-2060): As we look further out, the chaotic noise of internal variability begins to average out. The persistent push of external forcing starts to emerge more clearly. In this time frame, model uncertainty often becomes the dominant source of spread in projections. Different models, with their different structural assumptions and parameters, may disagree on the magnitude of feedbacks (like how clouds will respond to warming), leading to a divergence in their projections.
Long-Term (End of the Century and Beyond): On the longest timescales, a clear winner emerges: scenario uncertainty. Over many decades, the accumulated difference between a world that aggressively curbed emissions and a world that did not becomes enormous. It dwarfs the spread from both internal variability and model differences. This is a profound and ultimately hopeful conclusion. It tells us that over the long run, the single greatest uncertainty in the future of our climate is not an irreducible vagary of nature or a technical detail in a computer model. It is our collective choices.
So, we are faced with this menagerie of uncertainties. What can we do? The scientific process is not just about identifying uncertainty, but about actively trying to reduce and manage it.
One of the most elegant strategies for reducing epistemic uncertainty is the emergent constraint. The logic is akin to a clever piece of detective work. Suppose we want to narrow down the uncertainty in a future quantity we can't measure, like the Equilibrium Climate Sensitivity (how much the world will eventually warm from a doubling of ). We look across our ensemble of diverse models and search for a robust statistical relationship between this future quantity and a variable we can observe in the present-day climate (e.g., a specific pattern of cloud variability). If we find such a physically-grounded relationship, we have found an emergent constraint. We can then use our real-world, high-quality observations of the present-day variable to "constrain" the range of plausible future outcomes. We are using the present to learn about the future.
However, some uncertainty cannot be easily reduced or even quantified. This is particularly true for scenario uncertainty, which is often described as deep uncertainty. We cannot meaningfully assign a probability to future human political and economic choices. It is not a roll of the dice. To pretend we can, for instance by assigning equal probabilities to all scenarios, would create a false sense of precision.
In the face of deep uncertainty, the goal of science shifts from prediction to preparedness. Instead of trying to find the single optimal strategy for a "most likely" future, we use methods like robust decision-making. A city planning a new sea wall would use climate projections not to find the single best height, but to find a design and an adaptation strategy that works reasonably well across the full range of plausible futures—from the best-case to the worst-case scenarios. The aim is to build resilience, creating solutions that are robust to a future we cannot perfectly know. This marks a mature understanding of uncertainty: not as a failure of science, but as an essential feature of the world that we must learn to navigate with wisdom and foresight.
Now that we have grappled with the fundamental nature of climate uncertainty, you might be tempted to think of it as a frustrating fog that obscures our view of the future. But this is not the right way to look at it at all. In science, understanding the character of our ignorance is just as powerful as cataloging what we know. By quantifying uncertainty—by giving it a shape, a size, and a structure—we transform it from a mere nuisance into a powerful tool. It allows us to make predictions that are not just more honest, but more useful. It enables us to make decisions that are not brittle and prone to failure, but robust and resilient.
In this chapter, we will go on a journey to see how a clear-eyed view of uncertainty ripples through diverse fields, from the forests and fields of ecology to the bustling wards of public health and the global stage of economic policy. We will discover that the principles we have discussed are not abstract curiosities; they are the very tools that scientists and decision-makers are using to navigate a changing world.
Let’s begin with a seemingly straightforward question: if the world warms, where will the trees and animals go? An ecologist trying to answer this might build a Species Distribution Model (SDM), a sophisticated statistical tool that maps a species' current habitat based on environmental conditions like temperature and rainfall. The biological part can be modeled with great care. The real headache, it turns out, comes from the climate itself. When the ecologist tries to project the model into the future, they must choose a climate forecast. The trouble is, there isn't just one. Different General Circulation Models (GCMs), even when fed the exact same assumptions about future greenhouse gas emissions, will produce a range of different predictions for future climate. This is a prime example of model uncertainty. The result isn't a single future map for a species, but an ensemble of possible maps, a cloud of potential future homes.
This might seem like a step backward, but it is actually a giant leap forward. Instead of a single, likely-wrong prediction, we have a much more honest assessment of the possibilities. This becomes crucial when we ask a more pointed question: will a species go extinct? Conservation biologists tackle this with Population Viability Analysis (PVA), which is essentially a grand simulation of the future. They create a virtual population of an endangered animal and subject it to the challenges of the next century. They run this simulation not once, but thousands upon thousands of times. In each run, they pick a different plausible future: a different climate trajectory drawn from a GCM ensemble, a different set of biological parameters (like survival rates) drawn from their statistical distributions, and they even allow for the sheer chance of births and deaths (demographic stochasticity).
By running this gauntlet of simulated futures, they don't get a simple "yes" or "no" answer. Instead, they calculate the probability of extinction. This is an immensely more valuable piece of information. A chance of extinction calls for a different management strategy than a chance. The same logic applies to projecting how fast a species' range might shift. By combining multiple climate models in a principled way, such as through Bayesian Model Averaging, we can produce a predictive distribution for the rate of range shift, complete with a mean and a variance that properly accounts for every known source of uncertainty.
This way of thinking translates directly to human health. Consider the spread of a vector-borne disease, like one carried by mosquitoes. The mosquito's life cycle, and therefore the disease's ability to spread, is exquisitely sensitive to temperature and rainfall. The basic reproduction number, , which tells us how many new cases a single infected individual will cause, is a direct function of these climate variables. Therefore, the uncertainty in climate projections translates directly into uncertainty about future disease risk. By propagating climate model uncertainty through our epidemiological models, public health officials can map out not just the most likely future hotspots, but the range of plausible outbreaks they need to prepare for.
Understanding the shape of uncertainty is one thing; making a decision based on it is another. This is where the true power of these ideas comes to life. Imagine you are a public health official tasked with setting a temperature threshold for issuing heat advisories. A simple question, but the answer is buried under layers of uncertainty. First, there is structural uncertainty: is the relationship between temperature and mortality a straight line, or does it curve upward sharply at extreme heat? Choosing the wrong model form could lead to bad decisions. Second, there is parametric uncertainty: even if we pick a model, we don't know the exact values of its coefficients. Finally, there is stochastic uncertainty: the irreducible randomness of daily weather.
By carefully separating these different "flavors" of uncertainty, we can design more robust policies. We can test our heat advisory threshold against multiple model structures and across the plausible range of parameters to find a policy that performs well no matter which is true. We are no longer searching for a single "optimal" answer, but for a strategy that is resilient to our own ignorance.
This logic allows us to do something truly remarkable: put a price on knowledge. Suppose a conservation agency has to choose between two strategies to protect a forest, but the best choice depends on both the future climate and an uncertain biological parameter, . Using the tools of decision theory, the agency can calculate the expected benefit of each strategy given the current uncertainty. But they can also ask: what if we could eliminate the uncertainty about by conducting more research? They can calculate the Expected Value of Perfect Information (EVPI), which is the expected increase in benefit they would get from making the decision with perfect knowledge of . This EVPI gives a concrete monetary value—a budget—for how much they should be willing to spend on research to reduce that specific uncertainty. Uncertainty becomes just another variable in a cost-benefit analysis.
Now, let’s zoom out to the entire global economy. One of the most pressing questions is how to price carbon emissions. Economists use models to estimate the social cost of carbon, which informs the level of a carbon tax. A key input to these models is the climate sensitivity—how much the world will warm from a doubling of . As we’ve seen, this parameter is uncertain. A simple model can reveal something profound: uncertainty doesn't justify inaction; it demands a stronger response. The optimal carbon tax is not simply the tax calculated using the average climate sensitivity. The mathematics, using a second-order approximation, reveals that the tax should include additional positive terms. One term acts as a premium against the risk of unexpectedly high damages. Another, proportional to societal risk aversion (), acts as an insurance premium against the catastrophe of a high-sensitivity world. This is the "precautionary principle" emerging not from a vague feeling, but from the rigorous logic of economics under uncertainty.
What happens when the uncertainty is so profound that we cannot even agree on the probabilities of different futures? This is the domain of deep uncertainty, and it is the reality for many long-term planning problems, from water management to landscape restoration.
Consider a 50-year "rewilding" project to reintroduce an apex predator. The project's success hinges on climate decades from now, but different GCMs give wildly different pictures of that future. To assign a single set of probabilities to these scenarios would be an act of false precision. In these situations, we must change our entire approach to planning.
Instead of seeking a single "optimal" plan, we adopt a strategy of robust satisficing. The goal is no longer to find the absolute best outcome, which is unknowable, but to find a policy that is "good enough" (satisficing) across the widest possible range of plausible futures (robust). We stress-test our strategies against a whole ensemble of challenging scenarios, looking for the one that avoids catastrophic failure in as many of them as possible.
Furthermore, we abandon the idea of a static blueprint. We instead design adaptive pathways. This conceives of a plan as a flexible roadmap, not a fixed itinerary. The plan includes a series of actions, but it also specifies triggers—real-world indicators that we monitor over time. If a trigger is hit (for example, if a herbivore population drops below a critical threshold), it signals that we are heading into a future different from the one we expected, and the plan dictates a pre-agreed switch to a different set of management actions. This creates a policy that learns and adapts as the future unfolds, allowing us to navigate the deep uncertainty of the coming century with humility and foresight.
From the microscopic details of a statistical parameter to the grand sweep of global policy, the message is the same. Embracing and quantifying uncertainty is not a sign of weakness in our science. It is the very source of its strength and its relevance, providing the critical tools we need to understand our world and act wisely within it.