AI has stormed into weather and climate modeling, but it is still far from triggering a "revolution"

Translated by 36Kr Global, a team focused on technology, business, career, and lifestyle topics, with a spotlight on cutting-edge innovations, fresh perspectives, and emerging trends from global markets.

Editor's note: Machine learning is deeply integrating into weather and climate modeling systems. From short-term weather forecasting to long-term climate simulation, from parameter optimization to model substitution, AI is continuously penetrating the internal architecture of traditional physical models. However, it does not replace existing paradigms, but rather strikes a balance between efficiency improvement and physical constraints. The real change is not a revolution, but a gradual reconstruction. This article is a translated piece.

Image source: Getty Images

AI is everywhere — but is there real transformative change?

Nowadays, AI is nearly ubiquitous. You have barely typed half a sentence in an input box when the system is already ready to continue writing for you; you try to buy a fridge without internet connectivity, only to find that voice assistants have become the default configuration. As technological penetration accelerates, people can't help but wonder: are we experiencing a genuine technological leap, or merely surrounded by an amplified technological narrative?

This question becomes even more acute when AI enters the field of weather and climate modeling, even touching the very boundaries of scientific modeling itself.

Earlier this year, an office of the U.S. National Weather Service released a forecast map featuring non-existent city names in Idaho, such as "Whata Bod" and "Orangeotild". It was later confirmed that this was merely an AI-generated image intended for social media distribution, not an output from the official forecasting system. Even so, such incidents easily create a misperception that weather forecasting is being taken over by large language models. In reality, meteorologists and climate scientists have not been replaced by prompt engineers.

AI is indeed making inroads into this field, but in a non-radical way — closer to tool-level penetration rather than system-level replacement.

It's not LLMs that are making a difference, it's machine learning

In meteorology and climate science, the term "AI" primarily refers to machine learning, not large language models.

At its most fundamental level, what it does is straightforward: extract patterns from data and establish mapping relationships between inputs and outputs.

The most basic form is linear regression, which fits a trend line to data points. Moving up the complexity scale, there is nonlinear function fitting or higher-dimensional relationship modeling. The forms keep evolving, but the essence remains consistent: all are about discovering statistical patterns in data space.

The key advancement of machine learning is that it no longer relies on manually defining relationships one by one through explicit rules. Instead, it automatically learns structures from data via optimization algorithms, capturing complex correlations that are difficult to express explicitly.

The essence of the training mechanism is high-dimensional optimization

A typical machine learning model is usually trained from scratch. The model structure is defined first — for example, a neural network, which essentially forms a space of tunable parameters, with all its capabilities compressed within these parameters.

Then the system is fed a large volume of labeled data, such as thousands of annotated images. Through continuous iterations of parameter adjustment, the model gradually makes predictions that approach the real labels. From an external perspective, this is a learning process. Mechanistically, however, it is more akin to a high-dimensional optimization problem.

Behind performance improvements lie inherent structural limitations

As the capabilities of machine learning models grow, their structural limitations also remain clearly defined.

First, there is the boundary of generalization ability. Models cannot stably process data outside their training distribution; once inputs exceed the scope of their learned experience, performance degrades significantly. Second, data biases are amplified. If a certain type of object in the training data is always tied to a specific background, the model may mistakenly treat background information as a critical feature.

Then there is the interpretability problem. The internal structure of deep models usually cannot be reduced to human-readable logical chains, and their decision-making processes are closer to statistical coupling rather than rule-based deduction.

The essence of changes in meteorological models is efficiency improvement

In weather forecasting, the training method for machine learning resembles that of image recognition — except the data sources are meteorological observations taken at different points in time.

Since there is no need to solve physical equations point by point, such models are significantly faster in computation than traditional weather models. Companies including Google, Nvidia, Huawei, and Microsoft have launched related models and collaborated with academic institutions, achieving capabilities to compete with traditional forecasting systems in certain scenarios.

In February 2025, the European Centre for Medium-Range Weather Forecasts deployed its first machine learning model, running in parallel with the traditional IFS system. Trained on "reanalysis data" — which reconstructs global weather states by integrating observations and physical consistency — the model learns the evolution patterns of weather over the next 6 hours.

Model inputs include variables such as temperature, air pressure, wind speed, water vapor, and cloud cover. Instead of explicitly applying physical equations, the model directly learns how these variables change across space and time.

Critical challenges are beginning to emerge

The problem is that these models do not inherently understand physical constraints.

They do not realize that precipitation cannot be negative, nor do they automatically satisfy the laws of mass conservation and energy conservation. During the error minimization process, they may produce physically implausible results. This is why additional constraints are required — for example, directly correcting negative precipitation values to zero.

This "physics-informed constraint" is becoming a key design direction for machine learning-based meteorological models.

Yet their efficiency advantage is equally obvious. The traditional IFS model consumes approximately 1000 times more computational energy than AIFS, taking about 30 minutes to run, while the machine learning model only needs roughly 3 minutes — with the advantage becoming even more pronounced in ensemble forecasting.

Extreme weather remains a weak point for models

Conventional weather forecasting problems are relatively stable, but extreme weather forecasting is a completely different challenge.

Because extreme events themselves are extremely rare, the training data is inherently incomplete, making it difficult for models to learn about unprecedented extreme situations. Research shows that machine learning models often underestimate the frequency and intensity of extreme events, and their errors expand as the degree of extremity increases. The fundamental reason is that models cannot transcend their training distribution.

The result is that extreme weather is "compressed" and tends to converge toward average values.

Weather and climate are two fundamentally distinct categories of problems

Weather forecasting focuses on short-term state evolution, while climate models target long-term statistical structures — such as energy balance or the response of the climate system.

From a modeling perspective, one is an initial value problem, and the other is a boundary condition problem. For example, the impact of CO₂ emissions on long-term climate structure cannot be directly learned from historical data, which is why physical laws remain irreplaceable.

Hybrid models are becoming the mainstream direction

Tapio Schneider from Caltech is involved in the CliMA project, which aims to build a new generation of climate models that run on GPU and cloud architectures, using Julia instead of the traditional Fortran.

The core idea is not to replace physical models with machine learning, but to embed it into local process modules.

Climate models are composed of multiple subsystems, where a large number of microscopic processes cannot be resolved explicitly and can only be approximated via parameterization. CliMA is replacing some parameterized modules — such as snow accumulation processes — with machine learning, while incorporating water conservation constraints.

Within the current climate range, this approach performs well, as many patterns are transferable. However, in complex scenarios such as cloud systems, where future states may exceed the distribution of historical data, the model still faces the risk of failure.

This makes the strategy clear: use machine learning in learnable domains, and retain physical models in others.

Model optimization enters the "meta-problem stage"

Some physical model parameters are adjustable, a process known as model calibration. Machine learning can be used to optimize this process.

The NASA GISS team generated 450 sets of simulation results by adjusting parameter combinations, with each simulation running for one year, and compared them with observational data — including differences in cyclone counts and energy budgets. They then trained a machine learning model to learn the relationship between parameters and errors, and derived the optimal parameter combination, which is essentially a high-dimensional optimization problem.

Simulators and surrogate models

Machine learning can also be applied to model compression: training lightweight models to mimic complex physical models, thereby constructing surrogate systems.

Trained on high-cost simulation data, these models can be used to rapidly explore different emission scenarios without repeatedly consuming supercomputing resources. While their accuracy is slightly reduced, their computational costs are drastically cut down.

More importantly, this creates a closed-loop structure: simulators generate data, which in turn trains more efficient models.

This methodology is expanding to fields such as ice sheets, sea ice, and ocean circulation.

If we can achieve nearly half the performance with less than 1% of the computational cost, this approach becomes practically meaningful for engineering applications.

The black box problem persists

The core controversy around machine learning remains the black box issue. Every term in a physical model has a clear physical meaning, while the internal structure of a neural network is usually unexplainable.

The core value of a scientific model lies in understanding reality, not just predicting outcomes. If we cannot explain a model, we cannot identify its failure boundaries.

For this reason, the climate science community remains cautious when using machine learning. In practice, however, this methodology is still part of the scientific method system, as it ultimately requires experimental and numerical verification.

Explainable AI is working to reduce opacity

One approach is to trace the influence of input variables on outputs. For example, in a precipitation forecasting model, researchers found that the model initially relied heavily on lightning data. After removing this information, the model shifted to using infrared and water vapor data, and began to identify cloud structural features.

This allows researchers to verify whether the model aligns with physical mechanisms, thus partially mitigating the black box problem.

The true position of machine learning

Machine learning is part of big data science, with its core value lying in extracting structures from massive datasets.

Its impact depends on the application scenario, rather than following a one-size-fits-all conclusion.

In the weather forecasting domain, it has already significantly transformed computational methods and efficiency structures. In climate science, it is still in the exploratory stage, yet gradually penetrating the internal architecture of models.

As Laure Zanna from NYU puts it, it is currently just one of many factors, but it may play a more important role in hypothesis testing and simulation reliability in the future.

Translator: Xiaochuan