Translated by 36Kr Global, a team focused on technology, business, career, and lifestyle topics, we highlight cutting-edge innovations, fresh perspectives, and emerging trends in the global market.
Editor’s note: Machine learning is deeply integrated into weather and climate modeling systems. From short-term weather forecasts to long-term climate simulations, from parameter optimization to model replacement, AI is continually penetrating the internal architecture of traditional physical models. However, rather than replacing existing paradigms, it strikes a balance between efficiency gains and physical constraints. Real change is not a revolution, but a gradual restructuring. This article is a translated article.
Image source: Getty Images
AI is everywhere, but is it truly a transformative change?
AI is now almost ubiquitous. After typing only half a sentence into the input box, the system is already ready to continue writing. If you try to buy a refrigerator without an internet connection, you will find that the voice assistant is the default setting. As the spread of technology accelerates, people can’t help but wonder. Are we experiencing a true technological breakthrough, or are we simply surrounded by amplified technological narratives?
This problem becomes even more acute as AI enters the field of weather and climate modeling, reaching the boundaries of scientific modeling itself.
Earlier this year, the 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 just an AI-generated image for social media distribution and not the output of an official prediction system. Still, events like this easily create the misconception that weather forecasting is being taken over by large-scale language models. The reality is that meteorologists and climate scientists have not been replaced by nimble engineers.
AI is certainly making inroads in this space, but in a non-radical way, more of a tool-level penetration than a system-level replacement.
It’s machine learning, not LLMs, that’s making a difference
In meteorology and climate science, the term “AI” primarily refers to machine learning and refers to large-scale language models.
At the most basic level, its operation is simple. 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 involves nonlinear function fitting and high-dimensional relational modeling. Although the format continues to evolve, the essence remains consistent. It’s all about discovering statistical patterns within your data space.
The main advance in machine learning is that it no longer relies on manually defining relationships one by one through explicit rules. Instead, it automatically learns structure from the data through optimization algorithms to capture complex correlations that are difficult to express explicitly.
The essence of the training mechanism is high-dimensional optimization
Typical machine learning models are usually trained from scratch. The model structure is first defined. For example, neural networks essentially form a space of tunable parameters, and all functionality is compressed within these parameters.
The system is then fed a large amount of labeled data, including thousands of annotated images. By continuously iterating parameter adjustments, the model gradually approaches the predictions of the actual labels. From an outside perspective, this is a learning process. However, mechanically it resembles a high-dimensional optimization problem.
Behind performance gains are inherent structural limitations
As the capabilities of machine learning models improve, their structural limitations also remain well-defined.
First, there is a limit to generalization ability. The model cannot reliably handle data outside the training distribution. Performance degrades significantly when the input exceeds the range of learned experience. Second, biases in the data are amplified. If a certain type of object in the training data is always associated with a certain background, the model may incorrectly treat background information as an important feature.
Then there’s the question of interpretability. The internal structure of deep models is usually not reducible to human-readable logical chains, and their decision-making processes are more like statistical combinations than rule-based deductions.
The essence of changes in weather models is to improve efficiency
For weather forecasting, machine learning training methods are similar to image recognition training methods, except that the data sources are weather observations taken at different points in time.
Because physical equations do not have to be solved point by point, such models are much faster to compute than traditional weather models. Companies such as Google, Nvidia, Huawei, and Microsoft are launching related models and collaborating with academic institutions to enable capabilities that can compete with traditional predictive systems in certain scenarios.
In February 2025, the European Center for Medium-Range Weather Forecasts introduced the first machine learning model to run in parallel with the traditional IFS system. Trained on reanalysis data, which reconstructs global weather conditions by integrating observational data and physical consistency, the model learns patterns of weather evolution over the next six hours.
Model inputs include variables such as temperature, pressure, wind speed, water vapor, and cloud cover. Instead of explicitly applying physical equations, the model directly learns how these variables vary over space and time.
Serious challenges are beginning to surface
The problem is that these models inherently do not understand physical constraints.
They don’t understand that precipitation cannot be negative, nor can it automatically satisfy the laws of conservation of mass and conservation of energy. During the error minimization process, physically impossible results can occur. This requires additional constraints. For example, correcting negative precipitation values directly to zero.
This “physics-based constraint” is becoming an important design direction for machine learning-based weather models.
But their efficiency benefits are equally obvious. Traditional IFS models consume about 1000 times more computational energy than AIFS and take about 30 minutes to run, whereas machine learning models only require about 3 minutes. This advantage becomes even more pronounced in ensemble predictions.
Extreme weather remains a weakness in the model
While traditional weather forecasting problems are relatively stable, predicting extreme weather is an entirely different challenge.
Because extreme events themselves are so rare, training data is inherently incomplete, making it difficult for models to learn about unprecedented and extreme situations. Research shows that machine learning models often underestimate the frequency and intensity of extreme events, and their errors grow as the degree of extreme events increases. The underlying reason is that the model cannot exceed the training distribution.
As a result, extreme weather events tend to be “compressed” and converge to an average value.
Weather and climate are two fundamentally different problem categories
Weather forecasts focus on the evolution of short-term conditions, whereas climate models target long-term statistical structures such as the energy balance and 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 effects of CO₂ emissions on the long-term climate structure cannot be directly known from historical data, so the laws of physics remain irreplaceable.
Hybrid models are becoming mainstream
Tapio Schneider from the California Institute of Technology is participating in the CliMA project, which aims to build a new generation of climate models that run on GPUs and cloud architectures, using Julia instead of traditional Fortran.
The central idea is not to replace physical models with machine learning, but to incorporate it into local process modules.
Climate models are composed of multiple subsystems, and many microscopic processes cannot be resolved explicitly and can only be approximated through parameterization. ClMA replaces some parameterized modules, such as the snow accumulation process, with machine learning while incorporating water conservation constraints.
This approach works well because many patterns are transferable within the current climate range. However, in complex scenarios such as cloud systems where future states can exceed the distribution of historical data, models still face the risk of failure.
This clarifies the strategy of using machine learning in areas where it is learnable and preserving physical models in other areas.
Model optimization enters the “meta-problem stage”
Some physical model parameters can be adjusted, a process known as model calibration. Machine learning can be used to optimize this process.
The NASA GISS team adjusted parameter combinations to generate 450 sets of simulation results, running each simulation for a year and comparing them to observed data, including differences in cyclone numbers and energy budgets. They then trained a machine learning model to learn the relationship between parameters and errors to derive the optimal parameter combination, which is essentially a high-dimensional optimization problem.
Simulator and surrogate models
Machine learning can also be applied to model compression. That is, you train a lightweight model to mimic a complex physical model, thereby building a surrogate system.
These models, trained on expensive simulation data, can be used to quickly explore different emissions scenarios without repeatedly consuming supercomputing resources. Accuracy is slightly reduced, but computational cost is significantly reduced.
More importantly, this creates a closed-loop structure. The simulator generates data that trains more efficient models.
This methodology is spreading to areas such as ice sheets, sea ice, and ocean circulation.
Achieving nearly half the performance at less than 1% computational cost would make this approach practically meaningful for engineering applications.
The black box problem still persists
The central debate surrounding machine learning remains the black box issue. Although all the terms in a physical model have clear physical meanings, the internal structure of a neural network cannot usually be explained.
The core value of scientific models is to understand reality, not just predict outcomes. If a model cannot be explained, its failure boundaries cannot be determined.
For this reason, the climate science community remains cautious about using machine learning. However, in reality, this methodology is still part of the scientific method system and ultimately requires experimental and numerical validation.
Explainable AI is working to reduce opacity
One approach is to track the influence of input variables on the output. For example, in precipitation prediction models, researchers found that the models initially relied heavily on lightning data. After removing this information, the model transitioned to using infrared and water vapor data and began identifying structural features of the cloud.
This partially alleviates the black box problem by allowing researchers to verify whether their models are consistent with physical mechanisms.
The true position of machine learning
Machine learning is part of big data science, and its core value lies in extracting structure from large datasets.
The impact depends on the application scenario, rather than following a one-size-fits-all conclusion.
In the field of weather forecasting, calculation methods and efficiency structures have already undergone significant changes. In climate science, it is still in the exploratory stage, but it is gradually penetrating the internal architecture of models.
As New York University’s Laure Zanna says, this is just one factor among many for now, but it could play a more important role in hypothesis testing and simulation reliability in the future.
Translator: Xiaochuan
