When it comes to climate modeling, grid cell size matters. The smaller the cell, the more accurately the model can reproduce local weather phenomena – thunderstorms, hurricanes, and terrain features. However, there's a problem: traditional high-resolution physical models require enormous computing power. Simply put, to run such a model, you need a supercomputer.
The team at the Allen Institute for AI proposed a different approach. They released HiRO-ACE – a machine learning-based model that operates with a resolution of about a kilometer yet does not require specialized infrastructure. It can even be run on a standard computer.
What is Kilometer Resolution and Why is it Needed
What is Kilometer Resolution and Why is it Needed 🌍
In climatology, model resolution refers to the size of a single grid cell into which the atmosphere is divided. Global models usually operate with a resolution of tens or hundreds of kilometers. This is sufficient for general forecasts but insufficient for understanding local effects: how the climate will change in a specific valley, city, or coastal zone.
Kilometer resolution allows you to “see” such details. The model begins to distinguish convective storms, mountain airflows, and the influence of urban development. This is important not only for science but also for planning: where to build, how to prepare for extreme events, and which regions will be at risk.
The problem is that physical models with such resolution are slow and expensive to run. For example, to model the climate for several decades ahead, it might take several months of supercomputer work.
How the HiRO-ACE Climate Model Works
How HiRO-ACE Works
HiRO-ACE stands for High-Resolution Output – Ai2 Climate Emulator. It is an emulator trained on data from a physical model but runs orders of magnitude faster.
At its core lies a transformer architecture – the same basic idea found in language models, only applied to climate data. The model was trained on the results of runs of MPAS-A – a physical atmosphere model with a resolution of about 3.75 km. The team used the SSP3-7.0 scenario – one of the standard climate scenarios describing the possible evolution of emissions and temperatures through the end of the century.
After training, HiRO-ACE learned to predict the state of the atmosphere one step ahead, relying on current conditions. This allows it to “unroll” the forecast for decades, step by step.
An important detail: the model does not try to replace physics. It emulates the behavior of an already existing physical model but does so much faster. One could say it has learned the patterns that the physical model calculates directly.
HiRO-ACE Model Accuracy
How Accurate Is It
The team conducted a comparison with the original MPAS-A physical model. HiRO-ACE showed good agreement on key climate variables: temperature, precipitation, pressure, and humidity. The model reproduces spatial patterns, seasonal cycles, and long-term trends.
Of course, there are nuances. The emulator does not always accurately convey extreme events – rare but important phenomena like severe storms. This is expected: machine learning models handle typical situations better than outliers. But for most climate research, such accuracy is already useful.
Another advantage is speed. HiRO-ACE runs approximately 100 times faster than the physical model. What used to require months of computation can now be obtained in days or even hours.
Accessibility of HiRO-ACE Climate Model
Accessibility Is the Key Feature
The Allen Institute has made the model publicly available. It can be downloaded and run on standard hardware – no supercomputers or cloud clusters with hundreds of GPUs are needed. This is a game-changer.
Previously, detailed climate modeling was available only to large research centers. Now, independent scientists, small universities, startups, and even enthusiasts can experiment with kilometer-resolution data.
This does not mean that HiRO-ACE is suitable for all tasks. But it opens doors: one can quickly test a hypothesis, assess the impact of local factors, and prepare data for analysis without waiting for months.
Future Development of HiRO-ACE
What's Next
HiRO-ACE is not the finish line, but rather a beginning. The model is trained on a single climate scenario. In the future, the team plans to expand coverage: add other scenarios, improve accuracy regarding extreme events, and possibly increase the resolution even further.
The question of interpretation also remains. A machine learning model is a black box. It gives a result, but it is not always clear exactly why. Physical models are more transparent in this sense: one can trace which processes led to a particular conclusion. Emulators do not yet provide such depth of understanding.
Nevertheless, the direction is promising. The combination of physical models and machine learning can offer the best of both worlds: the accuracy of the former and the speed of the latter. HiRO-ACE shows that this works not only in theory but also in practice.
Who Benefits from HiRO-ACE Climate Data
Who Could Benefit From This
The obvious audience is climatologists and atmospheric researchers. But not only them. Detailed climate data is needed in other fields as well.
Urban planners can assess risks for specific neighborhoods. Engineers can design infrastructure taking future changes into account. Ecologists can model the impact of climate on ecosystems. Insurance companies can reassess risks. And so on.
The openness of the model makes it a tool for broad application. There is no need to negotiate access, pay for computations, or wait in line for a supercomputer. Download, run, get the result.
Of course, serious conclusions will still require expertise. The model is a tool, not the ultimate truth. But it significantly lowers the barrier to entry into a field that previously seemed inaccessible.