Modern Earth system models are accurate but computationally expensive. Machine learning promises to speed up simulations, but trusting such models requires evidence that they have learned physical laws rather than merely memorized statistics. Previous studies assessed neural network models on synoptic scales or short-term forecasts, but their behavior under long-term anomalous forcings remained unknown. Alongside this work on the preprint server are studies using neural networks for gravitational maps of galaxies, improving climate models, and identifying the cosmic web, highlighting a broader trend toward hybridizing physics and machine learning.
The authors designed three series of idealized experiments with the NeuralGCM model, specifically the version with stochastic noise, a resolution of 2.8 degrees, and 37 vertical levels. For cyclones, they selected 526 winter extratropical cyclone cases over the North Pacific from 1979 to 2020 and ran the model four days forward from genesis, comparing against ERA5 reanalysis and CMIP6 climate models. For El Niño, they used composites of anomalies from strong events defined by the Niño3.4 index. For warming, they conducted long simulations over 36 years with uniform ocean surface heating of 2 and 4 kelvin, as well as short ensembles of 20 members with heating from 1 to 4 kelvin. Skill was evaluated using spatial correlation coefficients and normalized root-mean-square errors.
On the synoptic scale, NeuralGCM reproduces the evolution of extratropical cyclones with a correlation coefficient above 0.8 for 500 hPa geopotential height in the first days, comparable to CMIP6 models. By day three, the model overestimates the spatial size of cyclones by 50–63%, linked to an overestimated gradient of potential vorticity. For El Niño, the model captures the PNA teleconnection with a correlation of 0.86 for 200 hPa geopotential height, falling within the spread of CMIP6 models, but overestimates the amplitude of the nonlinear asymmetry of the response. Under uniform warming, global surface temperature increases by 0.86 kelvin per kelvin of ocean heating, and precipitation by 4.14% per kelvin, close to CMIP6 estimates. However, in the upper troposphere and lower stratosphere, the model produces excessive cooling in the tropics and excessive warming in polar regions, causing the jet stream not to shift poleward as in physical models.
This work shows that a hybrid model with a dynamical core can generalize beyond its training data, something purely neural network models have not achieved. Unlike neighboring studies where neural networks merely emulate physical models or solve narrow tasks, this work tests full climate dynamics across three time scales. The stratospheric errors point to a fundamental limitation: the model did not receive carbon dioxide and ozone concentrations as inputs, so its stratospheric response is unphysical. This raises the question of the minimum physics required in hybrid models for trustworthy climate projections.
The field will likely move toward incorporating additional Earth system components, such as ocean, ice, and biogeochemistry, into hybrid models. Increasing resolution and expanding training data, including satellite observations of the stratosphere, could correct errors in the upper atmosphere. Standardized test suites for machine-learned climate models, similar to CMIP protocols, are likely to emerge.
The results will influence the development of the next generation of Earth system models, where machine learning replaces expensive parameterizations. They are also important for operational weather forecasting and climate risk assessments, where fast ensemble simulations are needed.
The next step will be to quantitatively assess the contributions of the machine learning component and the dynamical core to model errors, for example, by disabling the machine learning block. Additionally, tests need to be expanded to other types of forcings, such as changes in greenhouse gas concentrations.
This work directly addresses an unresolved problem in climate modeling: how to combine physical consistency with computational efficiency. The stratospheric errors reflect a long-standing difficulty in reproducing the vertical structure of the atmosphere, while the cyclone overestimation points to the issue of scale-dependent parameterization of eddies. For climatology, the key question is whether machine-learned models can be trusted outside their training distribution, and this work provides a partial affirmative answer.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The authors tested the hybrid NeuralGCM model at three levels: synoptic cyclones, the response to El Niño, and uniform ocean warming. The model was compared against the ERA5 reanalysis and CMIP6 models, evaluating correlations and errors. Systematic errors were identified: overestimation of cyclone sizes and an incorrect stratospheric response.
Where its strength is
The work shows that a hybrid model with a dynamical core can generalize beyond its training data, reproducing key climate responses. This is a strong argument for trusting such models for long-term forecasts.
Where it could go next
Worth trying
Why this may matter
If the identified systematic errors can be fixed, hybrid models could become a reliable tool for studying climate change, combining the speed of machine learning with physical credibility.
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.54, against a usual 0.64. It looks like few have drilled nearby.
The NeuralGCM model was trained on ERA5 reanalysis and IMERG satellite precipitation only for 2001–2018, yet it managed to reproduce the response to warming that was not present in that data. It is as if a student who learned physics from a textbook passed an exam on problems not covered in the book.
The idea of hybrid intelligence, where machine learning complements physical laws, echoes Stanisław Lem's novel 'Solaris', in which an ocean-mind generates models of visitors from memories. Here, the model generates climate responses based on learned patterns, but sometimes creates phantoms in the stratosphere.