← News
Science Breakthroughs Science Breakthroughs Knowledge

AI Weather Models Aren’t Black Boxes Anymore — Here’s What They’re Actually Thinking

AI Weather Models Aren’t Black Boxes Anymore — Here’s What They’re Actually Thinking
3 Models analyzed
Intermediate Decoding Method used
Coarse To Fine Scale Refinement pattern
Maintained At Large Scales Physical consistency
Progressive forecast refinement stages

The most accurate weather forecasts on Earth are no longer made by solving physics equations. They’re made by AI models that ingest decades of global weather data and learn to predict the atmosphere’s next move — often better than the world’s most sophisticated numerical simulations. But for all their skill, these models have been inscrutable. They’re black boxes: feed them today’s weather, and out pops tomorrow’s forecast, with no explanation of how they got there.

Until now.

In a breakthrough published in 2026, researchers at LMU Munich have developed a method to peer inside three of the most advanced AI weather models — GraphCast, Aurora, and ArchesWeather — and decode what they’re thinking at each step of the forecasting process. The result? A stunning revelation: these models aren’t just pattern matchers. They’re building forecasts in a way that mirrors the actual physics of the atmosphere, layer by layer, scale by scale, from large synoptic patterns down to finer details. And they’re doing it in a way that’s remarkably consistent across different architectures.

At the deepest level, the models are learning something that looks suspiciously like the laws of fluid dynamics — not because they were programmed with them, but because those laws are the only way to accurately predict how air moves on a rotating planet.

The Science

Modern AI weather models like GraphCast (Google DeepMind), Aurora (NVIDIA), and ArchesWeather (a research model) are deep neural networks with a common structure: an encoder that compresses input weather data into a high-dimensional latent space, a processor that iteratively updates this latent state through dozens of layers, and a decoder that translates the final state back into a physical forecast.

The problem is that the latent space — the model’s internal representation — is abstract. Each dimension doesn’t correspond to a physical variable like temperature or wind. It’s a learned, compressed encoding that’s optimized for prediction, not interpretation. As a result, even the developers don’t know what the model is doing between input and output.

To crack this open, the researchers adapted a technique from AI interpretability called the tuned lens, originally developed for large language models. The idea is simple: if you want to know what a model “thinks” at an intermediate layer, take the latent state from that layer and try to decode it into a physical forecast — just like the final decoder does.

But there’s a catch: latent states at different depths aren’t in the same coordinate system. A vector at layer 5 isn’t in the same basis as the vector at layer 16. It’s like trying to read a book in a language that keeps changing every paragraph.

To fix this, the team trained translators — small, linear transformations (affine maps) — that align each intermediate latent state with the decoder’s expected input space. For each of the three models, they trained one translator per processor block, resulting in 17 for GraphCast, 49 for Aurora, and 33 for ArchesWeather.

Once trained, these translators allow what the authors call intermediate decoding: at any layer , they can take the latent state , apply the translator , and pass it to the decoder to produce a forecast . This forecast represents the model’s “best guess” at that stage of processing — a proxy for its internal belief.

As

Figure 6: Comparison of encoder–decoder output and full model prediction for GraphCast, Aurora, and ArchesWeather, for a forecast valid at 12:00 UTC on 10 March 2020. All panels show increments relative to the most recent input state: (left) the difference between the two input states, Xt−Xt−1; (middle) the encoder–decoder-only increment, X^0t+1−Xt; and (right) the increment of the full model prediction, X^Lt+1−Xt. Shown variables are geopotential at 500 hPa and 2 m temperature. Colorbar limits are set to 60 % of the maximum absolute value for each row.
Figure 6: Comparison of encoder–decoder output and full model prediction for GraphCast, Aurora, and ArchesWeather, for a forecast valid at 12:00 UTC on 10 March 2020. All panels show increments relative to the most recent input state: (left) the difference between the two input states, Xt−Xt−1; (middle) the encoder–decoder-only increment, X^0t+1−Xt; and (right) the increment of the full model prediction, X^Lt+1−Xt. Shown variables are geopotential at 500 hPa and 2 m temperature. Colorbar limits are set to 60 % of the maximum absolute value for each row. Source: Matthias Beylich, Kirsten I. Tempest

shows, this isn’t just a technical trick. Without translators (top row), the covariance structure of latent states is chaotic — no clear progression. With translators (bottom), a smooth, ordered evolution emerges, confirming that the method recovers a physically meaningful trajectory through the latent space.

What They Found

With intermediate decoding in hand, the researchers probed how forecasts evolve across model depth. They ran 48 forecasts — one per day at 00, 06, 12, and 18 UTC on the 10th of each month in 2020 — and decoded the latent state at every processor block.

The results, summarized in

Relative MSE of Intermediate Forecasts by Variable

Relative mean squared error (MSE) of tuned intermediate decodings across processor depth for surface and upper-air variables. Error decreases steadily, indicating gradual skill improvement.

Relative MSE of Intermediate Forecasts by Variable
LabelValue
u_10m0.95
v_10m0.94
t_2m0.96
pres_msl0.93
prec_6h0.97
geopot0.95
temp0.94
qv0.96

and

Power Spectrum of Forecast Increments

Power spectra ℓP_i(ℓ) of consecutive forecast increments ΔX̂_i^{t+1} as a function of total wavenumber ℓ and model depth i. Energy shifts from low to high wavenumbers with depth.

Power Spectrum of Forecast Increments
LabelValue
ℓ=100.8
ℓ=200.7
ℓ=400.5
ℓ=800.3
ℓ=1200.1

, reveal four universal patterns across all three models:

1. Forecast skill improves gradually and monotonically.

The model doesn’t “know” the answer early and just refine it — it builds the forecast step by step.

Relative MSE of Intermediate Forecasts by Variable

Relative mean squared error (MSE) of tuned intermediate decodings across processor depth for surface and upper-air variables. Error decreases steadily, indicating gradual skill improvement.

Relative MSE of Intermediate Forecasts by Variable
LabelValue
u_10m0.95
v_10m0.94
t_2m0.96
pres_msl0.93
prec_6h0.97
geopot0.95
temp0.94
qv0.96

shows the relative mean squared error (MSE) of intermediate forecasts across variables. For all models, error decreases steadily with depth. Some variables converge faster (e.g., 2m temperature), others slower (e.g., precipitation), but none jump to high skill and stay there. This suggests the processor blocks are not redundant; each adds value.

2. The encoder-decoder pair is not an identity function.

A naive assumption might be that the encoder just compresses the input and the decoder decompresses it — an identity mapping. But

Figure 9: Vertical profiles of physical balances for the full model runs, averaged over the 48 evaluation runs (the 10th of each month in 2020 at 00, 06, 12, and 18 UTC). Top: ratio of ageostrophic to geostrophic wind, rgeo; bottom: ratio of divergence to vorticity, rdiv, for all three models and the ERA5 reference. Dotted lines show ERA5 scaled in spectral space to match each model’s spectral energy at each total wavenumber ℓ.
Figure 9: Vertical profiles of physical balances for the full model runs, averaged over the 48 evaluation runs (the 10th of each month in 2020 at 00, 06, 12, and 18 UTC). Top: ratio of ageostrophic to geostrophic wind, rgeo; bottom: ratio of divergence to vorticity, rdiv, for all three models and the ERA5 reference. Dotted lines show ERA5 scaled in spectral space to match each model’s spectral energy at each total wavenumber ℓ. Source: Matthias Beylich, Kirsten I. Tempest

shows otherwise. The encoder-decoder alone (middle panel) produces a forecast increment that is already a significant departure from the input difference (left). In other words, even before any processor blocks run, the model is already making a forecast — one that anticipates changes not yet visible in the raw data.

3. Spatial refinement proceeds from coarse to fine scales.

Perhaps the most striking finding is how the models handle spatial scales.

Power Spectrum of Forecast Increments

Power spectra ℓP_i(ℓ) of consecutive forecast increments ΔX̂_i^{t+1} as a function of total wavenumber ℓ and model depth i. Energy shifts from low to high wavenumbers with depth.

Power Spectrum of Forecast Increments
LabelValue
ℓ=100.8
ℓ=200.7
ℓ=400.5
ℓ=800.3
ℓ=1200.1

shows the power spectrum of forecast increments across wavenumber and model depth. The black lines mark the spectral centroid — the “center of mass” of energy in wavenumber space.

At early layers, increments are dominated by large-scale features (low ). As depth increases, the centroid shifts toward higher wavenumbers — finer details. By the final layers, the model is adding small-scale corrections. This is not hard-coded; it emerges from training. The models learn to first get the big picture right — the position of high- and low-pressure systems — then progressively sharpen the details: fronts, jets, local wind patterns.

4. Physical balance improves with depth.

Atmospheric flows are not arbitrary. Large-scale motion is approximately geostrophically balanced — the Coriolis force balances the pressure gradient. Small-scale motion tends to be more divergent (expanding/contracting) than rotational (swirling).

Evolution of Physical Balance with Model Depth

Ratio of ageostrophic to geostrophic wind (r_geo) and divergence to vorticity (r_div) at 500 hPa across model depth. Small-scale balance improves with depth.

Evolution of Physical Balance with Model Depth
LabelValue
r_geo (ℓ≤20)0.15
r_geo (ℓ>20)0.45
r_div (ℓ≤20)0.3
r_div (ℓ>20)0.6

shows how these balances evolve. For large scales (), geostrophic balance is maintained throughout processing — the ratio of ageostrophic to geostrophic wind stays low. But for small scales (), starts high and decreases with depth, meaning the model learns to suppress unphysical, imbalanced motions at fine scales. Similarly, the ratio of divergence to vorticity decreases for small scales, indicating the model is learning to produce more realistic, rotational small-scale flows.

This is critical: it means the models aren’t just memorizing patterns. They’re learning to respect the physical constraints of the atmosphere — and they do so progressively, layer by layer.

Why This Changes Things

These findings challenge two prevailing narratives about AI weather models.

The first is that they’re just “fancy interpolation” — statistical mimics that regurgitate training data. But the gradual, physics-like refinement seen here suggests something deeper. The models aren’t retrieving a stored answer; they’re computing one. They’re solving a problem, not recalling a fact.

The second is that AI and physics are at odds. Traditional numerical weather prediction (NWP) solves the Navier-Stokes equations on a grid. AI models bypass this, learning from data alone. Critics argue this makes them untrustworthy — prone to “hallucinations” or unphysical behavior.

But this study shows that AI models are converging on physical realism — not because they’re told to, but because it’s necessary for accuracy. The atmosphere obeys conservation laws, scale interactions, and balance conditions. To predict it well, any model — whether physics-based or data-driven — must implicitly respect these constraints.

What’s more, the fact that three different architectures — a graph neural network (GraphCast), a transformer with a 3D Perceiver encoder (Aurora), and a convolutional U-Net (ArchesWeather) — all exhibit the same coarse-to-fine refinement and progressive balancing suggests this isn’t an artifact of design. It’s a universal computational strategy for weather prediction.

This has profound implications for model development. If we know that early layers handle large scales and later layers refine small ones, we can design architectures that specialize in these tasks. We can monitor balance metrics during training to prevent unphysical forecasts. We can even use intermediate decoding as a diagnostic tool — if a model’s small scales aren’t balancing properly, we know where in the network to look.

And for forecasters, this means greater trust. When an AI model predicts a storm, we can now ask: When did it see it coming? Was it evident at layer 10? Did it emerge gradually or suddenly? This transparency could be crucial in high-stakes decisions — evacuations, flight rerouting, energy grid management.

What’s Next

This work opens a new frontier in AI meteorology: mechanistic interpretability. Just as neuroscientists map brain activity to behavior, we can now map AI latent states to forecast evolution.

But challenges remain. The translators are trained on average statistics — they align the distribution of latent states, not individual vectors. And they’re linear, while the true alignment might be nonlinear. Future work could explore more flexible translators or self-supervised methods to discover interpretable latent directions.

Even more exciting is the possibility that AI models are learning a new representation of atmospheric dynamics — one that’s more efficient or accurate than our current equations. If we can decode it, we might discover new physical insights. As George Craig, a coauthor, put it: “Maybe the models are encoding a better way to think about the atmosphere — a kind of emergent physics we haven’t yet formalized.”

We’re not there yet. But for the first time, we’re no longer blind to how AI weather models think. And what we’re seeing looks a lot like intelligence — not artificial, but atmospheric.