A new AI model learned the rules of wind so well it can “see” around turbines and hillsides in seconds instead of hours
Siting a wind farm or estimating structural loads on a solar array means running detailed fluid simulations — sometimes dozens of them — to understand how air moves over ridgelines, around turbine blades, and through gaps between panels. Each simulation can take hours. Run enough of them to compare layouts or test design variants, and the process becomes a serious bottleneck.
The physics involved leaves little room for error. Get the wind field wrong over complex terrain and the consequences range from underperforming turbines to structures that weren’t built to handle the real loads. A new approach may now offer a way out of that tradeoff.
The bottleneck slowing renewable energy siting
Placing wind turbines and solar installations isn’t just a land-use decision. It requires knowing, in detail, how air accelerates over ridges, decelerates in valleys, and twists through gaps between structures. Those details determine both energy yield and structural safety.
When engineers need to compare dozens of layout options, or test how a design change affects loads across varied terrain, that cost compounds quickly.
The standard tool is Reynolds-averaged Navier-Stokes simulation — RANS — typically run through software like OpenFOAM. RANS solves the governing equations of fluid motion across a three-dimensional grid, and it’s accurate. It’s also slow.
A single simulation can take hours. When engineers need to compare dozens of layout options, or test how a design change affects loads across varied terrain, that cost compounds quickly. Complex terrain makes it worse — hills, ridges, and valleys create flow patterns that flat-ground approximations simply can’t capture. The more irregular the landscape, the more simulations you need, and the harder it becomes to justify running all of them.
A surrogate model that learns the physics of airflow
The new approach, described in a preprint published in Wind Energy Science, trains a neural network to reproduce what OpenFOAM would output — but at a fraction of the computational cost. This kind of model is called a surrogate: it doesn’t solve the fluid equations from scratch. It learns to predict their answers.
The researchers developed two model families. One is a hybrid point-wise model combining a coordinate multilayer perceptron with two-dimensional convolutional terrain encoders. The other is a three-dimensional grid-based U-Net. Both architectures handle the full spatial complexity of wind and pressure fields.
Each model uses a two-stage cascade structure. A terrain-scale background model predicts the broad flow pattern across domains up to 3×3 km at 30 m resolution, and a structure-scale residual refiner then corrects that prediction near turbines or solar panels at 0.5 m resolution. This dual-resolution design handles landscape-level siting questions and fine-grained structural load estimates within a single framework.
Why physics penalties make the model more trustworthy
Training a neural network on CFD data alone teaches it to match outputs — it doesn’t guarantee those outputs make physical sense. The researchers addressed this by adding physics regularization: penalties applied whenever predictions violate mass and momentum conservation.
Penalties also apply when predictions depart from prescribed inflow profiles, outlet pressure conditions, and boundary constraints. The model is penalized not just for being wrong relative to reference data, but for being physically inconsistent. That’s a meaningful distinction.
The practical effect is measurable. Physics regularization roughly halved continuity and momentum residuals in structure regions and reduced them by 15–65% at terrain scale. Raw wind speed prediction accuracy didn’t shift much — the gain isn’t in the numbers. It’s in whether the predicted flow fields actually obey the laws governing airflow.
Active learning cuts the data needed to train the model
Building a reliable surrogate still requires CFD simulations for training data. Active learning addresses that cost by making the selection of new training cases strategic rather than arbitrary.
Starting from 120 simulation domains, the team expanded the training set to 292 by selecting new cases based on where the model was most uncertain. Score-based selection outperformed random selection at the same simulation budget. Terrain-scale test error dropped by about 30%, with wind speed normalized RMSE falling from 0.26 to 0.18. Fewer expensive simulations were needed to reach that accuracy level.
Seconds instead of hours — and what each model does best
The surrogate returns full three-dimensional wind and pressure fields in seconds. That makes repeated micro-siting comparisons and preliminary load studies practical in a way that hour-long CFD runs simply aren’t.
The two model families perform differently depending on the situation. The hybrid model achieves the best overall errors and supports flexible point-wise inference. The U-Net better resolves narrow channeling jets — the accelerated airflow that forms between closely spaced structures — which matters because those jets represent concentrated structural loads that volume-averaged error metrics can miss entirely.
The surrogate is positioned as a tool for preliminary and repeated studies, not a replacement for high-fidelity simulation at final design sign-off. As the method matures and training datasets grow, its role in early-stage wind farm development could expand considerably.
More information is available here: Gasser, J., Frischholz, Y., and Lehning, M.: Physics-regularised neural network surrogates for fine-scale wind and pressure fields over complex terrain and structures, Wind Energ. Sci. Discuss. [preprint], https://doi.org/10.5194/wes-2026-179, in review, 2026.
Daniel García is an Editor-in-Chief with strong expertise in structural work and engineering principles. He combines this technical foundation with deep knowledge of energy, spatial design, and emerging technologies, bringing a forward-thinking and analytical approach to editorial leadership.
