8-Hour CFD Cut to 3 Seconds: Neural Operator Wind Fields Save UAV Energy up to 12.7% in Headwind

GeoWind2Plan: Mission-Time 3D Urban Wind Prediction for Energy-Efficient UAV Planning

Shaoxiang Qin, Yucheng Zhao, Fuyuan Lyu, Di Zhou, Jiachen Yao, Xue Liu, Anima Anandkumar, Liangzhu Leon Wang, Xiongye Xiao

NeurIPS 2026 (Spotlight)

cs.AI, cs.RO

2026-09-29

A geometry-conditioned neural operator predicts corridor-local 3D urban wind in about 3 seconds versus 8 hours for CFD, cutting UAV headwind energy 12.7% versus wind-agnostic planning and recovering 86% of CFD-reference savings.

What problem this solves

Urban UAV energy depends on more than path length. Buildings reshape ambient wind into spatially varying 3D flow: speed-up through street canyons, slow wakes behind towers, redirected flow around dense clusters. Propulsion energy tracks air-relative velocity, so the geometrically shortest route can be the costliest one. Wind-aware planning therefore needs building-resolved 3D wind, and CFD is the natural source, yet one simulation solves one city under one inflow boundary condition. Change the wind direction or speed and you rerun it; a high-resolution 1.2 km urban case takes about 8 hours, while urban missions last minutes to tens of minutes. Prior energy-aware planners mostly assumed a wind field was already on hand, or used a single wind vector or a height-only profile, discarding the horizontal structure that drives energy cost.

Method

GeoWind2Plan takes three inputs: a background wind vector, 3D building geometry, and a start-goal pair.

• Direction adaptation is a coordinate rotation. The paper proves a yaw-covariance property: rotating the city into a reference-wind frame, predicting under a fixed reference inflow, and rotating the vector field back is equivalent to predicting for the new direction. The model learns a geometry-to-wind mapping, and direction variation is absorbed into the rotation, so no per-direction model is needed.
• Speed adaptation is a rescale. High-Reynolds-number urban mean flows are approximately self-similar: the normalized time-averaged field varies weakly with inflow speed, so the prediction is multiplied by U/Uref. Measured normalized vector differences across CFD fields at 2, 4, and 8 m/s stay between 6.7% and 11.2%.
• Corridor-localized inference. Full-city fields hold millions of grid cells, while a mission needs only a corridor around the start-goal line. Predicting just the intersecting patches and stitching them into a queryable field cuts inference from 15 s to 3 s at 20% corridor width.
• Planning runs 10 RRT warm starts, each refined by nonlinear trajectory optimization that jointly optimizes the 3D path and speed profile under obstacle, thrust, and acceleration constraints, about 2 s per start on a CPU core. The energy model is a physics-derived multirotor model that explicitly accounts for air-relative speed, drag, thrust, induced power, and blade profile power.

The wind predictor is a localized FNO (a neural operator that learns input-function-to-PDE-solution maps) conditioned on occupancy, SDF, and multi-directional distance features, applied to shared 45×45×45 patches. It trains on roughly twenty 1.2 km CFD cases in about 7 hours on a single V100.

The design goal is decision-useful wind, not pointwise CFD reconstruction: trajectories are planned on predicted wind and evaluated under high-fidelity CFD wind.

Results

Evaluation covers four held-out 1.2 km blocks and one 3 km block, 256 missions each, default inflow 4 m/s. Baselines are zero wind and a height-only profile; the CFD field serves as the oracle (planning directly on the true field). Energy is reported in Wh/km; on block A across all wind angles, the oracle sits at 12.3, wind-agnostic at 13.3, and this method at 12.5.

Averaged over the four 1.2 km blocks, savings over wind-agnostic planning reach 6.9% in tailwind, 12.7% in headwind, and 4.5% in crosswind, recovering 87.9%, 85.7%, and 75.0% of the oracle's savings. Stronger wind raises the value of local 3D structure. Trajectories adapt systematically: higher and faster in tailwind, lower and slower in headwind, with lateral bending that matches the CFD reference. Narrowing the corridor to 10% barely changes energy, but planning fails when a building taller than 120 m spans the corridor and blocks all feasible paths.

Why it matters

This is the first framework coupling wind-conditioned, building-resolved 3D wind prediction with CFD-evaluated UAV energy planning at mission time. For urban logistics or inspection fleets, it removes the need for a precomputed city-wide wind library: a 3-second wind field recovers 75% to 88% of what oracle planning would save, as a pure software gain with no hardware changes. The more transferable idea is the decision-level evaluation protocol: a surrogate's worth is measured by downstream trajectory energy under the true field, not pointwise error.

The absolute savings of 5% to 13% are moderate, but they cost nothing in hardware and still matter for range-constrained vehicles.

Limitations

Stated by the authors: the model uses time-averaged fields, so turbulence-induced power fluctuations and turbulence safety risks are out of scope, and background wind is assumed fixed during a mission, with no replanning under time-varying inflow. The speed-rescaling error grows to 11.2% at 8 m/s, and thermally stratified or gusty inflow is untested.

Reading closely: training uses about twenty CFD cases, which the authors call the main bottleneck, and generalization to irregular real-city geometry is unverified. All ground truth comes from the CityFFD solver, evaluation reuses the same energy model, and no real flight data backs the numbers. The oracle is a same-optimizer CFD-reference planner, not a certified global optimum.

Terms

• FNO (Fourier Neural Operator): an operator-learning network that mixes spatial information in the Fourier domain, often used to approximate PDE solution maps.
• SDF (signed distance function): signed distance from any point to the nearest building surface.
• MDDF (multi-directional distance feature): distance to obstacles along a set of rays, encoding local blockage and channeling.
• RRT (rapidly-exploring random tree): a sampling-based path planner used here to generate feasible initial routes.
• Yaw covariance: rotating the city geometry and inflow about the vertical axis rotates the flow solution accordingly.

Source

• arXiv 2609.36056v1
• PDF
• Hugging Face

What people are saying

• GeoWind2Plan: seconds-scale 3D urban wind prediction for energy-efficient drones AnimaAnandkumar

Related papers

• ToM2Vec adds 12 Overcooked points by aligning 10 latent tokens to a partner
• T2Quant reallocates a fixed 4-bit DLM budget over denoising steps, gaining up to 6.7 points
• MongoDB and Trane propose contestable governance: agents argue, a steward contests before commit
• Same contest problem, different MoEs: expert overlap ~60%, same as a new question
• Deterministic preferences nullify DPO's KL term; IPO pins the log-ratio gap
• SCOPE compiles policies into programs weak models run, matching frontier judges at 9-47x less cost

All paper explainers