Geometry, not simulation volume, carries the signal
FLUME's premise is that under a prescribed inflow and thermal forcing, most of the spatial variation in time-averaged wind and temperature is controlled by the surrounding building geometry. If that holds, a model does not need to memorize thousands of full CFD fields; it needs a compact description of the open space around each point and enough examples to learn how that space shapes flow.
The framework is built on a Fourier neural operator, which learns mappings between fields rather than solving the governing equations step by step. Instead of learning one mapping from a whole city to a whole 3D field, FLUME cuts simulations into smaller 3D patches and learns locally. That decomposition is what makes 23 CFD cases sufficient: each case contributes many patches, and each patch pairs a local geometry description with its CFD-derived velocity and temperature values.
The geometry description is the Multi-Directional Distance Feature (MDDF). For each grid point it records how far unobstructed space extends in different directions before a straight ray meets a building surface. The paper calls this line-of-sight free space. Airflow does not actually travel in straight rays, but the authors argue this visibility-based measure captures the local blockage, channeling, and wake formation that a fixed-size patch would miss, because open plazas and narrow gaps have no common fixed scale.
Twenty-three blocks of CFD become the training set
The ground truth comes from CityFFD, a GPU-based large-eddy simulation solver, run on 33 urban configurations drawn from real cities including London, Montreal, Tokyo, Singapore, and Hong Kong. Cases span 1.2 km × 1.2 km domains, with two larger 3 km × 3 km configurations, at a near-building mesh resolution of 4 m × 4 m × 1.5 m. Each 1.2 km case used one GPU for roughly eight hours on a domain of 14 million cells.
The forcing is fixed across the dataset: an open-terrain inflow profile with U = 4 m/s at 10 m height, building surfaces held at 40 °C, inlet air at 30 °C, and no radiation modelling. Time-averaged fields were extracted after one hour of spin-up from a 2.7-hour simulation. Of the 33 cases, only 23 were used for training; the rest supported testing and validation, including the two unseen configurations reported in the results.
This design choice matters for reading the accuracy numbers. The reference is a simulation, not a field measurement, so the reported errors measure agreement with CityFFD under one inflow and thermal setting, not agreement with observed air temperature or wind at a real site.
Streamwise velocity and temperature transfer best
On the unseen 1.2 km downtown Montreal configuration, FLUME reached R² of 0.97 for both streamwise velocity U and temperature T, with FAC2 values of 0.98 and 1.00. FAC2 is the fraction of predictions falling within a factor of two of the reference, or within a small absolute band near zero; it answers a different question from R², which describes how much of the spatial variance is reproduced. Both point the same way here.
The secondary velocity components were harder. R² was 0.86 for lateral velocity V and 0.83 for vertical velocity W, and all mean absolute errors stayed below the paper's practical thresholds of 0.5 m/s for velocity components and 1 °C for temperature. The paper treats this ordering, best for U and T and weaker for V and W, as physically consistent with the sensitivity of turning, separation, and recirculation to fine geometric detail.
The larger unseen case, a 3 km × 3 km Osaka domain, held up with only modest degradation: R² of 0.95 for U and 0.98 for T, 0.78 for V and 0.70 for W. Mean absolute errors were 0.31 m/s for U, 0.05 m/s for V, 0.09 m/s for W, and 0.18 °C for T, with FAC2 of 0.99, 0.86, 0.79, and 1.00 respectively. The abstract's headline figures of 0.2 m/s and 0.19 °C sit at the favourable end of this range; the 0.31 m/s streamwise error on the larger domain is the number to keep in view when judging scale-up.
Rooftop shear layers and wakes blur first
Spatial comparisons show the model preserves the large-scale organization: high-velocity corridors in open regions, sheltered low-speed zones inside the dense core, and coherent wake bands extending across the domain. Temperature agreement is the strongest of the four variables, with the warm urban core and cooler periphery reproduced closely and residuals concentrated near the domain edge and in poorly ventilated pockets.
The losses appear at sharp transitions. FLUME smooths some of the narrow gradients near rooftop level and immediately behind taller obstacles, where separation and reattachment create abrupt local changes. Vertical velocity is the most challenging variable, with wider scatter in regions of rooftop-induced vertical exchange and recirculation loops. The paper attributes this to the patch-based formulation, which reconstructs abrupt local features less exactly than broad ones.
The 3 km case also tested a data-composition question. Only one training case, Case 31, came from a comparably large domain; the rest were smaller 1.2 km cases. The authors trained a separate model without Case 31 to probe how much that single large-area example contributed, which makes the scale-up result conditional on having at least one training domain of comparable size rather than on patch decomposition alone.
Ensemble spread marks the hard regions
A deep ensemble supplies a practical proxy for epistemic uncertainty, and conformal prediction converts the ensemble spread into calibrated prediction intervals. The useful property is not just calibration but spatial structure: spread is positively related to local prediction error across all variables, so wider intervals tend to appear where the model is actually wrong more often.
The widening follows aerodynamic complexity. Intervals are narrowest in the incoming flow with limited geometric interference, and broaden in building wakes, channeling passages between buildings, roof-level shear layers, and for vertical velocity near the ground. Temperature intervals stay narrow above rooftop level and widen mainly near the surface where heated façades and sheltering shape the local thermal field.
Calibration was checked against nominal coverage: at the 70% level, observed coverage was 66.6–75.8%, and at the 30% level it was 28.8–32.3% for case 0. Notably, calibration used entirely separate cities (cases 8, 20, 24), which supports transferring the uncertainty calibration across geometries. The intervals are not a guarantee everywhere: at two channeling and wake locations, the CityFFD profiles fell outside the intervals over parts of the height for U, V, and W.
Seconds replace hours under fixed forcing
After training, FLUME produces a full 3D prediction in about 10 seconds, and total prediction time stays under one minute when MDDF computation is included. The comparison point in the paper is CFD: tens to hundreds of CPU or GPU hours for a building-resolving case of practical scale, or roughly eight hours per case even with the GPU-based CityFFD solver at the scales studied here.
That speed is what makes rapid scenario screening plausible, but the scope conditions are part of the result. All training and evaluation used a single inflow speed and direction and fixed thermal forcing, so the rotation and rescaling procedures described in the methods remain unvalidated. Generalization across inflow speed and direction, atmospheric stability, thermal boundary conditions, and mesoscale states is explicitly outside the paper's scope.
Resolution is a second boundary. The 4 m × 4 m × 1.5 m near-building mesh resolves dominant canopy flow but not the smaller turbulent structures relevant to pedestrian comfort in narrow streets or detailed flow around façade elements. The authors list coupling with higher-resolution CFD and testing how accuracy scales with grid refinement as future work.
The threshold to watch is 0.5 m/s and 1 °C
The paper judges practical acceptability against thresholds of 0.5 m/s for velocity components and 1 °C for temperature, drawn from urban microclimate and wind engineering literature for pedestrian wind assessment, thermal comfort, and urban energy modelling. Every reported mean absolute error sits below those thresholds, including the 0.31 m/s streamwise error on the larger unseen domain, which is the tightest margin among the headline numbers.
The more informative split is by variable and by region rather than by city. Streamwise velocity and temperature are the reliable outputs; lateral and vertical velocity are the ones to treat cautiously, especially in wakes, channeling passages, and rooftop shear layers, where the ensemble intervals widen and where reference profiles occasionally fall outside them. A user screening building layouts can lean on U and T fields and should read V and W as indicative, with the conformal intervals as the local reliability signal.
The paper's own framing is that FLUME is a first step toward a geometry-aware, uncertainty-aware emulator, not a validated replacement for CFD across conditions. Its distinctive contribution is the combination: patch-based learning with a visibility-defined geometry feature that makes 23 CFD cases enough, plus an uncertainty layer whose spread tracks the aerodynamic difficulty of each location. The practical priority that follows is to use the calibrated intervals, not the global error averages, to decide where a predicted field is trustworthy enough for a given design question.
Sources used in this answer
FLUME-FNO: data-efficient and scalable prediction of 3D wind and temperature fields in unseen urban morphologies
FLUME, a Fourier neural operator trained on CFD simulations of 23 urban blocks with a Multi-Directional Distance Feature geometry encoding, predicts 3D mean wind and temperature fields in unseen urban morphologies with mean absolute errors of 0.20–0.31 m/s for streamwise velocity and 0.18–0.19 °C for temperature, producing a full 3D prediction in about 10 seconds after training.