Reconstructing Boundary Layer Wind Profiles with Machine Learning: A Comparative Study on IGRA Radiosonde Data

AI summary

Overview

Research area: Machine learning for atmospheric science — specifically, data-driven reconstruction of atmospheric boundary layer (ABL) wind profiles from radiosonde observations, benchmarked against classical analytical parameterizations.

Technical level: Intermediate. The paper combines meteorological concepts (geostrophic balance, Ekman layer, Monin–Obukhov similarity) with standard tabular machine learning and neural operator methods. Readers with a background in either ML or atmospheric science can follow it; the physics section assumes some familiarity with boundary-layer theory.

Scope in one sentence: A benchmark study on 158,097 global IGRA radiosonde soundings that separates how much accuracy comes from the choice of predictors versus the choice of model architecture, estimates an empirical error floor for point-wise reconstruction, and tests a Fourier Neural Operator as a post-hoc corrector of vertical structure.

What This Paper Is About

Wind profiles in the atmospheric boundary layer matter for short-range forecasting, pollution transport, and wind-energy siting, but radiosondes are sparse in space and time, so practitioners must infer full vertical profiles from a small handful of physical predictors. Classical approaches — the power-law profile and Monin–Obukhov similarity theory — capture only first-order effects of roughness and stability and consistently underperform against actual observations. The paper asks whether modern machine learning can do better, and more importantly, whether any improvement comes from bigger models or from better input features.

Key Contributions

1. A predictor-versus-architecture decomposition on a global benchmark. On 158,097 IGRA soundings, the authors evaluate FT-Transformer, CatBoost, and TabM under two predictor sets (a "canonical" set and an "extended" set). Switching the predictor set moves RMSE from roughly 2.5 m/s to roughly 1.3–1.6 m/s, while switching the architecture moves it by less than 0.05 m/s. A leave-one-block-out ablation localizes the gain to thermodynamic predictors (temperature, pressure) and diurnal/seasonal categorical variables.

2. An empirical floor for point-wise reconstruction. A cohort analysis of within-bin profile dispersion yields an estimated lower bound of approximately 1.77 m/s for any model using only standard radiosonde features. This is independently consistent with the roughly 2 m/s residual reported by prior work that inverted eddy-viscosity coefficients from radiosonde data.

3. A 1D Fourier Neural Operator used as a post-hoc corrector. Applied on top of frozen tabular predictions for U.S. stations under the extended configuration, the corrector reduces near-surface RMSE from 1.5 m/s to 1.2 m/s while leaving upper-layer error essentially unchanged — evidence that there is coherent vertical structure the point-wise tabular models do not exploit.

4. Cross-region transfer diagnostics. Round-robin cross-validation (train on one region, test on all others) and a four-way partition of U.S. stations show models retain meaningful transferability across regions, particularly under the richer predictor set.

Main Findings

• ML beats analytical fits by roughly a factor of two. Power-law RMSE is 3.02 m/s and Monin–Obukhov is 3.21 m/s, against 1.3–1.6 m/s for the best ML models.
• Architecture barely matters. Under a fixed predictor set, CatBoost, TabM, and FT-Transformer differ by less than 0.05 m/s — smaller than the seed-to-seed spread reported in Table II (standard deviations of 0.02–0.05 m/s).
• Predictors dominate. Moving from Configuration 1 (geostrophic wind components, latitude, boundary-layer height) to Configuration 2 (adding longitude, surface altitude, surface/ABL-top temperature and pressure, season, part of day) drops RMSE by roughly 1 m/s: TabM goes from 2.50 to 1.31 m/s locally and from 2.60 to 1.58 m/s globally.
• Local models edge out global ones. Per-country models reach about 1.3 m/s; the global model reaches about 1.6 m/s.
• The ablation names the load-bearing predictors. Removing season and part-of-day categoricals costs +0.25 m/s (1.58 to 1.83); removing the thermodynamic block (T and p at surface and ABL top) costs +0.14 m/s (1.58 to 1.72); removing longitude and surface altitude costs only +0.03 m/s (1.58 to 1.61). Dropping back to Configuration 1 costs +1.02 m/s (1.58 to 2.60).
• There is a measurable information floor. Within-cohort dispersion falls from 3.40 m/s with geostrophic wind alone, to 3.20 with latitude, 2.70 with season, 1.98 with part of day, and 1.77 m/s once boundary-layer height is included. Reported model RMSE sits within 0.2 m/s of this 1.77 m/s figure.
• Error is largest near the surface. RMSE profiles decrease with altitude, consistent with tighter coupling to geostrophic flow aloft and unresolved turbulence, stratification, and surface effects near the ground.
• The FNO corrector helps only near the surface. The roughly 0.3 m/s gain is confined to lower layers; aloft, the tabular models already receive the geostrophic wind as an input, leaving little residual structure to recover.
• Cross-region degradation is modest under the extended predictor set. For Configuration 2, off-diagonal errors exceed diagonal values by less than 1 m/s; for Configuration 1, degradation reaches up to 1.5 m/s.
• Reported global numbers are likely optimistic. The 80/20 split is over individual soundings rather than stations, so spatially close profiles can appear in both partitions.

Methodology in Plain English

The authors start from a public archive of weather-balloon measurements (IGRA, the Integrated Global Radiosonde Archive) and clean it aggressively. Each sounding's boundary-layer height is diagnosed from where the potential temperature profile first matches the surface virtual potential temperature. Profiles are then resampled onto a fixed grid of 50 vertical levels using a spline in normalized height, and soundings that are too short (fewer than 20 measurements), too thin (boundary layer under 100 m), too flat (velocity variation under 2 m/s), or too fast (above 50 m/s) are removed. The surviving dataset has 158,097 soundings.

The prediction task is stated simply: given a compact set of inputs — geostrophic wind at the boundary-layer top, latitude, boundary-layer height, stability-related temperature and pressure differences, and categorical season/time-of-day tags — output the full vertical profile of zonal and meridional wind. Two input sets are defined, one minimal and physically motivated, one extended with geographic and thermodynamic context. Three model families are trained: CatBoost (gradient-boosted trees), TabM (a parameter-efficient neural ensemble), and FT-Transformer (a transformer adapted to tabular data). All are trained with the same unweighted level-wise mean squared error so the comparison with analytical fits stays fair, on both per-country and global data, with a 10% inner validation slice for hyperparameter selection and five random seeds to gauge variance.

Two additional experiments probe the limits. First, a cohort analysis bins profiles by shared predictor values and measures how much wind speed still varies within each bin — a direct empirical estimate of irreducible uncertainty. Second, a one-dimensional Fourier Neural Operator is trained as a corrector sitting on top of frozen tabular predictions, exploiting the fact that tabular regressors treat each height level independently and therefore cannot smooth or propagate structure vertically.

Why This Matters

The paper's central message is a resource-allocation argument for the field: on this task, additional model capacity buys essentially nothing, while additional physical predictors buy roughly a full meter per second of accuracy. That reframes where effort should go in ML-for-atmospheric-science work — toward feature engineering, data assimilation, and additional observational context rather than toward larger networks.

Real-world applications:

• Short-range weather forecasting: boundary-layer winds are a known weak point in near-surface forecasts; a 1.3–1.6 m/s reconstruction error benchmark gives a concrete target for learned downscaling or post-processing modules.
• Wind-energy siting and resource assessment: turbine hub heights sit inside the boundary layer, and the near-surface error profile the paper reports is exactly where siting uncertainty is largest.
• Pollutant and tracer transport: dispersion models need vertical wind structure, not just surface wind, and radiosonde sparsity is a persistent constraint.
• Aviation and surface transport operations: low-level wind profiles affect takeoff/landing planning and can be relevant to wind hazard warnings.

Industry relevance: the paper is directly relevant to wind-energy developers and operators, to numerical weather prediction groups considering learned boundary-layer parameterizations (the authors note recent work on neural-network emulation of schemes like YSU inside NWP models), and to teams building data-driven downscaling products. The explicit impact statement warns that global model numbers are systematically optimistic for stations underrepresented in training — primarily tropical, maritime, and Southern-Hemisphere regimes, which overlap with regions where wind capacity is projected to grow fastest — so deployment there requires station-blocked retraining rather than direct transfer.

Future Directions

1. Test whether the 1.77 m/s floor is fundamental or feature-bound. The authors propose adding predictors that exist in NWP pipelines but not in radiosonde reports — surface heat flux, terrain roughness, or a proxy for the Obukhov length — as a falsifiable test.
2. Ablate the FNO corrector against matched-compute alternatives. The current setup compares TabM to TabM+FNO but does not isolate operator-specific spectral mixing from generic vertical smoothing; a 1D convolution or small MLP on the same tiled tensor might recover a similar fraction of the 0.3 m/s gain.
3. Move from single-snapshot regression toward learned data assimilation. Providing neighbouring stations, recent reanalysis fields, or temporal context could close part of the residual currently attributed to intrinsic dispersion.
4. Retrain under a fully station-disjoint protocol and add comparisons against operational parameterizations (e.g., YSU within WRF) and against reanalysis-derived ABL winds, since the current "ML beats analytical" claim is only against simple closed-form fits.

Target Audience

Atmospheric scientists and meteorologists interested in what machine learning can and cannot add to boundary-layer modeling; ML researchers working on tabular regression or scientific machine learning who want a domain-specific benchmark where architecture is demonstrably not the bottleneck; wind-energy analysts and NWP developers evaluating learned post-processing for near-surface winds; and graduate students looking for a well-scoped example of how to design a predictor-versus-architecture ablation with an explicit uncertainty floor.

Authors’ abstract

Wind profiles in the atmospheric boundary layer (ABL) drive short-range forecasting and wind-energy siting, yet sparse radiosonde coverage forces practitioners to interpolate them from a handful of physical predictors. We benchmark modern tabular learners against analytical parameterizations on 158,097 IGRA soundings. CatBoost, TabM, and FT-Transformer reduce RMSE to 1.3-1.6 m/s, against approximately 3 m/s for power-law and Monin-Obukhov fits; once the predictor set is…