The index is a stand-in for a physiological model
The Universal Thermal Climate Index (UTCI) converts four weather inputs — air temperature, mean radiant temperature, wind speed and relative humidity — into an equivalent air temperature that describes how a person would feel. The full calculation runs the Fiala multi-node model of human heat transfer and thermoregulation coupled to a clothing model; the paper calls this combination Fiala-UTCI. It is computationally intensive and proprietary paid software, so in practice almost nobody runs it directly.
Instead, operational systems use a sixth-order polynomial approximation fitted to Fiala-UTCI outputs. The paper calls this Polynomial-UTCI. It sits inside common tools such as RayMan, pythermalcomfort, thermofeel and the official UTCI FORTRAN calculator, and it underpins the ERA5-HEAT global hourly product at 0.25° resolution. Because the approximation is so widely embedded, the literature often treats it as if it were the full model. The paper's central claim is that this conflation hides a consequential error.
The approximation, not the forecast, drives much of the error
On the Fiala-UTCI testing data, Neural-UTCI reaches an RMSE of 0.36°C (interquartile range 0.35–0.38°C) against 2.78°C (IQR 2.53–2.99°C) for Polynomial-UTCI, an 87% reduction. The distribution matters as much as the average: every Neural-UTCI error falls within ±2°C, while 16% of Polynomial-UTCI cases exceed that bound; 90% of Neural-UTCI errors sit within ±0.5°C versus 38% for the polynomial.
The gap widens where the polynomial is known to fail. Above 17 m/s wind speed, Polynomial-UTCI has an RMSE of 9.03°C with individual errors above 20°C, and the index is conventionally declared out-of-domain there. Neural-UTCI brings that high-wind RMSE down to 0.89°C, which is why the authors describe continuous global application rather than a range-restricted one.
The paper also reframes total operational error. Against the Fiala-UTCI model's own reported 1.04°C RMSE, Neural-UTCI's 0.36°C approximation error contributes about 19% of total operational UTCI error, whereas Polynomial-UTCI increases expected total error by 132%. In other words, the approximation stops being the dominant error source once the network replaces the polynomial.
Category boundaries turn small errors into different warnings
UTCI is used categorically: fixed thresholds sort the index into bands from extreme cold stress through no thermal stress to extreme heat stress, and those bands feed weather warnings and exposure studies. Because classification is threshold-based, an error only has to cross a boundary to change the label, so average accuracy understates the practical stakes near band edges.
Measured against Fiala-UTCI categories, Neural-UTCI misclassifies 1.7% of cases (IQR 1.4–2.0%) and Polynomial-UTCI 5.3% (IQR 4.8–5.7%), a 68% reduction. Neural-UTCI has lower absolute error in 85% of test cases and reduces error by more than 1°C in 35% of them. The largest mean improvement is in extreme cold stress, at 2.4°C, and the smallest is in the no thermal stress band, at 0.2°C. Among heat categories, very strong heat stress gains most, though its mean improvement is only 0.4°C.
The heat-stress numbers need a caveat the paper states plainly: the testing dataset contains no extreme heat stress observations and only 15 very strong heat stress cases. The stronger evidence for heat comes from the training data, where Polynomial-UTCI misclassifies 21% of extreme heat stress and 14% of very strong heat stress observations, against 0.6% and 2% for Neural-UTCI.
The network reproduces humidity and wind effects the polynomial misses
Accuracy alone does not show that a model behaves like the physiology it replaces, so the authors used SHAP decompositions to compare how each approximator responds to individual inputs. Both rank air temperature as the most important predictor of the UTCI offset, but they diverge afterwards.
Relative humidity is the least important predictor in Polynomial-UTCI. At high temperatures the polynomial barely varies with humidity, and at low temperatures high humidity lowers its output. Neural-UTCI ranks humidity second after air temperature: values below the 50% reference condition reduce the index and values above it raise it, which the authors describe as matching the Fiala model's heat–humidity relationship. That relationship matters because high humidity limits evaporative cooling from sweating.
Wind shows a similar pattern. Both models lower UTCI as wind speed rises, but Neural-UTCI makes the effect strongly dependent on air temperature — a marginal contribution at high temperatures and a much larger one in the cold — while Polynomial-UTCI diverges from the expected relationship around 20 m/s, which the authors link to its large high-wind errors. The SHAP comparison is about the direction and shape of responses; the paper notes magnitudes are not comparable between the two models.
Event-level heat stress counts shift substantially
The paper translates approximation differences into exposure metrics for two European events. During the June–August 2003 European heatwave in Rome, the number of very strong heat stress days rises from 15 to 35 when Neural-UTCI is used instead of an operational product such as ERA5-HEAT. The authors also examine the 2012 European cold spell, and report that Neural-UTCI classifies extreme cold stress reliably enough to support continuous global use.
These are comparisons of classification outputs from two approximators applied to the same event, not forecasts of health outcomes. Their value is diagnostic: they show that the choice of approximation can change how many days cross a warning threshold, which is the quantity that feeds heat-health alerts. The paper further reports that Polynomial-UTCI introduces systematic and spatially structured errors in global 2024 UTCI estimates relative to Neural-UTCI, altering global thermal stress exposure estimates.
Speed is comparable on CPUs and much better on GPUs
A replacement is only useful operationally if it fits existing workflows. For individual and batched inputs on a CPU, the two methods show a negligible speed difference. On a GPU, Neural-UTCI computes batched inputs 33 times faster than the CPU-based polynomial implementation, depending on batch size and GPU specifications. The network is released as an open-source Python package and is positioned as a drop-in replacement for Polynomial-UTCI.
The input domain is also wider in one respect: Neural-UTCI is validated for wind speeds from 0.5 to 30.3 m/s, whereas Polynomial-UTCI's valid range stops at 17 m/s. Both share the same limits on air temperature (−50°C to +50°C), mean radiant temperature (−80°C to +120°C) and relative humidity (5% to 100%), inherited from the training data.
The ceiling is the physiological model and its training grid
Neural-UTCI can only be as accurate as Fiala-UTCI itself. That model represents an average individual of 73.4 kg and 14% body fat, and it does not account for age, health conditions, activity level, acclimatization, cultural clothing differences or the lagged effects of thermal conditions over preceding days. Local adjustments to UTCI have therefore been proposed. A more faithful approximation of a model with these blind spots does not remove them.
The training data impose a second ceiling. The grid used to fit both approximators is synthetic, so it does not reproduce the correlations between inputs found in real atmospheric samples — for example, the roughly −0.4 correlation between wind speed and vapour pressure, mean radiant temperature and air temperature in an Earth-system model sample. Both approximators perform worst at high wind speeds and high relative humidity, where the training grid is sparsest. The test data also come from 1971–80 control simulations, which is why they contain no extreme heat stress cases; the authors argue that updated reference simulations spanning present-day and emerging extremes would be needed to know where the models fail.
Accuracy at the extremes is the priority, not average fit
The paper's argument is not that the polynomial is unusable but that its error is concentrated where warnings matter most. The largest categorical gains appear in extreme cold stress and in the high-wind range the polynomial is officially out-of-domain for, and the Rome comparison shows a threshold-crossing count more than doubling for very strong heat stress days. A fixed misclassification rate also translates into a growing absolute number of missed dangerous days as heat-stress days become more frequent in Europe, which is the authors' stated reason for treating this as a timely methodological fix.
For a reader deciding whether to switch, the relevant split is therefore by condition rather than by average performance: the case for Neural-UTCI is strongest at high wind speeds, in cold extremes and near category boundaries, and weakest where the underlying Fiala model or the synthetic training grid is itself least trustworthy — high humidity combined with high wind, and heat extremes absent from the test set. The 0.36°C versus 2.78°C gap and the 1.7% versus 5.3% misclassification gap describe approximation fidelity to Fiala-UTCI, not fidelity to observed human health outcomes.
Sources used in this answer
A neural network-based Universal Thermal Climate Index for reliable global thermal-stress classification across extreme weather
A compact neural network approximating the Fiala-UTCI model reduces approximation RMSE from 2.78°C to 0.36°C and thermal-stress misclassification from 5.3% to 1.7% relative to the operational polynomial, while extending valid wind speeds to 30.3 m/s.