PredHydro-Net: 3D Hydrometeor Forecasting

• The paper introduces a physics-guided dual-decoding approach that decouples thermodynamic and hydrometeor predictions to better preserve extreme event signals.
• It integrates wavelet decomposition, spectral amplitude matching, and PatchGAN adversarial training to counter over-smoothing and maintain high-frequency texture fidelity.
• Evaluations on a 72-h global ERA5 benchmark show improved extreme-event detection and climatological consistency compared to Earthformer, PredRNNv2, and operational systems.

Searching arXiv for the cited paper and related baselines to ensure fresh, correct citations. PredHydro-Net is a physics-guided dual-decoding framework for global three-dimensional hydrometeor prediction, introduced to address a persistent failure mode of standard deep learning optimization on cloud and precipitation variables: overly smooth forecasts that attenuate extreme events and spatial textures. The method is formulated for hydrometeor fields that are both strongly zero-inflated and long-tailed, and it combines a decoupled architecture, unidirectional physical modulation, wavelet-based multiscale supervision, spectral amplitude matching, and PatchGAN-based adversarial texture generation within a single training objective. In a 72-h global evaluation, it is reported to outperform Earthformer, PredRNNv2, and the operational Global Forecast System in extreme-event detection and spectral representation, while also showing strong climatological consistency with Global Precipitation Measurement satellite retrievals and reasonable reproduction of three-dimensional cloud structures in events such as Hurricane Ian (Chen et al., 7 Jun 2026).

1. Problem formulation and modeling target

PredHydro-Net is designed for forecasting three-dimensional hydrometeor fields, specifically cloud ice, liquid water, rain, and snow. The motivating observation is that these variables are strongly zero-inflated and long-tailed. Under a conventional end-to-end shared decoder trained with standard regression losses, the predicted fields tend to “average out” localized extremes and high-frequency structure, producing smooth outputs that may reduce pointwise error while degrading event realism (Chen et al., 7 Jun 2026).

The paper situates this problem within a broader contrast between continuous atmospheric variables and hydrometeors. While global data-driven models are described as excelling at predicting continuous atmospheric variables, hydrometeor forecasting remains challenging because the target distribution is not well matched to optimization schemes that favor mean behavior. In this formulation, the key technical objective is not only low global error, but a favorable trade-off between quantitative accuracy and spatial fidelity. This includes preservation of extreme-event signals, neighborhood-scale structure, and high-wavenumber spectral energy.

A central conceptual distinction in the method is between macroscopic thermodynamic and dynamic fields on the one hand, and hydrometeor generation on the other. PredHydro-Net treats these as related but non-identical prediction tasks. This suggests that the model’s architecture is intended to reduce optimization conflict among heterogeneous output variables rather than forcing all variables through a single decoder pathway.

2. Dual-decoder architecture and physics-guided coupling

The core architecture adopts a PredRNNv2-style spatiotemporal encoder consisting of a four-layer ST-LSTM stack. It ingests two past 6-hourly states,

$X\in\mathbb{R}^{2\times 93\times180\times360},$

and produces a latent state

$H_{\mathrm{lat}\in\mathbb{R}^{48\times114}.$

From this shared latent representation, two decoders unfold in parallel. The thermodynamic decoder reconstructs ten continuous fields, including temperature $T$, specific humidity $q$, winds, and geopotential, at five pressure levels. The hydrometeor decoder generates twenty three-dimensional hydrometeor contents—CIWC, CLWC, CRWC, and CSWC—at the same five levels (Chen et al., 7 Jun 2026).

The critical coupling between the branches is implemented through a unidirectional Feature-wise Linear Modulation layer called TQ2HydroFiLM. At a given decoder stage, feature maps from the thermodynamic branch, denoted $F_M\in\mathbb{R}^{C\times h\times w}$, are used to compute scale and shift parameters,

$$\Gamma \;=\;\sigma\bigl(W_\Gamma\,F_M\bigr), \quad B\;=\;W_B\,F_M, \tag{1}$$

where $W_\Gamma$ and $W_B$ are learned $1\times1$ convolutions and $\sigma$ is the sigmoid. These parameters modulate the hydrometeor features:

$H_{\mathrm{lat}\in\mathbb{R}^{48\times114}.$0

A stop-gradient on $H_{\mathrm{lat}\in\mathbb{R}^{48\times114}.$1 makes the interaction explicitly one-way: the macroscopic branch constrains the hydrometeor branch without reciprocal interference. In the terminology of the paper, this is the main mechanism by which the model is “physics-guided.” It does not mean that the forecast is generated by an explicit physical solver; rather, predicted macroscopic fields such as temperature and humidity are used to modulate hydrometeor generation. A common misunderstanding would be to treat PredHydro-Net as a single-decoder system with auxiliary losses. It is instead defined by architectural decoupling plus directed conditioning.

3. Wavelet decomposition, spectral supervision, and texture realism

PredHydro-Net uses three complementary mechanisms to counter over-smoothing: wavelet-based frequency decoupling, spectral amplitude matching, and adversarial training. Together, these are intended to separate large-scale envelopes from fine-scale textures and to constrain both the spatial-domain and frequency-domain character of the prediction (Chen et al., 7 Jun 2026).

For wavelet supervision, the model applies a three-dimensional discrete wavelet transform using the Haar mother wavelet to a hydrometeor field $H_{\mathrm{lat}\in\mathbb{R}^{48\times114}.$2:

$H_{\mathrm{lat}\in\mathbb{R}^{48\times114}.$3

where $H_{\mathrm{lat}\in\mathbb{R}^{48\times114}.$4 are the Haar wavelet filters corresponding to LL, LH, HL, and HH subbands. At the coarsest scale, the LL subband captures the low-frequency envelope, while LH, HL, and HH capture orthogonal high-frequency details. Each subband is supervised separately. The LL component receives an asymmetric $H_{\mathrm{lat}\in\mathbb{R}^{48\times114}.$5 penalty that penalizes under-prediction of precipitation more heavily, with the example given as twice the weight for misses versus false alarms. The high-frequency subbands are regularized toward sparsity to prevent random noise. The intended effect is explicit reconstruction of both envelope and texture rather than implicit averaging under an MSE-only objective.

The method also imposes spectral amplitude matching in horizontal Fourier space. Defining the spectral amplitude at spatial frequency $H_{\mathrm{lat}\in\mathbb{R}^{48\times114}.$6 by

$H_{\mathrm{lat}\in\mathbb{R}^{48\times114}.$7

the spectral amplitude loss is written as

$H_{\mathrm{lat}\in\mathbb{R}^{48\times114}.$8

The paper notes that the implemented form is an $H_{\mathrm{lat}\in\mathbb{R}^{48\times114}.$9 distance on the log-amplitude, while the conceptual role is to align predicted and reference power spectra across all resolved scales.

To sharpen local texture further, PredHydro-Net incorporates a PatchGAN discriminator $T$0 that sees only the 20 hydrometeor channels. The generator $T$1 is the full PredHydro-Net pipeline, including encoding, ST-LSTM dynamics, dual decoding, FiLM modulation, and the wavelet head. The adversarial losses are given in standard minimax form:

$T$2

The implementation is described as using lightweight convolutional layers for both $T$3 and $T$4, so that adversarial feedback enforces local texture realism without overwhelming the physical-space regression. This is important for interpreting the role of GAN training here: it is a subordinate component in a composite objective, not the primary driver of forecast skill.

4. Loss design and optimization strategy

The full generator loss combines grouped regression, wavelet supervision, spectral alignment, and adversarial feedback. The grouped $T$5 regression term, denoted $T$6, is defined in standardized space using a Z-score transform for $T$7 and an $T$8-Z-score transform for hydrometeors, together with a physical-space “magnitude correction” $T$9 term in original units. This mixed normalization strategy reflects the heterogeneous statistical structure of the predicted variables (Chen et al., 7 Jun 2026).

The total objective is

$q$0

The reported weights are $q$1 and $q$2. The discriminator is trained to minimize $q$3 in alternation with the generator.

Several structural points follow directly from this formulation. First, the model does not abandon direct regression; it augments it. Second, the wavelet and spectral terms target distinct failure modes: the former decomposes scale-specific structure in physical space, while the latter constrains aggregate spectral energy across wavenumbers. Third, the low adversarial weight indicates that the GAN component is used to regularize texture rather than to define the prediction objective outright. A plausible implication is that the framework is designed to preserve optimization stability while still improving high-frequency fidelity.

5. Evaluation protocol and reported performance

PredHydro-Net is evaluated on a held-out year, 2022, using global $q$4-resolution ERA5 data for a 72-h forecast spanning 12 lead times at 6-h intervals. The evaluation emphasizes two classes of metrics: extreme-event detection and spectral skill assessment (Chen et al., 7 Jun 2026).

For extreme-event detection, the threshold is the 95th percentile $q$5. The reported metrics are Probability of Detection, False Alarm Ratio, and Critical Success Index:

$q$6

$q$7

$q$8

PredHydro-Net is reported to maintain substantially higher $q$9 than Earthformer, PredRNNv2, and the operational GFS baseline, especially at long lead times. It also improves the neighborhood-scale Fractions Skill Score at $F_M\in\mathbb{R}^{C\times h\times w}$0:

$F_M\in\mathbb{R}^{C\times h\times w}$1

where $F_M\in\mathbb{R}^{C\times h\times w}$2 and $F_M\in\mathbb{R}^{C\times h\times w}$3 are neighborhood-averaged observations and predictions over radius $F_M\in\mathbb{R}^{C\times h\times w}$4.

For spectral skill, the assessment uses radially averaged power spectral density of hydrometeor fields at each pressure level, together with a spectral RMSE:

$F_M\in\mathbb{R}^{C\times h\times w}$5

The paper states that PredHydro-Net’s spectra remain much closer to ERA5 across wavenumbers $F_M\in\mathbb{R}^{C\times h\times w}$6, whereas MSE-only models attenuate high frequencies too sharply. In summary form, the reported result is a favorable trade-off between low RMSE at the global scale and faithful extreme-event and high-frequency structure in three-dimensional cloud and precipitation prediction.

6. Physical interpretability, empirical scope, and scientific significance

Beyond headline forecast metrics, the model is described as showing strong climatological consistency with Global Precipitation Measurement satellite retrievals. It also reasonably reproduces three-dimensional cloud structures in extreme weather events, with Hurricane Ian named as an example (Chen et al., 7 Jun 2026).

Feature attribution is reported to confirm dependence on physical precursors such as relative humidity and wind convergence. Within the framework of the paper, this serves two purposes. First, it supports the claim that the hydrometeor branch is responding to physically meaningful large-scale controls rather than only fitting texture statistics. Second, it provides a post hoc interpretability argument that is consistent with the one-way thermodynamic-to-hydrometeor modulation used in the architecture.

The significance of PredHydro-Net lies in how it organizes several methodological elements around the specific pathology of hydrometeor prediction. It is not presented merely as a larger recurrent predictor or merely as a GAN-augmented weather model. Instead, its defining contribution is the coordinated combination of decoupled decoding, FiLM-based physical guidance, wavelet-domain supervision, spectral alignment, and local adversarial texture priors. This suggests a broader methodological lesson: for atmospheric targets with severe zero inflation and long tails, forecast quality depends not only on minimizing aggregate error but also on controlling how information is distributed across physical scales.

At the same time, the empirical scope described in the paper is specific. The reported results are tied to a 72-h global evaluation on ERA5 at $F_M\in\mathbb{R}^{C\times h\times w}$7 resolution, with comparisons to Earthformer, PredRNNv2, and GFS. Broader claims beyond that setup would require additional evidence. The paper’s own framing is therefore best read as a robust, physics-informed approach to long-tailed atmospheric prediction rather than as a universal solution to all weather-forecasting tasks.