🤖 machine learning

QWRF-Net: A Quantum-Wavelet Framework with Rectified Flow for Short-Term Precipitation Nowcasting

QWRF-Net is a novel quantum-wavelet framework that integrates wavelet-based multi-scale decomposition, differentiated quantum-inspired modulation, and rectified flow generation to significantly improve the accuracy and structural preservation of short-term precipitation nowcasting, particularly for intense convective rainfall events.

Original authors: Zhuo Wang, Chaorong Li, Wenjie Luo, Chuanhu Deng

Published 2026-08-04

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhuo Wang, Chaorong Li, Wenjie Luo, Chuanhu Deng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are watching a time-lapse video of a storm rolling over a city. The clouds swirl, raindrops gather, and suddenly, a massive downpour hits a specific neighborhood. Predicting exactly where that rain will fall, how hard it will hit, and how long it will last is the job of "precipitation nowcasting." It's like trying to guess the next few moves in a chaotic game of chess where the pieces are made of water and the board is the sky. Scientists use giant radar dishes to take snapshots of these storms, but the pictures are messy. They contain huge, blurry rain clouds mixed with tiny, sharp, intense bursts of rain that can cause flash floods in minutes. The problem is that when computers try to predict the future of these storms, they often struggle. As they look further into the future (say, an hour ahead), the sharp edges of the rain get blurry, the intense centers fade away, and the prediction turns into a fuzzy, washed-out guess. This is dangerous because if you don't know exactly where the heavy rain is, you can't warn people in time to save their homes.

To solve this, researchers are building smarter computer models that act like weather detectives. They use a mix of old-school math tricks and brand-new AI techniques to keep the storm's details sharp. One key idea is to break the storm picture into different "layers" of detail, separating the big, slow-moving clouds from the fast, sharp edges. Another idea is to use a special kind of math that mimics how quantum computers think (even though these models run on regular computers) to tweak those layers just right. Finally, instead of guessing one minute at a time and hoping for the best, they use a smooth, continuous flow to draw the entire future storm at once. The goal is to keep the "danger zones" of the storm clear and bright, so emergency teams know exactly where to send help.

The Paper: QWRF-Net

In this paper, a team of scientists introduces a new AI model called QWRF-Net. Think of it as a super-powered weather forecaster designed specifically to keep the "spicy" parts of a storm from turning into "mild" mush as it looks into the future. The researchers built this model to tackle the two biggest headaches in short-term rain prediction: keeping the storm's complex shapes clear and making sure the prediction doesn't get worse the further out you look.

Here is how QWRF-Net works, broken down into three fun steps:

1. The Wavelet "Magic Lens" (Separating the Noise)
Imagine you have a messy pile of LEGOs representing a storm. Some are big, flat baseplates (the general rain area), and some are tiny, jagged spikes (the dangerous, heavy rain cores). Old models tried to look at the whole pile at once, which made the tiny spikes get lost in the noise. QWRF-Net uses a Wavelet Decomposition tool. This is like a magical lens that instantly sorts the LEGOs into four separate bins:

LL (Low-Low): The big, smooth background rain.
LH, HL, HH: The sharp edges, sudden changes, and intense spikes.
By sorting them first, the model can see the "danger spikes" clearly without them getting blurred by the "boring background."

2. The "Quantum" Tweak (The Special Sauce)
Once the LEGOs are sorted, the model gives each bin a special makeover. This is where the "Quantum" part comes in. Don't worry, there are no actual quantum computers or spooky physics happening here! The researchers used a Quantum-Inspired module. Think of this as a very clever, structured filter that acts like a quantum computer would, but it runs on a normal computer.

It looks at the "background rain" bin and gives it a gentle nudge to keep it steady.
It looks at the "sharp spike" bins and gives them a strong, specific boost to make sure those intense rain cores stay bright and sharp.
The paper argues that treating these different parts of the storm differently is crucial. If you treat the whole storm the same way, the intense parts get smoothed out, and the warning system fails.

3. The "Rectified Flow" (The Smooth Slide)
Finally, the model has to draw the future storm. Old methods often tried to guess the next minute, then use that guess to guess the minute after, and so on. This is like trying to walk a tightrope by looking at your feet; if you stumble once, you fall further and further off.
QWRF-Net uses a Rectified Flow method. Instead of stepping one by one, it draws a smooth, continuous slide from "now" to "60 minutes in the future." It calculates the speed and direction of the rain all at once and slides the storm forward in one go. This prevents the errors from piling up, keeping the prediction sharp even at the 60-minute mark.

What They Found

The team tested QWRF-Net on two real-world radar datasets: one from the Netherlands (KNMI) and a massive global storm dataset (SEVIR). They compared it against other top models like RainNet, NowcastNet, and DiffCast.

The Results: QWRF-Net didn't just win by a little bit; it was consistently better at predicting the intense, heavy rain. While other models got fuzzy and lost the "core" of the storm as time went on, QWRF-Net kept the heavy rain centers sharp and in the right place.
The "Extreme" Test: They even tested it on the worst storms (extreme events). In these tough cases, QWRF-Net was the most accurate at keeping the structure of the storm intact.
The "Why": When they took apart their own model (an "ablation study"), they found that all three parts were necessary.
- If they removed the "Wavelet Lens," the model got messy.
- If they removed the "Quantum Tweak" and just used normal math, the model wasn't as good at boosting the sharp edges.
- If they removed the "Smooth Slide" and went back to guessing step-by-step, the prediction got blurry faster.

What This Means (And What It Doesn't)

The authors suggest that QWRF-Net offers a promising new way to handle short-term rain warnings. By keeping the "danger zones" of the storm clear, it could help emergency teams know exactly where to send sandbags or evacuate people.

However, the paper is careful to note a few things:

It's a 60-minute window: The model is designed for the next hour. It doesn't claim to predict the weather for next week.
It's a simulation: The "Quantum" part is a clever math trick running on normal computers, not a real quantum computer.
It's a starting point: The model predicts the rain, but it doesn't yet simulate the flooding itself. The authors hope that in the future, this accurate rain prediction can be fed into flood models to save even more lives.

In short, QWRF-Net is like giving a weather forecaster a pair of high-definition glasses and a steady hand, ensuring that when a storm hits, the warning is as sharp and clear as the rain itself.

Technical Summary: QWRF-Net for Short-Term Precipitation Nowcasting

1. Problem Statement

Short-term precipitation nowcasting is critical for hydrometeorological early warning, particularly for mitigating high-impact hazards like urban flooding and flash floods triggered by intense convective rainfall. The paper identifies two primary, coupled challenges in this domain:

Complex Multi-Scale Representation: Radar precipitation fields contain strongly coupled multi-scale structures, ranging from broad precipitation organization to localized intense cores and fine-scale boundaries. Existing models often struggle to represent these intertwined structures adequately within a single feature space, leading to a loss of structural fidelity in intense rainfall cores.
Progressive Degradation in Multi-Step Forecasting: In warning-oriented settings, forecast quality often degrades at later lead times. Small prediction errors accumulate over recursive multi-step predictions, resulting in blurred structures, weakened intensities, and reduced stability. This degradation is particularly detrimental when the forecast horizon extends to the 30–60 minute window required for emergency response.

Current deep learning methods, while faster than numerical weather prediction (NWP) models, often suffer from these issues. Deterministic models tend to smooth intense echoes, while generative models (e.g., GANs, Diffusion) may introduce high inference costs or still fail to preserve structural details over longer horizons.

2. Methodology: QWRF-Net

The authors propose QWRF-Net, a conditional generative framework designed to jointly improve multi-scale precipitation representation and stable multi-step generation. The architecture follows a "representation-to-generation" strategy, consisting of three main components:

A. Overall Architecture

The model adopts a U-Net-style encoder-decoder backbone. It takes the past 6 radar frames (30 minutes) as input ( $X_{in}$ ) and generates the next 12 frames (60 minutes) as output ( $X_{tar}$ ). Historical context is injected throughout the encoder via a gated fusion operator that adaptively balances main feature streams with conditioning branches.

B. Quantum–Wavelet Bottleneck (Scale-Aware Representation)

Placed at the deepest level of the U-Net, this module reorganizes mixed latent features to separate scales before nonlinear modulation:

Wavelet Decomposition: A Discrete Wavelet Transform (DWT) decomposes the bottleneck feature map into four sub-bands:
- LL: Low-frequency background information (broad precipitation organization).
- LH, HL, HH: High-frequency directional details (boundaries, local gradients, intense cores).
Quantum-Inspired Modulation: Each sub-band is processed independently by a learnable transformation simulated on classical hardware. This involves:
- Compressing the sub-band via Global Average Pooling (GAP).
- Encoding the descriptor into a quantum state using angular embedding.
- Processing via a Variational Quantum Circuit (VQC) composed of parameterized rotation and entangling gates.
- Measuring the output via Pauli-Z expectations to generate a classical modulation term.
- Note: The authors clarify this is a classically simulated structured nonlinear operator, not a claim of practical quantum advantage.
Reconstruction: The modulated sub-bands are fused back into the spatial domain using an Inverse Discrete Wavelet Transform (IDWT) and a residual connection to preserve the underlying latent representation.

C. Rectified-Flow-Based Generation (Non-Autoregressive Decoding)

Instead of recursive frame-by-frame prediction, the model uses a rectified flow formulation to generate the future sequence:

Training: The model learns a conditional velocity field $v_\theta$ that transports an initial Gaussian noise sample ( $Z_0$ ) to the target precipitation field ( $Z_1$ ) via a continuous-time interpolation $Z_t = tZ_1 + (1-t)Z_0$ . The objective minimizes the difference between the predicted velocity and the true velocity ( $Z_1 - Z_0$ ).
Inference: The future sequence is generated by solving the conditional Ordinary Differential Equation (ODE) using an Euler solver. This non-autoregressive approach aims to reduce error accumulation across the forecast horizon.

3. Key Contributions

Joint Formulation: The paper formulates nowcasting as a joint problem of conditional representation and future-sequence generation, explicitly targeting the preservation of multi-scale structures relevant to warning-oriented forecasting.
Quantum–Wavelet Bottleneck: Introduction of a novel bottleneck that disentangles latent features into wavelet sub-bands and applies differentiated, quantum-inspired structured modulation. This allows for distinct processing of background organization versus localized intense structures.
Rectified-Flow Integration: Integration of the refined representation with a rectified-flow-based non-autoregressive decoder to reduce degradation in later lead times.
Unified Evaluation: Demonstration of favorable performance on KNMI and SEVIR benchmarks under a unified protocol, showing consistent gains in medium-to-high precipitation thresholds and extreme-event subsets.

4. Experimental Results

The model was evaluated on the KNMI (Netherlands radar) and SEVIR (satellite/radar VIL) datasets using a 6→12 frame prediction protocol.

Quantitative Performance:
- KNMI: QWRF-Net achieved the highest Critical Success Index (CSI) and Heidke Skill Score (HSS) across all thresholds (0.5 to 30 mm/h), along with the lowest MAE/RMSE and highest SSIM. Improvements were particularly notable at higher thresholds ( $r \ge 10$ and $r \ge 30$ mm/h).
- SEVIR: The model outperformed strong baselines (e.g., NowcastNet, CoDiCast, DiffCast) in CSI and HSS at medium-to-high VIL thresholds and achieved the highest SSIM.
- Extreme Events: On a subset of SEVIR containing extreme events (peak VIL $\ge$ 219), QWRF-Net achieved the lowest MAE and RMSE and the highest SSIM, indicating robustness in preserving intense cores.
Qualitative Analysis: Visual comparisons showed that QWRF-Net better preserved the morphology, intensity distribution, and boundary sharpness of precipitation cores compared to discriminative and generative baselines, especially at later lead times (T+40 to T+60 min).
Ablation Studies:
- Removing the wavelet bottleneck (QWRF-Net-A) caused a significant performance drop.
- Using only wavelets without quantum modulation (QWRF-Net-B) or using shared processors (QWRF-Net-D) yielded lower performance than the full model, confirming the necessity of both scale disentanglement and differentiated sub-band modulation.
- Replacing the flow-based decoder with a discriminative head (QW-Net) resulted in lower performance at high thresholds and later lead times, validating the benefit of non-autoregressive generation.

5. Significance and Claims

The paper claims that QWRF-Net offers a promising direction for warning-oriented short-term precipitation nowcasting. Its significance lies in:

Structural Preservation: By explicitly separating scales and modulating them differently, the framework better preserves the location and intensity of convective cores, which are critical for flood warnings.
Stability: The rectified-flow generation mechanism provides more stable forecasts over the 60-minute horizon, reducing the progressive degradation common in recursive models.
Downstream Utility: The authors suggest that these improvements in structural fidelity and intensity accuracy may provide a more useful precipitation basis for downstream hydrological applications (e.g., runoff generation, inundation modeling), although the study does not directly simulate these downstream processes.

The authors maintain a modest stance, noting that the quantum component is a classically simulated operator and that the current study is limited to a 60-minute horizon. They propose that future work should explore longer horizons, higher resolutions, and direct coupling with hydrological models to assess practical decision-support value.