PIANO: Physics-informed Dual Neural Operator for Precipitation Nowcasting

AI summary

Overview

Research area: Machine learning for weather forecasting, specifically satellite-based precipitation nowcasting using physics-informed neural operators.

Technical level: Advanced. The paper assumes familiarity with neural operators, physics-informed neural networks (PINNs), partial differential equations (advection-diffusion), generative adversarial networks, and standard nowcasting metrics such as the critical success index (CSI).

Scope: A proof-of-concept framework, PIANO, that couples a data-driven time-stepping operator with a physics-informed velocity-extraction operator to forecast satellite imagery and then translate those forecasts into radar-based precipitation predictions.

What This Paper Is About

Accurate short-term precipitation prediction (nowcasting) matters for disaster early warning, but traditional numerical weather prediction (NWP) models are computationally heavy with low spatial resolution, and radar-based machine learning methods only work where radar coverage exists. Satellite imagery offers near-global coverage, but prior satellite-based nowcasting studies have largely ignored the underlying physics of atmospheric motion. PIANO addresses this by enforcing the advection-diffusion equation during training of a dual neural operator, then converting the predicted satellite frames into radar images for precipitation nowcasting.

Key Contributions

1. A dual-operator architecture. PIANO splits the task into a Time-stepping Neural Operator (T-NO) that learns purely data-driven temporal evolution and a Velocity-extraction Neural Operator (V-NO) that learns the underlying flow field. Decoupling the two lets the authors enforce physical consistency without limiting the expressiveness of the forecasting operator.

2. Physics constraints via the advection-diffusion equation. The V-NO is trained with a PINN loss derived from the advection-diffusion equation, discretized with finite differences, using the diffusion coefficient D (initial value 1) and source term R (initial value 0) as learned H × W scalar matrices.

3. A satellite-to-radar translation pipeline. A Pix2Pix ResNet trained in a GAN structure, with a ResNet generator and a ResNet-based discriminator, converts PIANO's predicted satellite images into radar images usable for precipitation nowcasting, following the motivation of Park et al. (2025).

4. A baseline evaluation on the Sat2RDR dataset. The authors benchmark PIANO against PhyDNet and the Neural Prediction Model (NPM) on CSI for moderate (above 4 mm/h) and heavy (above 8 mm/h) precipitation, plus a sensitivity analysis over the physics-loss weight α.

Main Findings

• Better moderate-precipitation skill: PIANO outperforms both PhyDNet and NPM on CSI 4mm at every reported lead time from 1h to 8h. Values range from 0.757 (1h) to 0.763 (6h, 7h, 8h), versus NPM's 0.750 to 0.756 and PhyDNet's 0.737 to 0.763.

• Strong early heavy-precipitation skill, then a crossover: On CSI 8mm, PIANO leads in the first two hours (0.611 at 1h, 0.612 at 2h) but PhyDNet becomes more accurate at longer lead times, rising from 0.601 at 1h to 0.654 at 8h, while PIANO stays in the 0.611 to 0.616 range. The paper notes the same trend of PhyDNet outperforming NPM at later frames was observed by Park et al. (2025).

• Physically plausible velocity fields: The V-NO output on January 1st, 2024 shows a generally rightward and downward velocity trend, which the authors state matches the previously known atmospheric trend over South Korea (Baek et al., 2015).

• Low seasonal variability: CSI 4mm performance was consistent across all months. The authors note the degree of seasonal change is low relative to NPM, which showed variation up to 0.1 in Park et al. (2025).

• The physics-loss weight α must be tuned: Satellite prediction MSE at 8h lead time is 0.592 for α = 5.0, 0.569 for α = 1.0, 0.580 for α = 0.2, and 0.576 for α = 0.0. Both over-constraining (α = 5) and under-constraining (α = 0.2) hurt, but a well-chosen α improves longer-lead predictions.

• The first hour is not helped by the PINN loss: T-NO performs best at a one-hour lead time (0.202 MSE with α = 0.0, 0.205 with α = 1.0), which the authors attribute to the PINN loss not constraining the first-hour prediction at all.

• Caveat on absolute CSI values: The paper explicitly states its CSI values are "incredibly larger" than those reported by Andrychowicz et al. (2023) and Park et al. (2025), possibly because differently sampled Sat2RDR data led to relatively simpler predictions.

Methodology in Plain English

The framework runs two neural operators in series. T-NO takes eight past satellite frames plus a digital elevation map and predicts the next eight frames, learning only from data via a mean squared error loss. V-NO takes frames and outputs eight velocity fields, trained with a physics loss that checks whether the frames evolve according to a discretized advection-diffusion equation: the change between consecutive frames should equal a diffusion term minus an advection term plus a source term.

Training happens in two steps: V-NO and T-NO are pre-trained in parallel, then combined and fine-tuned with a total loss that adds the data loss to the physics loss scaled by a weight α. Because finite differencing is used, gradients and Laplacians are computed with convolution matrices.

After prediction, a Pix2Pix ResNet trained with a GAN converts the forecast satellite images into radar images, which are what the precipitation metrics are computed on.

Data comes from the Sat2RDR dataset: GK2A satellite infrared (10.5 µm), upper-level water vapor (6.3 µm), lower-level water vapor (7.3 µm), and DEM data for South Korea. The authors subsample January–August 2020–2024, excluding September–December because of missing 0600-hour values that would make training temporally unbalanced. Training patches are 768 × 768 random crops with season-aware sampling yielding twenty frames per month; 2023 is validation and 2024 is test. The Pix2Pix ResNet uses 224 × 224 crops with flip augmentation. Combined precipitation performance is evaluated on January–August 2022 and 2023. Baselines are PhyDNet and NPM, trained on a single RTX 5090 with batch size three. Performance is measured with CSI = TP / (TP + FP + FN).

Why This Matters

Research impact: The paper shows that injecting a governing physical equation as an inductive bias into satellite-based nowcasting can improve moderate-precipitation skill and stabilize performance across seasons, and it offers an explicitly labeled baseline for future physics-informed nowcasting work. It also sidesteps the supercomputer-scale ERA5 dependency of models like Pangu-Weather, FourCastNet, and GraphCast.

Real-world applications:

• Early warning of extreme precipitation, flooding, landslides, and soil erosion in regions without radar coverage.
• Agricultural planning, where the paper links extreme precipitation to reduced agricultural productivity.
• Public health risk reduction from precipitation-related hazards.
• Nowcasting services for developing countries and other areas that lack radar infrastructure but are covered by satellites.

Industry relevance: The framework is described as simple to train and deploy on satellite imagery, and it was developed partly at TelePIX, a company, suggesting direct relevance to commercial Earth-observation and weather service providers. The generative satellite-to-radar step also makes output compatible with existing radar-based precipitation products.

Future Directions

• Tune model hyperparameters, including the physics-loss weight α, whose value the paper shows materially changes performance.
• Run comprehensive evaluations against strong baselines to confirm whether PIANO's advantage holds under other data sampling and evaluation setups, given the authors' own caveat that their CSI values are much higher than previously reported.
• Address the long-lead heavy-precipitation gap, where PhyDNet overtakes PIANO beyond the first two hours at the 8 mm/h threshold.
• Investigate why the PINN loss does not improve first-hour predictions, since it does not constrain that first step, and determine whether a reformulation could help.

Target Audience

Researchers and practitioners in machine learning for weather and climate, particularly those working on neural operators, physics-informed learning, and remote-sensing-based nowcasting. It will also interest operational meteorology teams and Earth-observation companies evaluating satellite-only alternatives to radar- and NWP-based pipelines, and readers who want a compact proof-of-concept baseline for physics-informed precipitation nowcasting.

Authors’ abstract

Precipitation nowcasting, key for early warning of disasters, currently relies on computationally expensive and restrictive methods that limit access to many countries. To overcome this challenge, we propose precipitation nowcasting using satellite imagery with physics constraints for improved accuracy and physical consistency. We use a novel physics-informed dual neural operator (PIANO) structure to enforce the fundamental equation of advection-diffusion during training to predict satellite imagery using a PINN loss. Then, we use a generative model to convert satellite images to radar images, which are used for precipitation nowcasting. Compared to baseline models, our proposed model shows a notable improvement in moderate (4mm/h) precipitation event prediction alongside short-term heavy (8mm/h) precipitation event prediction. It also demonstrates low seasonal variability in predictions, indicating robustness for generalization. This study suggests the potential of the PIANO and serves as a good baseline for physics-informed precipitation nowcasting.