Organizations: Artificial Intelligence Innovation and Incubation Institute, Fudan University · Shanghai Academy of Artificial Intelligence for Science · Shanghai Innovation Institute · National Marine Environment Forecasting Center · Department of Atmospheric and Oceanic Sciences, Fudan University
Current AI weather forecasting models predict conventional atmospheric variables but cannot distinguish between cloud microphysical species critical for aviation safety. We introduce AviaSafe, a hierarchical, physics-informed neural forecaster that produces global, six-hourly predictions of these four hydrometeor species for lead times up to 7 days. Our approach addresses the unique challenges of cloud prediction: extreme sparsity, discontinuous distributions, and complex microphysical interactions between species. We integrate the Icing Condition (IC) index from aviation meteorology as a physics-based constraint that identifies regions where supercooled water fuels explosive ice crystal growth. The model employs a hierarchical architecture that first predicts cloud spatial distribution through masked attention, then quantifies species concentrations within identified regions. Training on ERA5 reanalysis data, our model achieves lower RMSE for cloud species compared to baseline and outperforms operational numerical models on certain key variables at 7-day lead times. The ability to forecast individual cloud species enables new applications in aviation route optimization where distinguishing between ice and liquid water determines engine icing risk.
Figures & tables
Figure 1: Architecture of our forecasting framework. The model takes historical atmospheric states Xt−1 and Xt as input to predict Xt+1 . (a) The overall architecture comprises two modules: 1) The Prediction Backbone (red) with Encoder E , Swin Transformer blocks, and decoupled Decoders ( DCloud , D ), outputting feature Fbackbone and the final forecast. 2) The Physics-Informed Guidance module (green) with IC Block, Encoder EMask , Mask Predictor, and a convolutional layer (Conv). It uses physical masks and Fbackbone to produce the guidance signal ( Maskt+1 ) for cloud prediction. (b) IC Block: Illustrates the IC Function with temperature ( T ) and humidity ( Q ) inputs. (c) Mask Predictor: Details the internal structure including MLP, Swin Block, and Deconvolutional Decoder (Deconv).
Figure 2: Performance comparison between our model (AviaSafe) and ECMWF HRES model. Results are averaged over the 00 UTC and 12 UTC initialization times on the test set. The top row shows the latitude-weighted Root Mean Square Error (RMSE, lower is better), and the bottom row shows the latitude-weighted Anomaly Correlation Coefficient (ACC, higher is better) for five key variables at 500hPa. The x-axis represents the forecast lead time up to 7 days (28 steps). In all panels, the red line indicates our proposed AviaSafe model, while the blue line represents the ECMWF HRES model for comparison.
Figure 3: Mean NRMSE improvement of AviaSafe over the Baseline across all variables for a 7-day forecast. Each row represents different variable, averaged over all 13 pressure levels. Blue indicates that our model is better (lower RMSE). The data shows that AviaSafe outperforms the baseline in 93.7% of all variable/time-step combinations shown.
Figure 4: Time series of the mean NRMSE improvement for CIWC over a 15-day forecast. The line shows the NRMSE improvement (%) for the CIWC variable, averaged over all pressure levels, where negative values indicate an improvement.
Cloud Microphysics
Background Variables
Model
CIWC50
CLWC100
CRWC250
CSWC600
U500
V500
Z500
T600
Q600
( ×10−8 g/kg)
( ×10−8 g/kg)
( ×10−8 g/kg)
( ×10−5 g/kg)
(m/s)
(m/s)
(m 2 /s 2 )
(K)
( ×10−4 g/kg)
Baseline
1.059
1.318
1.010
4.359
2.857
2.903
139.52
0.898
6.908
w/o (MP, IC)
1.012
0.892
0.889
4.346
2.857
2.899
138.95
0.900
6.899
w/o IC
1.053
0.968
0.961
4.341
2.832
2.874
137.23
0.890
6.886
AviaSafe Model (Ours)
0.956
0.875
0.863
4.340
2.826
2.871
135.81
0.889
6.855
Table 1: Ablation study results (average RMSE over the first 5 days of the forecast). Lower is better. Best results are in bold.
Figure 5: CNOP-derived spatial importance map for the moist energy norm. Shading shows the CNOP initial-perturbation pattern; black contours denote the initial 500 hPa geopotential height; the red box marks the target region.
Figure 6: Comparison of 72-h forecasts at 500 hPa for CTRL vs CNOP-perturbed initial conditions. Initialization: 2024-01-04 12:00 UTC. The map domain spans 82°E–140°E in longitude and 15°N–50°N in latitude. The red bounding box (105°E–121°E, 21°N–30°N) highlights the target region of interest. Rows (top to bottom): Q, T, CLWC, and CIWC. Columns: control forecast (Control, left), forecast with a CNOP initial perturbation (Perturbed, center), and the Perturbed Control difference (right). Units are as indicated by the color bars.
Figure 7: Spatiotemporal evolution of the control forecast at 500 hPa. This figure displays the 72-hour forecast trajectory starting from the unperturbed initial state. The columns represent forecast lead times from +12h to +72h. The rows, from top to bottom, illustrate: specific humidity (Q500) overlaid with wind vectors, temperature (T500) with geopotential height contours, cloud liquid water content (CLWC500), and cloud ice water content (CIWC500). The red bounding box highlights the target region for verification. This baseline evolution serves as the reference for assessing the impact of initial state perturbations.
Figure 8: Spatiotemporal evolution of the CNOP-perturbed forecast at 500 hPa. This figure shows the forecast results initialized with the CNOP-perturbed state ( x0+δx0∗ ), using the same variable layout and time intervals as the control forecast.
Weather forecasting predicts future atmospheric conditions, from local storms to global patterns, using physical models, observations, and machine learning.
We introduce AIFS-COMPO, a skilful medium-range data-driven global forecasting system for aerosols and reactive gases. Building on the ECMWF Artificial Intelligence Forecast System (AIFS), AIFS-COMPO employs a transformer-based encoder-processor-decoder architecture to jointly model meteorological and atmospheric composition variables. The model is trained on Copernicus Atmosphere Monitoring Service (CAMS) reanalysis, analysis, and forecast data to learn the coupled dynamics of weather, emissions, transport, and atmospheric chemistry. We evaluate AIFS-COMPO against a range of atmospheric composition observations and compare its performance with the operational CAMS global forecasting system IFS-COMPO. The results show that AIFS-COMPO achieves comparable or improved forecast skill for several key species while requiring only a fraction of the computational resources. Furthermore, the efficiency of the approach enables forecasts beyond the current operational horizon, demonstrating the potential of AI-based systems for fast and accurate global atmospheric composition prediction.
Paula Harder, Johannes Flemming, Mihai Alexe +3
European Centre for Medium-Range Weather Forecasts (ECMWF)
While global data-driven models excel at predicting continuous atmospheric variables, three-dimensional hydrometeor forecasting remains challenging due to the zero-inflated, long-tailed distributions of these variables. Standard deep learning optimization often yields overly smooth forecasts, attenuating extreme events and spatial textures. We propose PredHydro-Net, a physics-guided dual-decoding framework that mitigates this smoothing. To resolve multi-variable optimization conflicts, it employs a decoupled architecture where macroscopic thermodynamic and dynamic fields unidirectionally modulate hydrometeor generation. By integrating wavelet-based frequency decoupling, spectral amplitude matching, and adversarial training, the model achieves a favorable trade-off between quantitative accuracy and spatial fidelity. In a 72-h global evaluation, PredHydro-Net outperforms both spatiotemporal deep learning baselines (Earthformer and PredRNNv2) and the operational Global Forecast System (GFS) in extreme-event detection and spectral representation. Furthermore, it demonstrates strong climatological consistency with Global Precipitation Measurement (GPM) satellite retrievals. The model reasonably reproduces the three-dimensional cloud structures in extreme weather events, such as Hurricane Ian. Feature attribution confirms its dependence on physical precursors such as relative humidity and wind convergence, offering a robust, physics-informed approach to long-tailed atmospheric prediction.
Tropical cyclone (TC) forecasting is critical for disaster warning and emergency response. Deep learning methods address computational challenges but often neglect physical relationships between TC attributes, resulting in predictions lacking physical consistency. To address this, we propose Phys-Diff, a physics-inspired latent diffusion model that disentangles latent features into task-specific components (trajectory, pressure, wind speed) and employs cross-task attention to introduce prior physics-inspired inductive biases, thereby embedding physically consistent dependencies among TC attributes. Phys-Diff integrates multimodal data including historical cyclone attributes, ERA5 reanalysis data, and FengWu forecast fields via a Transformer encoder-decoder architecture, further enhancing forecasting performance. Experiments demonstrate state-of-the-art performance on global and regional datasets.
Lei Liu, Xiaoning Yu, Kang Chen +4
University of Science and Technology of China, Hefei, China · Hefei University of Technology, Hefei, China