Operational weather forecasting centers increasingly rely on a diverse collection of numerical weather prediction models to anticipate severe weather events, yet synthesizing conflicting predictions into a single meteorologically coherent forecast remains an enduring challenge. While conventional multi-model combinations rely on arithmetic or weighted ensemble averaging, simple averaging frequently distorts atmospheric dynamics by weakening pressure gradients, diffusing localized wind fields, and splitting unified atmospheric systems such as tropical cyclones and frontal boundaries into artificial dual structures.
The operational landscape of modern meteorology encompasses both traditional physics-based dynamical models and an emerging class of artificial intelligence systems, presenting forecasters with an unprecedented volume of predictive data. Operational centers worldwide maintain specialized suites of physics-based models, including the European Centre for Medium-Range Weather Forecasts Integrated Forecasting System, the UK Met Office Unified Model, the National Oceanic and Atmospheric Administration Global Forecast System and High-Resolution Rapid Refresh, and the Japan Meteorological Agency Global Spectrum Model and Meso-Scale Model. Concurrently, data-driven deep learning models such as FourCastNet, GraphCast, Pangu-Weather, FengWu, Aurora, GenCast, and the Artificial Intelligence Forecasting System have demonstrated competitive accuracy and rapid execution times. However, when these diverse models produce divergent forecasts of comparable historical accuracy, forecasters cannot determine in advance which individual model will prove superior. Averaging outputs has traditionally served to cancel out random errors, yet the resulting fields routinely produce uninterpretable artifacts, leaving meteorologists torn between choosing a single, structurally intact model that might be inaccurate or an averaged field that destroys realistic atmospheric features.
In operational practice, this tension creates acute procedural friction, particularly in national weather services where strict hierarchical protocols govern the issuance of official advisories and public warnings. At agencies such as the Japan Meteorological Agency, the National Hurricane Center, and the Joint Typhoon Warning Center, consensus forecasts derived from averaging the center coordinates of tropical cyclones establish the authoritative track guidance that takes precedence over other operational outputs. Because no single numerical weather prediction model natively conforms to this averaged consensus track, operational forecasters handling general, marine, and aviation forecasts are forced to mentally adjust two-dimensional wind, pressure, and moisture fields to align with the official storm center. This manual mental recalibration demands considerable time and cognitive effort during critical warning scenarios, creating a bottleneck that impedes workflow standardization. DeepMedcast directly alleviates this operational burden by automatically producing continuous, two-dimensional mean sea-level pressure and surface wind fields that naturally center atmospheric systems along the consensus pathway without distorting the surrounding circulation patterns.
The foundational training concept underlying DeepMedcast departs from standard machine learning post-processing paradigms by eliminating reliance on observational target data during training. Conventional post-processing approaches like model output statistics often struggle with dataset homogeneity because operational numerical models undergo periodic structural and resolution upgrades that alter their underlying systematic bias profiles. DeepMedcast circumvents this data-maintenance dilemma by training exclusively on temporal sequences extracted from a single numerical model, specifically the Japan Meteorological Agency Global Spectrum Model spanning nine years from January 2013 through December 2021, validated against data from 2022. During the network training phase, the system takes two forecast lead times from the same model run—specifically lead time t minus delta t and lead time t plus delta t—as input channels, using the intermediate forecast at lead time t as the supervisory ground truth. Because both inputs and target originate from the same physically consistent model run, the learning process avoids observational noise, while channel-order swapping ensures symmetrical representations that encourage intermediate features to locate at the exact arithmetic mean.
Architecturally, DeepMedcast is built upon a fully convolutional U-Net framework designed to capture multi-scale spatial representations across complex atmospheric fields. The network accepts two input channels representing the input states and outputs a single channel representing the intermediate prediction. The encoder pathway employs convolutional layers with a three-by-three kernel, unit stride, and unit padding, interspersed with two-by-two max-pooling layers with a stride of two that progressively downsample feature maps across successive stages. This downsampling expands the channel capacity from an initial 256 channels up to 2048 channels while reducing the spatial resolution to one-eighth of the original grid dimensions. The decoder pathway symmetrically reconstructs spatial resolution through two-by-two transposed convolutional layers that double spatial dimensions and halve channel depths at each stage. Rectified linear units provide nonlinear activation throughout hidden layers, whereas the output layer applies a sigmoid function to bound values between zero and one. Data normalization using domain-wide extrema precedes network ingestion on a 121-by-151 target grid around Japan, extended to 128-by-158 dimensions for processing and subsequently denormalized after optimization via the Adam algorithm under a mean square error loss function.
To maintain computational efficiency across diverse operational parameters, separate deep neural networks were trained independently for distinct physical variables, including surface wind components, temperature, relative humidity, mean sea-level pressure, and upper-air variables at the 700-hectopascal level. Although the primary architecture computes an intermediate forecast between two input models, the researchers established a recursive processing protocol capable of combining four, eight, sixteen, or more disparate forecast models. Under this hierarchical workflow, intermediate fields are initially synthesized between distinct model pairs, after which the resulting intermediate outputs serve as inputs to subsequent DeepMedcast layers. Comprehensive evaluation demonstrated that the model exhibits remarkable structural robustness regardless of the order in which input models are introduced. While minor variations emerge due to slight numerical asymmetries within the trained convolutional weights, the overall predicted cyclone tracks, core central pressures, and regional wind fields remain qualitatively consistent across alternate input permutations.
The structural efficacy of the framework was validated through detailed case studies beginning in January 2023, completely independent of the training and validation datasets. In an evaluation focusing on Typhoon Lan in August 2023 at a 51-hour lead time, the Global Spectrum Model placed the typhoon center at 33.3 degrees north, 137.1 degrees east with a central pressure of 960 hectopascals, while the Meso-Scale Model placed it southwest at 32.7 degrees north, 135.7 degrees east with an identical 960 hectopascal central pressure. An arithmetic mean of these two fields produced an unrealistic split system with two distinct low-pressure centers, an attenuated central pressure of 974 hectopascals, and maximum winds degraded from 79 and 68 knots down to 58 knots. In contrast, DeepMedcast generated a single, sharply resolved cyclonic vortex centered precisely at 33.0 degrees north, 136.4 degrees east, maintaining a realistic central pressure of 960 hectopascals and maximum sustained winds of 69 knots without displaying unnatural spatial elongation.
A parallel investigation of a stationary frontal zone south of Tokyo in June 2024 at a 30-hour lead time further illustrated the model’s meteorological consistency. While the Global Spectrum Model positioned the front south of Tokyo with cool northeasterly surface winds and the Meso-Scale Model positioned it north of Tokyo with warm southwesterly flow, arithmetic averaging smeared the frontal zone into two separate, weakened wind shear lines and uncoupled the thermal boundary from the pressure field. DeepMedcast successfully synthesized a single, sharp frontal boundary situated midway between both inputs, precisely aligning intense directional wind shear with the 21-degree Celsius surface isotherm. In an even more challenging test involving Typhoon Shanshan in August 2024, the input models diverged by approximately 1,000 kilometers after a 75-hour lead time, placing the center either south of Kanto or east of Hokkaido. While arithmetic averaging generated two disconnected depressions separated by an unphysical dry zone, DeepMedcast synthesized a single, coherent low-pressure system at 39.5 degrees north, 143.0 degrees east off Tohoku, accompanied by an unbroken, physically consistent moisture envelope at 850 hectopascals.
The quantitative forecasting skill of DeepMedcast was rigorously established through an extensive, year-long statistical verification against real-world observations from the Automated Meteorological Data Acquisition System, an automated weather station network operated across Japan by the Japan Meteorological Agency. Evaluating surface wind forecasts initialized four times daily throughout 2023 across lead times extending from 3 to 39 hours, the analysis verified root mean square errors for both wind speed and wind direction. DeepMedcast consistently outperformed its input models, attaining lower root mean square errors for wind speed and wind direction across every evaluated lead time compared to both the Global Spectrum Model and the Meso-Scale Model individually. Wind direction root mean square error curves displayed characteristic six-hourly diurnal fluctuations linked to initialization cycles, yet DeepMedcast maintained a persistent accuracy margin throughout the entire forecast horizon, demonstrating that the deep learning interpolation extracts genuine predictive value rather than merely smoothing field variance.
Beyond predictive skill, DeepMedcast introduces substantial maintainability advantages for operational forecast centers. Because the system functions strictly as a geometric and dynamical synthesizer rather than an empirical bias-correction scheme, models trained purely on 20-kilometer Global Spectrum Model data transferred seamlessly to updated 13-kilometer model runs and entirely different forecasting systems, including the 5-kilometer Meso-Scale Model, the 0.5-degree Integrated Forecasting System, and 0.25-degree AI platforms GraphCast and Pangu-Weather, without requiring retraining or fine-tuning. The researchers noted several operational boundaries and avenues for future enhancement, including the current network’s tendency to slightly underestimate peak cyclonic wind speeds, its equal one-to-one weighting constraint which complicates odd-numbered model ensembles, and the omission of precipitation fields. Moving forward, the development team plans to integrate multi-variable network inputs, explore transformer and diffusion architectures, incorporate variable weighting ratios, and develop dedicated intermediate post-processing modules for quantitative precipitation forecasting.
Subject of Research: Deep learning generation of intermediate meteorological forecasts among multiple numerical weather prediction models
Article Title: DeepMedcast: A Deep Learning Method for Generating Intermediate Weather Forecasts among Multiple NWP Models
Article References: KUDO, A. (2025). DeepMedcast: A Deep Learning Method for Generating Intermediate Weather Forecasts among Multiple NWP Models. Journal of the Meteorological Society of Japan. Ser. II, 103(5), Article 2025-031. https://doi.org/10.2151/jmsj.2025-031
Image Credits: AI Generated
Keywords: DeepMedcast, numerical weather prediction, deep learning, U-Net architecture, intermediate weather forecasts, ensemble post-processing, tropical cyclone forecasting, Global Spectrum Model, Meso-Scale Model, Integrated Forecasting System, GraphCast, Pangu-Weather, wind speed verification, AMeDAS observation network, meteorological front prediction, Japan Meteorological Agency, atmospheric circulation modeling, consensus weather forecasting
Cite Scienmag News
Blake Davidson. (November 20, 2026). Deep Learning Method Generates Coherent Intermediate Weather Forecasts Across Multiple Numerical Models. Scienmag. https://scienmag.com/deep-learning-method-generates-coherent-intermediate-weather-forecasts-across-multiple-numerical-models/
Blake Davidson. "Deep Learning Method Generates Coherent Intermediate Weather Forecasts Across Multiple Numerical Models." Scienmag, 20 November 2026, https://scienmag.com/deep-learning-method-generates-coherent-intermediate-weather-forecasts-across-multiple-numerical-models/. Accessed 20 November 2026.
Blake Davidson. "Deep Learning Method Generates Coherent Intermediate Weather Forecasts Across Multiple Numerical Models." Scienmag. November 20, 2026. https://scienmag.com/deep-learning-method-generates-coherent-intermediate-weather-forecasts-across-multiple-numerical-models/








