NWF4CH: Neural Weather Forecasting

Approach / Solution

Forecasting precipitation in mountainous regions is challenging because precipitation arises from multiscale, nonlinear processes that global models cannot resolve at an affordable cost. Rather than re-training a regional AI weather model from scratch, SwAIther-Precip post-processes existing AIFS forecasts in two complementary steps :

1. The first step is a deterministic bias correction at coarse resolution. A U-Net takes the full multivariate AIFS forecast  (precipitation, winds, humidity, temperature, geopotential) and produces a corrected coarse precipitation field. The model is conditioned on forecast lead time through a Feature-wise Linear Modulation (FiLM) mechanism, so a single network handles all horizons from 6 hours to 6 days.
2. The second step is a generative super-resolution. A conditional diffusion model (based on NVIDIA's CorrDiff) takes the bias-corrected coarse field, together with topographic information, and generates an ensemble of realistic km-scale precipitation patterns. Because the bias correction has already absorbed lead-time dependence, this step can be trained in a perfect-prognosis setting on observations alone, keeping training costs low and the same model usable across all forecast horizons.

The pipeline was developed and evaluated using AIFS forecasts and the MeteoSwiss CombiPrecip radar–gauge product, an unusually rich combination of operational AI forecasts and high-resolution observational data.

Across all metrics, SwAIther-Precip outperforms both classical post-processing baselines and single step deep-learning alternatives. It reduces the Continuous Ranked Probability Score (CRPS) by 48% relative to raw AIFS and 11% relative to the strongest classical baseline. It also reproduces observed precipitation variance from synoptic to convective scales (effective resolution of ~4 km on a 1 km grid). Multi-horizon training also proves  to be more effective than training horizon-specific specialists, particularly at longer lead times.

Figure 1: High-resolution precipitation samples generated by the full bias correction + super resolution pipeline for our two SwAIther-Precip variants at a 2-day lead time, for a random time step in the test set.

Impact

SwAIther-Precip demonstrates that lightweight, targeted post-processing of global AI weather forecasts can deliver km-scale, probabilistic precipitation forecasts at significantly reduced cost compared to dynamic regional modeling. This is particularly relevant for operational forecasting agencies and for downstream applications such as flood early warning, hydropower management, agriculture, and infrastructure planning.

The framework is designed to be generalizable: while the present case study is focused on Switzerland, the methodology can easily be applied to other regions and to other global AI driving models, provided sufficiently high-resolution observational data is available for training.

In essence, NWF4CH illustrates the power of collaboration between data scientists, meteorological services, and academic researchers to translate the rapid progress of foundation-scale weather AI into practical tools for national-scale weather services and the broader scientific community.

‍