Research map / Image generation / Fast diffusion sampling

Tyche: One Step Flow for Efficient Probabilistic Weather Forecasting

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Guidance for conditional diffusion · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Tyche is a one-step (1-NFE) conditional flow model for probabilistic global weather forecasting that learns a destination-aware average velocity, then finetunes with autoregressive CRPS curriculum to calibrate ensembles.

Problem

Probabilistic weather ensemble forecasting with diffusion/flow models is expensive because inference cost scales with rollout length, ensemble size, and the number of denoising steps per transition.

Approach

Adapts MeanFlow-style interval-averaged velocity learning with a JVP-based rectified target to a conditional geophysical setting, paired with an isotropic Swin transformer with axial RoPE, and a two-stage training (one-step transport pretraining then curriculum CRPS finetuning on autoregressive ensemble rollouts) enabling 1-NFE ensemble generation.

Key results

• Table 1: Tyche has lowest 3-day RMSE on u10, t2m, mslp (2.13, 1.68, 2.49) versus GraphCast (2.21, 2.28, 2.86).
• Tyche generates a 20-member 15-day trajectory in about 62-65 seconds with 1 NFE, reported as 12-45x faster than probabilistic baselines.
• Ablation: at 1-NFE Tyche gets 0.103 Step-1 RMSE, matching the 10-NFE EDM result (0.108 Step-1 RMSE) with one tenth of the sampling budget.

Surprising findings

• GraphCast and Pangu outperform the probabilistic models at short (1-day) lead times, indicating deterministic models retain an advantage early in forecasts.

Limitations

• High GPU memory overhead of the model at inference.
• Stage II curriculum is short (only two rollout horizons, R=1 and 2), so limited iterative finetuning was explored.
• Baselines like GenCast, NeuralGCM-ENS and IFS-ENS are taken from saved WeatherBench2 outputs rather than retrained, limiting controlled comparison.

Experimental setup

Datasets: ERA5 WeatherBench2

Benchmarks: ERA5 2020 test split (WeatherBench2)

Metrics: latitude-weighted RMSE CRPS SSR ensemble spread

Baselines: Pangu GraphCast ARCI Swift ArchesWeatherGen GenCast NeuralGCM-ENS IFS-ENS EDM TrigFlow sCM

Builds on: MeanFlow flow matching Swin Transformer GenCast CRPS

Closest papers by meaning

• Continuous Ensemble Weather Forecasting with Diffusion models (ICLR 2025 · similarity 0.88)
• Probabilistic Weather Forecasting with Hierarchical Graph Neural Networks (NeurIPS 2024 · similarity 0.86)
• Probabilistic Forecasting with Stochastic Interpolants and Föllmer Processes (ICML 2024 · similarity 0.86)
• Probablistic Emulation of a Global Climate Model with Spherical DYffusion (NeurIPS 2024 · similarity 0.86)
• Flow Matching with Gaussian Process Priors for Probabilistic Time Series Forecasting (ICLR 2025 · similarity 0.85)
• OmniCast: A Masked Latent Diffusion Model for Weather Forecasting Across Time Scales (NeurIPS 2025 · similarity 0.85)