Stormscope: High-Resolution Nowcasting Model

• Stormscope is a data-driven, transformer-based generative diffusion model that predicts storm-scale weather using high-resolution satellite and radar data.
• It employs a DDPM framework with a diffusion transformer and multi-head neighborhood attention to generate nowcasts without needing classical data assimilation.
• The model demonstrates competitive forecasting skill against HRRR through ensemble techniques and robust deterministic and probabilistic evaluation metrics.

Stormscope is a data-driven, transformer-based generative diffusion model designed for explicit storm-scale weather prediction directly from high-resolution geostationary satellite imagery and ground-based weather radar. Developed to address the limitations of classical Lagrangian extrapolation and convection-allowing numerical weather prediction (NWP), Stormscope produces competitive short-range forecasts at 6 km spatial and 10 min temporal resolution, without requiring classical data assimilation infrastructure or expensive physical modeling codes (Pathak et al., 24 Jan 2026).

1. Objectives and Motivation

Stormscope’s primary objective is explicit short-term prediction (“nowcasting” and “nearcasting”) of clouds and precipitation at sub-hourly to 6-hour lead times. Traditional nowcasting (0–2 h) relies on Lagrangian extrapolation (e.g., optical-flow methods), which offer low computational demand but lose predictive skill beyond 1–2 hours. Operational mesoscale NWP (e.g., HRRR) maintains skill up to 6–12 hours, but is computationally intensive and dependent on complex, resource-heavy data assimilation pipelines. Stormscope bridges this gap by learning storm-scale evolution directly from observations, specifically using multi-band GOES satellite data and ground-based MRMS radar, exploiting their high spatiotemporal resolution. No data assimilation system is required: the model is initialized directly from observations, facilitating rapid, scalable deployment (Pathak et al., 24 Jan 2026).

2. Input Modalities

Prediction is conditioned on multi-modal input fields structured as follows:

Modality	Data Source	Channels/resolution (spatial, temporal)
Satellite	GOES-16 ABI	8 bands: 3 visible, 1 shortwave IR, 3 water vapor, 1 IR; 6 km / 10 min
Radar	MRMS composite	Composite reflectivity; 6 km / 10 min
NWP fields†	ERA5, GFS	Z500 (500 hPa geopotential), simulated reflectivity; 6 km / 1 hr

†NWP fields (Z500, GFS reflectivity) are used as additional conditioning for “nearcast” (lead times up to 6 hr) mode only.

GOES-16 Advanced Baseline Imager (ABI) data are remapped to a 3 km Lambert grid, then downsampled to 6 km. MRMS radar reflectivity is interpolated to this grid. ERA5 is used for training, GFS for inference; both fields are bilinearly interpolated as required. These inputs capture high-frequency convective evolution and mesoscale context over North America (Pathak et al., 24 Jan 2026).

3. Model Architecture and Diffusion Framework

Stormscope deploys a denoising diffusion probabilistic model (DDPM/EDM style) to estimate the conditional distribution for the next observed field $x^+$ given past observations $x^-$ and optional conditioning $c$. The generative process comprises the following steps:

• Forward process (noising):

$$q(x_t | x_{t-1}) = \mathcal{N}\left(x_t; \sqrt{1 - \beta_t} x_{t-1}, \beta_t I\right)$$

Equivalently,

$$x_t = \sqrt{\bar\alpha_t} x_0 + \sqrt{1-\bar\alpha_t} \epsilon,$$

with $\epsilon \sim \mathcal{N}(0, I)$, $\bar\alpha_t = \prod_{s=1}^t (1 - \beta_s).$

• Reverse process (denoising):

$$p_\theta(x_{t-1} | x_t) = \mathcal{N}\left(x_{t-1}; \mu_\theta(x_t, t), \Sigma_\theta(x_t, t)\right),$$

where $\mu_\theta$ and the noise predictor $\epsilon_\theta$ are parameterized by a transformer.

The core diffusion transformer operates on a $512 \times 896$ spatial grid (6 km), with each $4\times4$ region embedded as a token (yielding $128 \times 224$ tokens of 768 dimensions). Temporal history, auxiliary fields (e.g., latitude/longitude, solar-zenith angle, Z500, simulated reflectivity), and diffusion time are concatenated and embedded via strided convolution. Sixteen transformer blocks employ multi-head neighborhood attention with $31 \times 31$ windows, combining efficiency with large-scale spatial context accumulation. No explicit positional encodings are used; instead, spatial metadata is provided via input channels. The output is a direct projection onto the $512\times896$ grid, producing either the clean satellite or radar increment (Pathak et al., 24 Jan 2026).

4. Training Regimen and Loss Specification

The loss function follows the EDM-style weighted denoising objective:

$$L(\theta) = \mathbb{E}_{x, \sigma, n} \left[\lambda(\sigma) \lVert \mathcal{D}_\theta(x + n_\sigma; x^-; c; \sigma) - x \rVert_2^2 \right]$$

with $n_\sigma \sim \mathcal{N}(0, \sigma^2 I)$ and $\lambda(\sigma)$ an EDM-specific weighting.

A “multi-expert” strategy is implemented: the noise variance range $[\sigma_{\min}, \sigma_{\max}]$ is split at $\sigma_{\rm int}$, training separate “coarse” and “fine” experts on log-uniform samples within their respective ranges. Training uses Stable AdamW (initial learning rate $5 \times 10^{-4}$, cosine decay), batch size 32 on 32 H100 GPUs, and approximately 6 million iterations (~48 hours wall-time). No auxiliary losses are introduced (Pathak et al., 24 Jan 2026).

5. Forecast Generation, Autoregression, and Ensembling

• Temporal output: Nowcasting mode generates forecasts every 10 minutes, autoregressively advancing by cycling forecasts as new input. Nearcasting predicts every 1 hour with NWP-based conditioning.
• Spatial output: Each forecast is a $512 \times 896$ field at 6 km (CONUS).
• Sampling: An ODE-based deterministic sampler with stochastic perturbations (cf. Karras et al. 2022) is used, with 100 denoising steps transitioning from coarse to fine experts at $\sigma_{\rm int}$.
• Ensemble forecasts: Stochastic diffusion initialization enables ensembling. Five-member ensembles are benchmarked against a five-member lagged HRRR for CRPS; 24-member ensembles are used for reliability and object-based evaluation (Pathak et al., 24 Jan 2026).

6. Performance Benchmarks and Verification

Stormscope’s skill is evaluated on deterministic and probabilistic metrics:

• Deterministic Scores:
  • Fractions Skill Score (FSS) for composite reflectivity (20 dBZ, 30 dBZ, pooled over 18 km/42 km) and IR brightness temperature (10.35 μm channel at 235 K, 225 K).
  • Stormscope outperforms HRRR up to 3 h in FSS and remains competitive to 5–6 h.
• Probabilistic Scores:
  • Continuous Ranked Probability Score (CRPS) for rain-rate (derived from reflectivity using $Z=300R^{1.4}$), averaged across 360 initializations.
  • 5-member Stormscope ensembles surpass lagged HRRR ensembles in CRPS up to 6 h; 24-member reliability is well-calibrated with minor overconfidence at early leads (1–3 h).
• Nowcasting Metrics (<2 h):
  • Baselines: pySTEPS (Lagrangian extrapolation), HRRR.
  • Metrics: FSS (6–30 km scales), object-based probability of detection (POD), success ratio (SR), critical success index (CSI), frequency bias (FB) for storms >40 dBZ.
  • Stormscope exceeds both baselines beyond 20 min at large scales and shows uniformly higher object-based metric scores (Pathak et al., 24 Jan 2026).

7. Limitations and Prospects

Stormscope’s limitations include:

• Limited synoptic context in nearcast mode (only Z500 and simulated reflectivity used), which constrains skill at longer lead times.
• The 6 km resolution, a deliberate trade-off, may underresolve convective detail relative to 3 km or 1–2 km grids, though parallelization demonstrates feasibility at higher resolutions.
• Parallax errors at large satellite zenith angles are uncorrected.
• Lack of direct comparison to radar-only diffusion nowcasting models (e.g., NowcastNet).
• Global/oceanic and developing-world deployment is feasible owing to reliance on geostationary satellites and could democratize short-term high-resolution forecasting where radar and classical NWP assimilation infrastructure are limited or unavailable (Pathak et al., 24 Jan 2026).

Through integration of multi-modal inputs, a diffusion transformer backbone, and probabilistic ensemble generation, Stormscope enables robust, high-skill storm-scale forecasting at 10 min/6 km resolution, with performance leading deterministic and probabilistic verification metrics for short-range convective-scale meteorology (Pathak et al., 24 Jan 2026).