Conditional Flow Matching for Probabilistic Downscaling of Maximum 3-day Snowfall in Alaska

Douglas Brinkerhoff, Elizabeth Fischer · 2026

Read Conditional Flow Matching for Probabilistic Downscaling of Maximum 3-day Snowfall in Alaska by Douglas Brinkerhoff, Elizabeth Fischer (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-25172.txt.

Introduction

Brinkerhoff and Fischer present WxFlow, a flow-matching model for probabilistic downscaling of snowfall in Alaska. Climate models run at 50 to 100 km resolution, missing terrain-driven detail at scales of a few kilometers, while regional models cost months per scenario. WxFlow maps coarse climate output plus high-resolution terrain to calibrated ensembles of fine-scale fields. Applied to 4 km simulations of maximum 3-day snowfall over southeast Alaska, it improved the spectral match by 87.8 percent over bicubic downscaling. It generated 50-member ensembles in seconds on a laptop.

Catalog record

CREATORDouglas Brinkerhoff, Elizabeth Fischer
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE17582
LICENSEhttp://creativecommons.org/licenses/by-sa/4.0/
RIGHTSCC BY SA
SOURCEarXiv

Source: https://arxiv.org/abs/2604.25172

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Conditional Flow Matching for Probabilistic Downscaling of Maximum 3-day Snowfall in Alaska Douglas Brinkerhoff, Elizabeth Fischer Abstract Precipitation in complex terrain is governed by orographic processes operating at scales of a few kilometers, yet climate models typically run at resolutions of 50–100 km where this topographic detail is absent. Dynamical downscaling with high-resolution regional models such as WRF can resolve these processes, but the computational cost—months of wall-clock time per scenario—precludes the large ensembles needed for uncertainty quantification. We present WxFlow, a conditional generative model based on flow matching that learns to map coarse-resolution climate model output and high-resolution topography to calibrated probabilistic ensembles of fine-scale precipitation fields. Applied to 4 km WRF simulations of maximum 3-day snowfall over southeast Alaska, WxFlow achieves 87.8% improvement in spectral fidelity and dramatically lower Continuous Ranked Probability Scores relative to conventional lapse-rate-corrected bicubic downscaling, while generating 50-member ensembles in seconds on a laptop. Ensemble spread is spatially coherent and governed by topography, reflecting physically plausible uncertainty structure. All code is available at https://github.com/glide-ism/wrf-flow . 1 Problem and Overview Mountain precipitation is a critical input for climate adaptation planning, water resource management, and avalanche hazard assessment. However, climate models operate at \sim 100 km resolution and fail to represent orographic precipitation in complex terrain. Dynamical downscaling via regional climate models such as the Weather Research and Forecasting (WRF) model ( 10 ) can capture these fine-scale processes, but the computational cost—often requiring months of wall-clock time per climate scenario—severely limits the number of scenarios that can be evaluated in decision-support applications. We develop WxFlow, a conditional generative model that learns to map coarse-resolution climate model output and high-resolution topographic information to ensembles of plausible high-resolution precipitation fields. WxFlow uses conditional flow matching ( 6 , CFM;) , a framework for training generative models by learning velocity fields that interpolate between noise and data distributions. By training a neural network velocity field via conditional flow matching, we enable rapid generation of probabilistic ensembles of downscaled precipitation in seconds rather than months, thereby enabling uncertainty quantification for climate adaptation at computationally tractable cost. 2 Approach Figure 1: Reference map showing regions in Alaska where WxFlow was applied. Tiles correspond to locations shown in Figs. 2–5 To demonstrate its efficacy, we applied WxFlow to the regions of coastal Alaska shown in Fig. 1 . Beginning with WRF simulation outputs at 4 km horizontal resolution ( 5 ) , we coarsened data to simulate 64 km resolution climate model output. We partitioned the temporal domain into 120 time steps for training and 30+ independent time steps for evaluation. We then used standard flow matching (Appendix A ) to create a generative model of these high-resolution outputs conditioned on topography and low-resolution climate. The velocity network has approximately 23M parameters and operates on 64\times 64 pixel tiles at 4 km resolution ( 256\times 256 km). The dataset comprises maximum 3-day snowfall accumulations from multiple atmospheric forcings: CFSR reanalysis (1981–2010) for the historical baseline and GFDL CM3 and CCSM4 climate models (2031–2060) for future projections. Initial results, which took a few seconds to generate on a laptop computer with NVIDIA RTX4070, are shown in Figure 2 . Figure 2: Samples drawn from the CFM-based distribution of 3 day maximum snowfall. The hi-res WRF output (a–c) was coarsened to produce a low-res input for WxFlow (g–i). WxFlow combined the elevation map and low-res input to produce a high-res output (d–f), which shares many qualitative and spectral properties with the physics model. 3 Ensemble Statistics and Probabilistic Predictions Figure 3: Samples drawn from WxFlow for tile B (Malaspina Glacier area) with ensemble mean removedr (a–c) alongside the low-resolution forcing (d) and ensemble mean (e). Deviations from the mean are spatially coherent, reflecting topographic effects such as rain-shadowing and orographic enhancement. Figure 4: Uncertainty quantification for the St. Elias mountain range as quantified by sample standard deviation, centered on tile B (Malaspina Glacier). By generating multiple independent realizations with different noise samples, we obtain an ensemble that characterizes the conditional probability distribution over the downscaling problem (Figure 3 ). The spatial pattern of ensemble spread provides diagnostic information about which regions have high predictability (low spread) and which regions have intrinsic unpredictability relative to the coarse forcing (high spread). In particular, WxFlow produces spatially coherent samples in which precipitation anomalies on lee and stoss sides of mountain ridges are correlated across the ridge, reflecting physically plausible rain-shadow effects. Ensembles can also be characterized by their summary statistics (Fig. 4 ), which show that regions of extreme topography and precipitation have greater standard deviation than lower, flatter regions. 4 Spatial Forecast Skill Figure 5: CRPS computed with WxFlow (a, f, k) and the lapse-rate corrected bicubic baseline method (b, g, l) over Tiles a–c in Fig. 1 drawn from the test set, alongside the WRF ground truth (c, h, m), ensemble mean (d, i, n), and baseline mean (e, j, o). Fields are overlaid on high-resolution topography. WxFlow exhibits better performance than the baseline with respect to CRPS, particularly over high-relief regions. To evaluate WxFlow, we compare it to the baseline bicubic lapse rate method, in which coarse climate data are refined via lapse rate correction. We used Continuous Ranked Probability Score ( 3 , CRPS;) to evaluate each of the two methods against the original WRF output. CRPS is a proper scoring rule that measures the calibration of probabilistic predictions, penalizing both forecast bias and overconfidence while balancing distance to the observation (reliability) against ensemble spread (resolution). A lower CRPS score indicates better forecast skill. Results (Figure 5 ) show that WxFlow dramatically improves accurate placement of precipitation in Southeast Alaska, where orographic precipitation effects are important. 5 Power Spectral Density and Spatial Structure Fidelity Figure 6: Log-log plot of Power Spectral Density for WxFlow and the baseline bicubic algorithm. WxFlow shows a marked improvement over the baseline, with a little more challenge at high frequencies. This is consistent with most generative models, which tend to have an easier time learning smooth features. While CRPS assesses point-wise forecast calibration, the spatial structure of downscaled precipitation is equally important for applications in hydrology and hazard assessment. We evaluate spatial structure fidelity via the radially-averaged two-dimensional power spectral density (PSD), which decomposes variance as a function of spatial wavelength. The PSD reveals a scale-dependent variance cascade (Figure 6 ). Precipitation fields that are over-smoothed exhibit a deficit of power at small scales (short wavelengths \lambda<8 km), while fields with spurious noise show a spectral excess. The WxFlow ensemble span (one standard deviation across 50 members) reveals the uncertainty inherent in the probabilistic prediction at each scale—narrow bands indicate confident predictions, while wide bands indicate high model uncertainty. WxFlow Bicubic baseline Mean abs. log spectral error 0.141 ( \times 1.38) 1.152 ( \times 14.18) Spectral bias (signed) - 0.065 ( \times 0.86) - 1.148 ( \times 0.07) Table 1: Spectral evaluation metrics (lower magnitude is better). WxFlow achieves 87.8% improvement in spectral fidelity over the baseline. Multiplicative factors in parentheses indicate the ratio of model power to truth power implied by the log-space error. We also evaluate the PSD via two scalar metrics: mean absolute log spectral error and spectral bias (Table 1 ). Under both,

Read the complete text