Climate Downscaling with Stochastic Interpolants (CDSI)

Erik Larsson, Ramon Fuentes-Franco, Mikhail Ivanov, Fredrik Lindsten · 2026

Read Climate Downscaling with Stochastic Interpolants (CDSI) by Erik Larsson, Ramon Fuentes-Franco, Mikhail Ivanov, Fredrik Lindsten (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-2603-03838.txt.

Introduction

Erik Larsson and others introduce Climate Downscaling with Stochastic Interpolants, or CDSI, a data-driven method for turning coarse global climate output into fine local projections. Global climate projections come from Earth System Models, which are costly to run and so work at coarse spatial resolution. Regional Climate Models refine them for areas of interest, but they are expensive too and limit how many runs can be done at once.

The authors build the method on stochastic interpolants, recent chance-based machine learning tools. They show that CDSI converts coarse model output into accurate detailed regional forecasts at a fraction of the cost of standard regional models. Extensive checking produces accurate regional ensembles, which sharpens uncertainty estimates and widens access to detailed climate information for regional impact work.

Catalog record

CREATORErik Larsson, Ramon Fuentes-Franco, Mikhail Ivanov, Fredrik Lindsten
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42402
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.03838

Browse the full library · About the Tower of Babel

Full text

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

Climate Downscaling with Stochastic Interpolants (CDSI) Erik Larsson, Ramon Fuentes-Franco, Mikhail Ivanov, Fredrik Lindsten Abstract Global climate projections rely on computationally demanding Earth System Models ( ESM s), which are typically limited to coarse spatial resolutions due to their high cost. To obtain high-resolution projections for regions of interest, it is common to use Regional Climate Models ( RCM s), which are driven by data produced by ESM s as boundary conditions. While more efficient than running ESM s at fine resolution, RCM s remain expensive and restrict the size of ensemble simulations. Inspired by recent advances in probabilistic machine learning for weather and climate, we introduce a data-driven climate downscaling method based on stochastic interpolants. Our approach efficiently transforms coarse ESM output into high-resolution regional climate projections at a fraction of the computational cost of traditional RCM s. Through extensive validation, we demonstrate that our method generates accurate regional ensembles, enabling both improved uncertainty quantification and broader use of high-resolution climate information. † † affiliationnotice: Equal contribution 1 Introduction As climate change intensifies, there is a growing need for high-resolution regional climate information to assess local impacts and extreme events. Global climate projections from Earth System Models ( ESM s), such as those in CMIP6, typically operate at coarse spatial resolutions (around 100\text{\,}\mathrm{km} ) due to computational constraints ( 10 ) . Such coarse grids cannot resolve critical mesoscale processes (e.g., convective storms, complex topography) and thus cannot capture fine-scale variability of climate extremes at scales of 1– 10\text{\,}\mathrm{km} ( 10 ; 25 ) . This limitation necessitates downscaling to produce regional projections fit for local impact assessments. The Coordinated Regional Downscaling Experiment ( CORDEX ) addresses this gap by coordinating regional climate modeling at higher resolutions. The European branch, EURO- CORDEX , has produced ensembles of regional climate simulations at 0.1°(around 12\text{\,}\mathrm{km} ) resolution. These high-resolution simulations are widely used by scientists and stakeholders, forming a critical basis for impact assessments and adaptation planning. Finer spatial detail is essential for simulating climate extremes; for example, high-resolution precipitation data are needed to assess flood risks ( 25 ) . By resolving topography and coastlines, RCM downscaling can reduce biases in regional precipitation and temperature ( 25 ) . Importantly, RCM s produce physically consistent fields at fine scales ( 10 ) , which is crucial for analyzing multi-variable extreme events coherently. Despite these benefits, RCM s are computationally intensive, limiting the number of scenarios or ensemble members that can be downscaled ( 10 ) . A single RCM run spanning decades at 0.1°requires orders of magnitude more computing resources than one at 50\text{\,}\mathrm{km} ( 25 ) . Our main contributions are: 1. We propose Climate Downscaling with Stochastic Interpolants (CDSI) , a multivariate probabilistic data-driven downscaling framework specifically developed to downscale global low-resolution climate simulations to high-resolution regional climate simulations. 2. We show that CDSI can directly downscale ESM outputs to high-resolution regional projections, producing ensembles with performance competitive to state-of-the-art diffusion-based approaches while avoiding the multi-model, sequential pipelines used by competitive baseline methods. 3. We demonstrate that CDSI can generalize to unseen future conditions and across new climate model realizations, indicating robustness under distribution shift. 2 Background In this study, we use the HARMONIE-Climate ( HCLIM ) regional climate modeling system as our reference dataset, serving as both the ground truth for validation and the training target for our emulator. HCLIM is a community-developed, high-resolution modeling framework designed for regional climate simulations across a wide range of spatial scales. Built upon the HIRLAM–ALADIN Numerial Weather Prediction system, it integrates advanced physical parameterizations, data assimilation techniques, and surface modeling via the SURFEX land model. The system includes multiple configurations optimized for different resolutions: AROME for convection-permitting ( \sim 2.5\text{\,}\mathrm{km} ), ALARO for intermediate ( \sim 410\text{\,}\mathrm{km} ), and ALADIN for coarser ( \sim 1050\text{\,}\mathrm{km} ) grids. In this work, we employ simulations from the HCLIM38 configuration using ALADIN physics at 12\text{\,}\mathrm{km} resolution over the EURO- CORDEX domain. HCLIM has been extensively validated and applied in coordinated international projects such as EURO- CORDEX and CORDEX FPS, and is recognized for its ability to simulate temperature and precipitation variability across Europe with high spatial fidelity. Given its robustness, process realism, and widespread adoption in the regional modeling community, HCLIM provides an ideal benchmark for evaluating downscaling approaches like CDSI . Problem Definition. In this work, we address the problem of downscaling, where the goal is to recover a high-quality (HQ) field from a corresponding low-quality (LQ) input. Specifically, we aim to learn a mapping from LQ to HQ that can enhance spatial resolution and correct systematic errors or biases present in the low-quality data. In addition to LQ, auxiliary static features such as land–sea masks and topography, as well as precomputed dynamic features like time of day or year, may be incorporated to provide additional information to the model. Downscaling vs. Simulation. An alternative approach to obtaining HQ fields from LQ inputs is to perform regional simulation ( 13 ; 18 ; 17 ; 19 ; 1 ; 12 ) . For regional simulations, the task differs from downscaling. Instead of directly learning a mapping from LQ to HQ, the HQ fields are generated through simulation, where the model evolves trajectories within the HQ domain driven by the LQ dynamics and previous HQ states. This approach offers the advantage of producing temporally consistent trajectories, where each HQ ensemble member naturally evolves from its predecessor. However, it also involves the computational cost of simulating full trajectories, making it impractical for long-term applications such as climate prediction, where simulations may span on the order of a hundred years. In addition, maintaining stable and physically plausible model behavior over long roll-outs can be challenging for data-driven regional simulation approaches, as small errors may accumulate over time. This approach is also impractical when the objective is to generate high-resolution outputs at specific time instances, as it requires unnecessary simulation of the entire temporal evolution. In contrast, downscaling allows HQ fields to be inferred directly from LQ inputs at any desired time point without requiring the simulation of full trajectories. For this reason, we focus on downscaling as it provides greater flexibility and computational efficiency, particularly for large-scale or long-horizon climate applications. Downscaling vs. Super-resolution. Superresolution is a well-studied problem in computer vision, where the objective typically is to recover a high-resolution image from a low-resolution observation, typically by inverting a known or implicit degradation such as blurring or subsampling ( 7 ) . Downscaling in geophysical and climate applications differs fundamentally from classical super-resolution. The low-quality input is not obtained by a simple or known degradation operator acting on the high-quality field. Instead, it often originates from a different numerical model, resolution, or parameterization, and therefore exhibits systematic biases, structural errors, and mismatches in physical variability. As a result, downscaling entails not only increasing spatial resolution but also correcting model-dependent biases and inconsistencies. Furthermore, when multiple variables are downscaled simultaneously, the model must learn complex cross-variable dependencies while respecting the differing spatial scales, spectra, and statistical properties of each field, making downscaling a substantially more challenging problem than standard image super-resolution. Consequently, downscaling cannot be treated as a straightforward extension of image super-resolution, but rather

Read the complete text