About

Dale Durran is an Adjunct Professor of Atmospheric Sciences at the University of Washington, affiliated with the Department of Atmospheric Sciences. His fields of interest include atmospheric dynamics and predictability, machine learning, and numerical methods. He is involved in research related to atmospheric sciences, contributing to the understanding of atmospheric behavior and the development of computational techniques in the field.

Research topics

• Political Science
• Computer Science
• Artificial Intelligence
• Business
• Law
• Earth science
• Environmental ethics
• World Wide Web
• Geology

Selected publications

• Kilometer-scale convection-allowing model emulation using generative diffusion modeling

  Science Advances · 2026-01-30

  articleOpen access

  Storm-scale convection-allowing models (CAMs) explicitly resolve convective dynamics within the atmosphere to predict the evolution of thunderstorms and mesoscale convective systems that result in damaging extreme weather. Deep learning models have, thus far, not proven skillful in this regime of kilometer-scale atmospheric simulation, despite being competitive at coarser resolutions with state-of-the-art global, medium-range weather forecasting. We present a generative diffusion model called StormCast, which emulates the High-Resolution Rapid Refresh (HRRR)-National Oceanic and Atmospheric Administration's state-of-the-art 3-kilometer operational CAM. StormCast autoregressively predicts 99 state variables at the kilometer scale using a 1-hour time step, with dense vertical resolution in the atmospheric boundary layer, conditioned on 26 synoptic variables. We show successfully learned kilometer-scale dynamics including competitive 1- to 6-hour forecast skill for composite radar reflectivity alongside physically realistic convective cluster evolution, moist updrafts, and cold pool morphology. These results present opportunities for improving kilometer-scale regional ML weather prediction and future climate hazard dynamical downscaling.

  PublisherDOI

• Imposing the Fundamental Dynamical Constraint of Hydrostatic Balance to Improve Global ML Weather Prediction

  ArXiv.org · 2025-06-09

  preprintOpen access

  Forecasting weather accurately and efficiently is a critical capability in our ability to adapt to climate change. Data driven approaches to this problem have enjoyed much success recently providing forecasts with accuracy comparable to physics based numerical prediction models but at significantly reduced computational expense. However, these models typically do not incorporate any physics priors. In this work, we demonstrate improved skill of data driven weather prediction approaches by incorporating physical constraints, specifically in the context of the DLWP model (Karlbauer et. al. 2024). Near hydrostatic balance, between the vertical pressure gradient and gravity, is one of the most fundamental and well satisfied constraints on atmospheric motions. We impose this balance through both hard and soft constraints, and demonstrate that the soft constraint improves the RMSE of many forecast fields, particularly at lead times beyond 7-10 days. The positive influence of hydrostatic balance is also clearly evident in improving the physicality and strength of a 10-day forecast for hurricane Irma. These results show that adding appropriate physical constraints can improve the skill and fidelity of data driven weather models in a way that does not impose any significant additional memory capacity or scalability challenges.

  PublisherOA PDFDOI

• A Practical Probabilistic Benchmark for AI Weather Models

  Geophysical Research Letters · 2025-04-08 · 17 citations

  articleOpen access

  Abstract Since the weather is chaotic, it is necessary to forecast an ensemble of future states. Recently, multiple AI weather models have emerged claiming breakthroughs in deterministic skill. Unfortunately, it is hard to fairly compare ensembles of AI forecasts because variations in ensembling methodology become confounding and the baseline data volume is immense. We address this by scoring lagged initial condition ensembles—whereby an ensemble can be constructed from a library of deterministic hindcasts. This allows the first parameter‐free intercomparison of leading AI weather models' probabilistic skill against an operational baseline. Lagged ensembles of the two leading AI weather models, GraphCast and Pangu, perform similarly even though the former outperforms the latter in deterministic scoring. These results are elaborated upon by sensitivity tests showing that commonly used multiple time‐step loss functions damage ensemble calibration.

  PublisherOA PDFDOI

• A Deep Learning Earth System Model for Efficient Simulation of the Observed Climate

  AGU Advances · 2025-08-01 · 7 citations

  articleOpen access

  Abstract A key challenge for computationally intensive state‐of‐the‐art Earth System models is to distinguish global warming signals from interannual variability. Here we introduce Deep Learning Earth System Model (DL ESy M), a parsimonious deep learning model that accurately simulates the Earth's current climate over 1000‐year periods with minimal smoothing and no drift. DL ESy M simulations equal or exceed key metrics of seasonal and interannual variability—such as tropical cyclogenesis over the range of observed intensities, the cycle of the Indian Summer monsoon, and the climatology of mid‐latitude blocking events—when compared to historical simulations from four leading models from the sixth Climate Model Intercomparison Project. DL ESy M, trained on both historical reanalysis data and satellite observations, is an accurate, highly efficient model of the coupled Earth system, empowering long‐range sub‐seasonal and seasonal forecasts while using a fraction of the energy and computational time required by traditional models.

  PublisherOA PDFDOI

• Long-Range Distillation: Distilling 10,000 Years of Simulated Climate into Long Timestep AI Weather Models

  arXiv (Cornell University) · 2025-12-28

  preprintOpen access

  Accurate long-range weather forecasting remains a major challenge for AI models, both because errors accumulate over autoregressive rollouts and because reanalysis datasets used for training offer a limited sample of the slow modes of climate variability underpinning predictability. Most AI weather models are autoregressive, producing short lead forecasts that must be repeatedly applied to reach subseasonal-to-seasonal (S2S) or seasonal lead times, often resulting in instability and calibration issues. Long-timestep probabilistic models that generate long-range forecasts in a single step offer an attractive alternative, but training on the 40-year reanalysis record leads to overfitting, suggesting orders of magnitude more training data are required. We introduce long-range distillation, a method that trains a long-timestep probabilistic "student" model to forecast directly at long-range using a huge synthetic training dataset generated by a short-timestep autoregressive "teacher" model. Using the Deep Learning Earth System Model (DLESyM) as the teacher, we generate over 10,000 years of simulated climate to train distilled student models for forecasting across a range of timescales. In perfect-model experiments, the distilled models outperform climatology and approach the skill of their autoregressive teacher while replacing hundreds of autoregressive steps with a single timestep. In the real world, they achieve S2S forecast skill comparable to the ECMWF ensemble forecast after ERA5 fine-tuning. The skill of our distilled models scales with increasing synthetic training data, even when that data is orders of magnitude larger than ERA5. This represents the first demonstration that AI-generated synthetic training data can be used to scale long-range forecast skill.

  PublisherDOI

• Long-Range Distillation: Distilling 10,000 Years of Simulated Climate into Long Timestep AI Weather Models

  ArXiv.org · 2025-12-28

  articleOpen access

  Accurate long-range weather forecasting remains a major challenge for AI models, both because errors accumulate over autoregressive rollouts and because reanalysis datasets used for training offer a limited sample of the slow modes of climate variability underpinning predictability. Most AI weather models are autoregressive, producing short lead forecasts that must be repeatedly applied to reach subseasonal-to-seasonal (S2S) or seasonal lead times, often resulting in instability and calibration issues. Long-timestep probabilistic models that generate long-range forecasts in a single step offer an attractive alternative, but training on the 40-year reanalysis record leads to overfitting, suggesting orders of magnitude more training data are required. We introduce long-range distillation, a method that trains a long-timestep probabilistic "student" model to forecast directly at long-range using a huge synthetic training dataset generated by a short-timestep autoregressive "teacher" model. Using the Deep Learning Earth System Model (DLESyM) as the teacher, we generate over 10,000 years of simulated climate to train distilled student models for forecasting across a range of timescales. In perfect-model experiments, the distilled models outperform climatology and approach the skill of their autoregressive teacher while replacing hundreds of autoregressive steps with a single timestep. In the real world, they achieve S2S forecast skill comparable to the ECMWF ensemble forecast after ERA5 fine-tuning. The skill of our distilled models scales with increasing synthetic training data, even when that data is orders of magnitude larger than ERA5. This represents the first demonstration that AI-generated synthetic training data can be used to scale long-range forecast skill.

  PublisherOA PDF

• Earth, Wind, and Fire: Are Boulder’s Extreme Downslope Winds Changing?

  Bulletin of the American Meteorological Society · 2025-03-12

  articleOpen access

  Abstract A Denver newspaper in 2016 reported that a new Colorado all-time record peak wind gust of 148 mph was recorded on 18 February 2016, on Monarch Pass in the Colorado Rockies near 11 000 ft above sea level. The article stated that this broke the previous record of 147 mph set on 25 January 1971 at the National Science Foundation (NSF) National Center for Atmospheric Research (NCAR) Mesa Laboratory, at an altitude of 6077 ft, on the western edge of Boulder, Colorado. Though there is no actual official peak gust record in Colorado, this raised the issue that Boulder had not recently experienced winds of the magnitude of the megadownslope windstorms that wracked the area in the 1960s, 1970s, and 1980s when extreme wind gusts recorded at the NSF NCAR Mesa Laboratory were not unusual. Due to Boulder’s location at the eastern foot of a north–south mountain range ( Earth ), it is susceptible to destructive downslope winds ( wind ) often accompanied by fires ( fire ) such as the downslope wind-driven Marshall Fire just east of Boulder on 30 December 2021 that destroyed nearly 1100 homes. But after the 1990s, the weather station anemometer at NSF NCAR did not record a peak gust much over 100 mph. What changed? This detective story describes the search for causes of the apparent decrease in strength of extreme windstorms at NSF NCAR and their impacts in the Boulder area. The suspects in Boulder include a change in instrument location, changes in building codes, and increasing roughness length from tree growth. But climate change emerges as a chief culprit. Significance Statement National Science Foundation (NSF) National Center for Atmospheric Research (NCAR) was at the epicenter of megadownslope windstorms that wracked Boulder in the 1960s, 1970s, and 1980s when extreme windstorms were not unusual. But after the 1990s, the weather station anemometer at NSF NCAR, which replaced the previous anemometer that recorded the huge gusts, did not record a peak gust much over 100 mph. What changed? This detective story describes the search for causes of the apparent decrease in strength of extreme winds at NSF NCAR and their impacts in the Boulder area. Changing instrument location is part of the story, but climate change emerges as a key culprit.

  PublisherOA PDFDOI

• Advancing Parsimonious Deep Learning Weather Prediction Using the HEALPix Mesh

  Journal of Advances in Modeling Earth Systems · 2024-08-01 · 20 citations

  articleOpen access

  Abstract We present a parsimonious deep learning weather prediction model to forecast seven atmospheric variables with 3‐hr time resolution for up to 1‐year lead times on a 110‐km global mesh using the Hierarchical Equal Area isoLatitude Pixelization (HEALPix). In comparison to state‐of‐the‐art (SOTA) machine learning (ML) weather forecast models, such as Pangu‐Weather and GraphCast, our DLWP‐HPX model uses coarser resolution and far fewer prognostic variables. Yet, at 1‐week lead times, its skill is only about 1 day behind both SOTA ML forecast models and the SOTA numerical weather prediction model from the European Center for Medium‐Range Weather Forecasts. We report several improvements in model design, including switching from the cubed sphere to the HEALPix mesh, inverting the channel depth of the U‐Net, and introducing gated recurrent units (GRU) on each level of the U‐Net hierarchy. The consistent east‐west orientation of all cells on the HEALPix mesh facilitates the development of location‐invariant convolution kernels that successfully propagate weather patterns across the globe without requiring separate kernels for the polar and equatorial faces of the cube sphere. Without any loss of spectral power after the first 2 days, the model can be unrolled autoregressively for hundreds of steps into the future to generate realistic states of the atmosphere that respect seasonal trends, as showcased in 1‐year simulations.

  PublisherDOI

• Can Observation Targeting Be a Wild Goose Chase? An Adjoint-Sensitivity Study of a U.S. East Coast Cyclone Forecast Bust

  Journal of the Atmospheric Sciences · 2024-12-23

  articleSenior author

  Abstract Efforts to improve midlatitude-cyclone forecasts by deploying supplemental observations in localized target regions often fall short of expectations. We consider a potential contributing factor to these underwhelming results by investigating the initial-condition sensitivity of the 15 November 2018 East Coast cyclone forecast bust. We use a moist adjoint model to compute the initial-condition perturbations that minimize the large 48–72-h synoptic-scale forecast errors associated with this storm. The adjoint-optimal perturbations, which have maximum amplitudes of about 2 K in temperature and 1 m s −1 in horizontal wind speed, are widespread, extending throughout the troposphere and along a ridge–trough pattern covering much of North America. We investigate the most impactful components of the perturbations by truncating them in physical and spectral space and rescaling them to be equal in a domain-integrated energy norm to the full, unmodified perturbations. When the perturbations are confined to a localized target region of strongest sensitivity, they have weaker impacts on the forecast than when the perturbations within the target region are removed and the rest of the perturbations are retained. Additionally, when the perturbations are filtered to retain only wavelengths longer than 1000 km, they have stronger impacts on the forecast than when the perturbations are filtered to retain only wavelengths shorter than 1000 km. These results suggest that midlatitude-cyclone forecast improvements from targeted observations can be overwhelmed by smaller-amplitude but widespread and large-scale initial-condition sensitivities outside of the target region. Significance Statement Poor forecasts of midlatitude cyclones can cause tremendous socioeconomic disruption via unexpected heavy precipitation and damaging winds. One approach to improving these forecasts involves targeting observations in localized regions where initial-condition errors are expected to be most harmful to forecast accuracy. These efforts are expensive, yet they typically produce only minor forecast improvements. By examining a recent poorly forecast midlatitude cyclone, we find that a potential contributing factor to these underwhelming results is that small, but widespread changes to the initial state can be more impactful than the big, but localized changes that targeting is designed to make. This suggests that efforts to reduce initial-condition errors over broad areas can be more economical for improving midlatitude-cyclone forecasts than targeted observations.

  PublisherDOI

• Kilometer-Scale Convection Allowing Model Emulation using Generative Diffusion Modeling

  arXiv (Cornell University) · 2024-08-20 · 7 citations

  preprintOpen access

  Storm-scale convection-allowing models (CAMs) are an important tool for predicting the evolution of thunderstorms and mesoscale convective systems that result in damaging extreme weather. By explicitly resolving convective dynamics within the atmosphere they afford meteorologists the nuance needed to provide outlook on hazard. Deep learning models have thus far not proven skilful at km-scale atmospheric simulation, despite being competitive at coarser resolution with state-of-the-art global, medium-range weather forecasting. We present a generative diffusion model called StormCast, which emulates the high-resolution rapid refresh (HRRR) model-NOAA's state-of-the-art 3km operational CAM. StormCast autoregressively predicts 99 state variables at km scale using a 1-hour time step, with dense vertical resolution in the atmospheric boundary layer, conditioned on 26 synoptic variables. We present evidence of successfully learnt km-scale dynamics including competitive 1-6 hour forecast skill for composite radar reflectivity alongside physically realistic convective cluster evolution, moist updrafts, and cold pool morphology. StormCast predictions maintain realistic power spectra for multiple predicted variables across multi-hour forecasts. Together, these results establish the potential for autoregressive ML to emulate CAMs -- opening up new km-scale frontiers for regional ML weather prediction and future climate hazard dynamical downscaling.

  PublisherOA PDFDOI

Recent grants

• Troposphere-Stratosphere Coupling Processes

  NSF · $438k · 2002–2007

• The Co-Evolution of Mesoscale Airflow over Mountains and the Larger Scale Flow

  NSF · $562k · 2005–2009

• Mesoscale Airflow over Mountains: Wave Drag and Orographic Precipitation

  NSF · $714k · 2011–2016

• Mountain-Wave Evolution and Orographic Precipitation

  NSF · $733k · 2016–2019

• Collaborative Research: Efficient Unstructured Discontinuous Galerkin Methods for Global Nonhydrostatic Atmospheric Modeling

  NSF · $358k · 2012–2017

Frequent coauthors

• Russ S. Schumacher

  Colorado State University

  25 shared

• Jeffrey D. Kelley

  National Centre for Atmospheric Science

  25 shared

• David M. Schultz

  University of Manchester

  25 shared

• Gerard H. Roe

  University of Washington

  17 shared

• Jonathan A. Weyn

  14 shared

• Craig C. Epifanio

  Texas A&M University

  12 shared

• Thomas P. Ackerman

  University of North Carolina School of the Arts

  11 shared

• Rajul Pandya

  11 shared

Education

• Ph.D., Mathematics

  University of Washington

  1984

• M.S., Mathematics

  University of Washington

  1980

• B.S., Mathematics

  University of California, Berkeley

  1977