Aaron Hill

Machine Learning for High-Impact Weather

Postprocessing numerical weather prediction (NWP) model output to improve point-based and localized forecasts has been a topic of research for over 50 years (e.g., MOS). Statistical models are used to deduce relationships between pre-defined predictors and the phenomena of interest. Machine learning (ML) has emerged within atmospheric science as a new method to postprocess large NWP datasets to explicitly predict events that are unresolvable by NWP model output (e.g., extreme precipitation, tornadoes). One benefit of ML models is they are capable at learning the underlying spatio-temporal biases in NWP model output, thereby "correcting" biases in ML-based forecasts. I am interested in applying ML and deep learning (DL) techniques with global convection-parameterizing and convection-allowing models to create probablistic hazard outlooks (e.g., Hill et al. 2020a, Hill and Schumacher 2021, Schumacher et al. 2021, Hill et al. 2023) that are beneficial to operational forecast centers (e.g., Storm Prediction Center, Weather Prediction Center). This work is being supported by two grants from the National Oceanic and Atmospheric Administration Joint Technology Transfer Initiative program.

Predictability of Convection

Predictability of convective processes is easily discerned through the application of Ensemble Sensitivity Analysis (ESA), which relates changes in model initial conditions to changes in a forecast metric through univariate regressions. ESA is capable of diagnosing dynamic regions and features that are critical to successful forecasts of convection initiation and intensity (e.g., Hill et al. 2016). I continue to employ ESA to examine dynamic sensitivies of convection forecasts with convection-allowing models (e.g., Hill et al. 2016, Hill et al. 2021) and am interested in exploring how sensitivities vary across convection regimes and climates.

Targeted Observing

Observations of small-scale atmospheric prcesses become critical for convection-allowing model forecasts when convective processes are explicitly simulated. However, the selection of additional observations to improve forecasts is often subjective, even though objective methods exist to select observations in dynamically relevant regions. Ensemble sensitivity analysis (ESA) provides a framework to determine additional observing locations that dynamically relate earlier state variables (e.g., 2-m temperature) and a particular forecast metric (e.g., reflectivity) to reduce forecast uncertainty after the new observation is assimilated. I have examined ESA-based targeting at convection-allowing resolutions in idealized simulations (e.g., Hill et al. 2020b) and continue to explore its utility for real-data cases and storm-scale applications, for instance, targeting in-storm observations with Unmanned Aerial Systems and assimilating them into CM1 model simulations.

Data Assimilation

Data assimilation procedures, which combine first-guess four-dimensional snap shots of the atmosphere with observations, have evolved significantly over the last two decades to improve forecasts at all scales of motion. I am particularly interested in the ensemble Kalman filter (EnKF), which has a broad application at convective scales, and methodological parameters that control inflation, localization, and ensemble spread, which help to create reliable ensemble forecasts. Specifically, as a component of the VORTEX-SE project, I am investigating the influence of near-surface thermodynamic observations on deep convection in the southeast U.S. through retrospective data assimilation experiments with the High Resolution Rapid Refresh Ensemble (e.g., Hill et al. 2021).

Severe Storm Dynamics

I have been involved in a number of field programs that aimed to better understand the dynamics of severe storms, either through targeted deployments of observational assets in and near supercell thunderstorms (e.g., Targeted Observations by Radars and UAS of Supercells, TORUS) or deployments of mobile and portable near-surface platforms to observe heterogenieties preceeding convection in the southeast U.S. (e.g., VORTEX-SE). These experiences have propelled an interest in engineering platforms designed to observe severe storms (e.g., small, lightweight platforms) as well as the observations obtained that may elucidate new dynamic information to help advance warnings of severe-storm hazards. These interests are being supported by a National Oceanic and Atmospheric Administration grant through the VORTEX-SE Program.

Medium-Range Severe Weather Forecasts with ML

Improved Excessive Rainfall Forecasts into Extended Ranges

Tornadogenesis in Quasi-linear Convective Systems

How does storm predictability change in warming climates?

Severe Storm Dynamics

Cold-pool dynamics, supercell dynamics, tornadogenesis, tornado sturcture, radar remote sensing, microsphyics, electrification, single-cell thunderstorms, multi-cell thunderstorms

Mesoscale Meteorology

Parcel theory, fluid dynamics, atmospheric waves, dryline dynamics, frontogenesis, stability, shallow and deep convective processes, mesoscale convective systems, sounding analyses, low-level jets

Atmospheric Modeling and Data Assimilation

Discretization and finite differencing, governing dynamic equations, parameterization schemes, deterministic and ensemble forecasting, verification, data assimilation, chaos theory, predictability, WRF model, initial condition problems

Data Analytics and Numerical Methods

Least squares problem, maximum likelihood estimation, statistical filtering, objective analysis, coordinate transformations, random forests, neural networks, convolutional neural networks, advanced machine learning/deep learning methods

Introduction to Atmospheric Sciences

Energy cycle, atmospheric forces, vertical structure of the atmosphere, atmospheric composition, stability, thermodynamics and clouds, air masses, cyclones, climate cycles, thunderstorms, tornadoes, forecasting

Scientific Computing in the Atmospheric Sciences

Linux, bash scripting, Python, version control (git, GitHub), object-based programming, meteorological data packages