Ren Stengel

Ren Stengel

I am a Ph.D. student at the University of Colorado researching efficient numerical methods in the PhyPID group under Prof. Jed Brown. Applications of interest: Computational Neuroscience, biomechanics, proteomics.

Talks and Presentations

WCCM22 - Ratel - a Portable and Scalable Package for Solid and Fluid Mechanics

July 31, 2022

Talk, World Conference on Computational Mechanics (virtual), Yokohama, Japan

Ratel is a new, open-source package built on libCEED and PETSc capable of solving complex solid/fluid mechanics problems without sacrificing computational performance. This package provides a flexible yet intuitive user interface that reduces the prerequisite effort required for use, encouraging community involvement and development. Similar to libCEED and PETSc, Ratel is both performance portable and scalable allowing for effective material simulations on a variety of computing systems. Notably, Ratel uses single-source physics implementations with Just-In-Time compilations (rather than with domain specific languages or templates) and supports matrix-free high order elements. Additionally, we discuss high-performance automatic differentiation using Enzyme to simplify the development of new material models with consistent Jacobians. We discuss Ratel’s solver functionality including quasi-static and dynamic solvers and composite material support. In this presentation we also investigate Ratel’s performance on CPUs and GPUs, number of nodes, and solver degree for various physical simulations taken from Ratel’s example suite. Finally, Ratel is structured as a library to support uncertainty quantification or optimization studies.

AISRS22 - Super Resolution of Climatological Data with Generative Adversarial Networks

February 24, 2022

Talk, Carnegie Mellon, Department of Science, Pittsburgh, Pennsylvania

Accurate and high-resolution data reflecting different climate scenarios are vital for policymakers when deciding on the development of future energy resources, electrical infrastructure, transportation networks, agriculture, and many other societally important systems. However, numerical global climate models are unable to resolve the high resolution spatiotemporal characteristics required for such decisions. Here we discuss a generative adversarial neural network capable of super resolving (SR) variables such as wind velocity and solar irradiance outputs from global climate models to scales sufficient for renewable energy resource assessment. Using adversarial training to improve the physical and perceptual performance of our networks, we demonstrate up to a 50× spatial resolution enhancement of wind and solar data while maintaining physical relevance. We also extend this model to generate an ensemble of unique but physically correct SR outputs which are useful for uncertainty quantification studies. Finally, we show that a simultaneous 24x temporal SR and 10x spatial SR better captures the advection of fronts and produces more realistic wind ramp rates that are crucial for grid integration studies.

AGU 2019 - Physics-Informed Super Resolution of Climatological Wind and Solar Resource Data

December 12, 2019

Talk, Moscone Center, San Francisco California

Many aspects of modern society—including agriculture, transportation, emergency preparation, and resource planning—rely on high resolution (HR) weather and climate data. However, due the complex nature of weather and climate models, HR meteorological data is only available at local scales and even low resolution (LR) global climate models (GCM) are extremely computationally expensive. While GCM’s can forecast long term climatological trends, their resolutions are insufficient to study the impact on wind and solar energy production. In this work, we apply a deep learning image transformation technique, known as super-resolution (SR), to enhance GCM wind velocity and solar irradiance data in future climate scenarios. We propose a physics-informed variation to the super resolution generative adversarial network (SRGAN) model, which extends proven performance on super resolution of natural images to scientific datasets. Our model is trained on coarsened wind velocity data from the Wind Integration National Dataset (WIND) Toolkit, which includes a variety of meteorology data over the continental United States at a 2km resolution. The model learns the complex, nonlinear mapping from the LR input data to the associated HR output and is able to perform 50x SR in a manner that preserves the underlying turbulent flow physics in the data better than traditional SR methods. The trained model is then applied to global wind velocity data from the National Center for Atmospheric Research’s Community Climate System Model 4 (CCSM4). Our model was able to generate perceptually-realistic and physically-consistent wind velocity data fields at 2km resolutions from the original 100km resolution. Additionally, we trained a similar network on direct normal irradiance (DNI) and diffuse horizontal irradiance (DHI) data from the National Solar Radiation Database (NSRDB) to demonstrate its ability to increase CCSM DNI and DHI data from 100 km to 4km. Thus, this model has the potential to be utilized as an efficient method for enhancing coarse climate data from GCMs, enabling local energy resource assessment and grid resiliency studies to be performed for different climate scenarios.

RMFM 2019 - Physics-Informed Super-Resolution of Climatological Wind Data

July 29, 2019

Talk, University of Colorado Boulder, Boulder, Colorado

Many aspects of modern society including agriculture, transportation, emergency preparation, and resource planning rely on high resolution (HR) weather and climate data. However, due the complex nature of weather and climate models, HR meteorological data is only available at local scales and even low resolution (LR) global climate models (GCM) are extremely computationally expensive. In this work, we apply a deep learning image transformation technique, known as super-resolution (SR), to enhance GCM wind velocity data. Our model, based off of the SRGAN model, is trained on coarsened wind velocity data from the Wind Integration National Dataset (WIND) Toolkit. Wind resource planning would greatly benefit from having GCM data at a comparable resolution to the WIND Toolkit. The model successfully increased the spatial resolution of the data 50x while preserving the underlying physics. When applied to the global wind velocity trained model was able to generate perceptually-realistic and physically-consistent wind velocity data fields at 2km resolutions from the original 100km resolution that preserve a number of turbulent flow physics like energy spectra and velocity gradients.