Research · Colin Grudzien

Research

Research profile links

Short research description

Data assimilation (DA) refers to techniques used to combine the data from physics-based, numerical models and real-world observations to produce an estimate for the state of a time-evolving random process and the parameters that govern its evolution. Owing to their history in numerical weather prediction, DA systems are designed to operate in an extremely large dimension of model variables and observations, often with sequential-in-time observational data. As a long-studied “big-data” problem, DA has benefited from the fusion of a variety of techniques, including methods from Bayesian inference, dynamical systems, numerical analysis, optimization, control theory and machine learning. I am developing data assimilation methodology for the prediction of atmospheric rivers and other extreme precipitation events in the Western USA.

Please find my computational resources and pre-prints linked below by subject area.

Ongoing and past projects

DataAssimilationBenchmarks.jl

DataAssimilationBenchmarks.jl is my personal data assimilation benchmark research code with an emphasis on testing and validation of ensemble-based filters and sequential smoothers in chaotic toy models. The code is meant to be performant, in the sense that large hyper-parameter discretizations can be explored to determine structural sensitivity and reliability of results across different experimental regimes, with parallel implementations in Slurm. This includes code for developing and testing data assimilation schemes in the L96-s model currently, with further models in development.

Example code for the algorithms developed in our project and toy data can be found in my github Electric Grid Code Repository . Visualizations are linked below.

Visualizations

The following visualizations demonstrate the force directed layout of the fully reduced power grid network model, where nodes represent repelling bodies and edges represent springs. We randomly initialize the positions for the triangle reduced network and run the positions to a stabilized configuration. Random initial positions are shown in the first figure and a stabilized result is shown in the second. Finally, we overlay the degree two reduction (performed before the triangle reductions) on the positions for the stabilized triangle reduction. Enabling the physics on the third figure demonstrates the declustering of the degree two reduced network from the triangle reduced network, as shown in the manuscript. These visualizations are best viewed in Chrome.