New York, New York, United States
Prani Nalluri is a PhD student in Applied Mathematics at Columbia University and Lamont-Doherty Earth Observatory. Their research centers on using machine learning to improve the accuracy and physical consistency of weather and climate models, with a focus on learning representations of unresolved "sub-grid" air–sea interactions for coarse-resolution simulations. They build deep learning pipelines (including CNNs, U-Nets, diffusion models, and ANNs) to predict surface flux corrections using coupled atmosphere–ocean simulation datasets. In addition to ML-based parameterization, they have experience applying machine learning to downscaling and exploring alternative parameterization approaches such as equation discovery. Their work integrates applied mathematics, physical modeling, and large-scale data engineering (cloud-based Zarr workflows) to reduce model error under computational constraints and support reliable data-driven weather and climate prediction.
Developing a machine learning-based parameterization to account for contributions from subgrid spatial heterogeneity to air-sea turbulent heat flux estimates. This parameterization has been developed using CM2.6 data and has two working versions: one with a CNN and one with an ANN. Both methods perform well across a variety of resolutions. Current work is focused on polishing these parameterizations and implementing them into the National Center for Atmospheric Research’s (NCAR) Community Earth System Model (CESM).
Developing more efficient numerical methods to computationally model fluid flows. Improving on codes to visualize Julia outputs from fluid flow problems, using packages such as Trixi.jl. Computationally developing and verifying numerical methods of quasi-1-D flow problems.
In the fall of 2020, I learned about pre- and post-processing in computational fluid dynamics research. Namely, I became familiar with both softwares that can create finite-element meshes for various geometries and visualization softwares that can properly display the results of a simulation. Using these skills, I worked with a pre-processing meshing software in order to generate an improved mesh from a deformed mesh at a given timestep.
Through the summer, I worked on modeling the behavior of light through anisotropic media from a computational standpoint. This MATLAB code simulates the physics of light as it passes through media that have variation in the prominence of a physical property based on the direction in which data recordings of the material are taken.
Under Dr. Ruby Fu, I investigated the snow metamorphism, the evolution of snow structure on a microscopic level. Snow metamorphism is key to predicting macroscopic properties of a snowpack, a large body of snow. Unfortunately, there is yet to be a unified theoretical model for snow metamorphism. To deepen scientists’ understanding of this process, my work randomly generated initial geometries of ice particles with variable location and size on a small domain to investigate how ice particle geometries evolve at different isothermal domain temperatures. To accomplish this task, I utilized a C code with finite-element code and a PETSc nonlinear solver to approximate a solution to a set of theoretical coupled partial differential equations for the system. Next, I created a MATLAB script to post-process simulation and determine geometric parameters (e.g. average radius, average curvature) of ice particles in the system. Using this data, I determined the coarsening rates of ice particles across various domain temperatures using the Lifshitz-Slyozov-Wagner theory. Through this process, I am able to add to the theory behind dry snow metamorphism and possibly advance the conceptual knowledge of scientists in the field.
Through the summer of 2021, I worked with Paul Hovland at Argonne National Laboratory. My project developed native Julia ports of test problems from the CUTEst, a collection of test problems in mathematical optimization. At the start of my internship, the CUTEst test problems were only available in Julia via an interface that generates Fortran code from a problem description, then compiles and invokes that Fortran implementation. My work developed a Julia module with a number of ports for CUTEst test problems. This module supports certain capabilities that the older Fortran-dependent infrastructure for using CUTEst problems in Julia supports; it also is able to support further capabilities, such as computing in arbitrary precision.