Vivek.portfolio

Current Research

University of Georgia - GAIM Lab

Deep Learning for Statistical Downscaling

Current work at UGA's GAIM Lab focuses on deep-learning systems that map coarse atmospheric data into higher-resolution weather fields.

Scientific ML Geospatial AI Weather modeling Statistical downscaling PyTorch Hydra

Research Workflow

From coarse fields to evaluated reconstruction

Conceptual statistical downscaling pipeline Coarse meteorological inputs and terrain context pass through spatial and temporal modeling to produce a high-resolution reconstruction for evaluation. Inputs ERA5 + terrain Modeling U-Net / temporal paths Target PRISM-scale grid Evaluation loop RMSE border / center error gradient behavior fine-detail retention Diagnostics compare aggregate error with spatial behavior.
A high-level view of weather-data preparation, model training, and spatial diagnostics.

Problem Context

Downscaling is a modeling problem and an evaluation problem

Coarse atmospheric fields can contain useful signals, but regional use often depends on recovering finer spatial structure.

Meteorological models and reanalysis products describe the atmosphere on grids that are often too coarse for regional analysis.

Statistical downscaling asks whether models can reconstruct finer spatial structure while respecting geography, data-product differences, and validation limits.

Research Pipeline

Experiment infrastructure is part of the research

The workflow connects weather data handling, model configuration, checkpointed training, diagnostics, and interpretation.

  1. Weather inputs

    Prepare gridded atmospheric data such as ERA5, PRISM, HRRR, and severe-weather variables when available.

  2. Geospatial preparation

    Align grids, generate manifests, preserve normalization boundaries, and maintain repeatable train/validation workflows.

  3. Deep models

    Extend and compare CNN, ConvLSTM, U-Net, Enhanced U-Net, and PixelShuffle-style reconstruction models.

  4. Diagnostics

    Evaluate aggregate error alongside border behavior, gradients, local contrast, and high-frequency detail.

  5. Interpretation

    Compare outputs against baselines and report limitations before treating a result as a research claim.

Research Themes

Themes that shape the work

Statistical Downscaling

Mapping coarse weather fields into higher-resolution spatial reconstructions with clear validation boundaries.

Geospatial Machine Learning

Working with gridded weather data, terrain features, and spatial error patterns as part of the model design.

Scientific Deep Learning

Building neural architectures and experiment workflows that make model comparisons repeatable.

Spatial Reconstruction

Diagnosing blur, border artifacts, gradient behavior, and high-frequency detail recovery beyond aggregate RMSE.

Severe-Weather Modeling

Exploring high-resolution weather variables and model infrastructure relevant to future severe-weather workflows.

Reproducible ML Systems

Using configuration, normalization boundaries, checkpoints, tests, and evaluation scripts as part of the research workflow.

Modeling Approaches

Previous work, current framework, and active directions

Completed modeling work informs the current GAIM Lab framework, where new architecture and data workflows are under active development.

Previous Modeling Work

CNN ConvLSTM ERA5-to-PRISM temperature workflow

Current Framework Work

U-Net Enhanced U-Net PixelShuffle Hydra configuration PyTorch Lightning checkpoints HRRR preparation pytest regression tests

Active Research Directions

diffusion-style modeling spatial diagnostics severe-weather reconstruction

Scientific Tooling

Research tooling as experimental infrastructure

The work depends on practical tools for model configuration, weather-data preparation, training recovery, and reproducible evaluation.

Modeling

PyTorch PyTorch Lightning Enhanced U-Net CNN ConvLSTM U-Net PixelShuffle Diffusion-style modeling

Experiment Management

Hydra Checkpoint resume pytest regression tests Structured evaluation

Scientific Data

xarray Herbie NumPy Pandas rasterio NetCDF

Geospatial / Weather

HRRR ERA5 PRISM Severe-weather data

Research Questions

Questions driving the next iteration

These questions shape current implementation and evaluation work.

  1. How much fine-scale spatial detail is recoverable from coarse meteorological inputs and terrain covariates?
  2. Which decoder and padding choices reduce boundary artifacts without simply improving aggregate RMSE?
  3. How should temporal histories be reintroduced once spatial reconstruction limits are better understood?
  4. What experiment infrastructure makes weather-model architecture comparisons reproducible enough to audit?
View Robust Earth Forecast View Publications