Vivek.portfolio

Scientific Machine Learning / Climate Downscaling

Robust Earth Forecast

Robust Earth Forecast studies how coarse ERA5 meteorological fields can be transformed into higher-resolution PRISM temperature grids. The implementation combines boundary-aware spatial reconstruction, U-Net and PixelShuffle experiments, terrain-aware residual modeling, and diagnostic evaluation, building on earlier CNN and ConvLSTM downscaling work.

Active research codebase; temporal modeling is archived while spatial reconstruction limits are being diagnosed.

Python PyTorch PyTorch Lightning xarray rasterio U-Net PixelShuffle CNN ConvLSTM Jupyter
Repository

Architecture

System view

ERA5 to PRISM downscaling project pipeline ERA5 inputs pass through preprocessing, model comparisons, terrain-aware residual reconstruction, and PRISM-scale diagnostics. ERA5 coarse fields Preprocess alignment normalization split discipline Models U-Net PixelShuffle archived ConvLSTM PRISM target fine grid Terrain residuals DEM-derived channels direct vs residual topography boundary-aware diagnostics Evaluation RMSE, border/center error gradient and high-frequency detail multi-seed checks
Conceptual diagram of coarse meteorological fields resolving into a finer regional grid. Source: venkatavivekp-debug/robust-earth-forecast

Evidence

Measured outcomes and constraints

Project metrics are kept close to their original context and paired with limitations where needed.

Residual topography RMSE

1.3858 +/- 0.0564

GitHub README reports this across seeds for the residual topography experiment.

GitHub

U-Net vs plain encoder

1.8939 vs 2.2313

Medium core4_h3 seed-42 comparison reported in the public repository.

GitHub

Historical RMSE reduction

2.37 -> 1.57

Resume-reported ConvLSTM-era result, preserved as historical evidence rather than the current active direction.

Resume

Metric Visualization

Resume-reported historical RMSE

Baseline / previous

2.37 RMSE

Model

1.57 RMSE

Approximately 33% lower RMSE; the GitHub README now frames current work around spatial detail and boundary diagnostics.

Bar lengths are scaled within this comparison only.

Workflow

How the system is organized

The workflow shows the main technical pieces and how they connect.

  1. ERA5 inputs

    Coarse atmospheric fields such as near-surface temperature, wind, and surface pressure.

  2. Geospatial alignment

    ERA5 and PRISM grids are prepared with training-only normalization and reproducible splits.

  3. Model comparisons

    Plain encoder-decoder, U-Net, padding variants, PixelShuffle, and archived CNN/ConvLSTM experiments.

  4. Terrain-aware residuals

    DEM-derived elevation, slope, aspect, and terrain-gradient channels are tested as spatial covariates.

  5. Diagnostics

    RMSE, border/center error, gradient behavior, variance ratio, high-frequency detail, and local contrast.

Problem

Regional climate and weather workflows often need spatial detail beyond what coarse reanalysis fields directly provide.

The current research question is what limits fine-scale spatial reconstruction in terrain-aware ERA5-to-PRISM downscaling.

Context / Data

ERA5 provides coarse atmospheric state. PRISM provides finer spatial temperature structure informed by station observations and terrain interpolation.

The public repository is Georgia-focused and does not claim broad geographic transfer. It treats PRISM and ERA5 as distinct products whose differences require careful interpretation.

What I Built

Earlier work compared CNN and ConvLSTM modeling for supervised downscaling and produced the documented RMSE improvement.

Current repository work diagnoses spatial reconstruction limits through controlled U-Net, padding, decoder, PixelShuffle, direct-topography, and residual-topography experiments.

The project treats metrics as experiment evidence rather than state-of-the-art claims.

Architecture

Data pipeline, dataset, model, training, evaluation, script, notebook, and research-documentation directories keep experiments organized.

Training-only normalization, repeatable seeds, and structured evaluation reduce leakage and make architecture comparisons easier to audit.

Diagnostics separate center, border, edge, corner, gradient, and high-frequency-detail behavior so the model is judged beyond aggregate RMSE.

Evaluation / Results

The README reports a medium seed-42 comparison of U-Net RMSE 1.8939 against PlainEncoderDecoder RMSE 2.2313.

Residual topography experiments report RMSE 1.3858 +/- 0.0564 and gradient ratio 0.5665, while the README also notes that fine detail remains difficult to recover.

A historical ConvLSTM-era result reports an RMSE improvement from approximately 2.37 to 1.57, about a 33% reduction.

Technical Lessons

The current bottleneck is not simply choosing a larger model; it is understanding recoverable spatial detail, boundary artifacts, and the information available in the input products.

Scientific ML work is strongest when baselines, leakage boundaries, repeatability, ablations, and limitations are visible.

A strong research system can be honest about unresolved issues while still demonstrating technical depth.

Limitations

Georgia-only scope in the public repository; no transfer claim is made.

Temporal modeling is archived while spatial behavior is being understood.

Fine-scale detail recovery remains limited, especially in the 4-8 km band reported by the README.