1 of 12

Can Machine Learning Produce Climate Invariant Downscaling?

 

1

12/16/2025

2 of 12

High-Res Climate Data Importance

2

12/16/2025

3 of 12

Why ML(hybrid) Downscaling?

  • Model for Prediction Across Scales - Atmosphere (Skamarock et al., 2012)

​

  • Run 139x simulations of 36-hour events:
    • Low-Res: 240-48 km (500 CPU hours) ; 60-3 km (240,000 CPU hours)

3

12/16/2025

4 of 12

Till when would the ML-learnt downscaling functions still work?

  • Dn-SPAL (Saha & Ravela, 2022)

​

​

​

​

​

​

  • CorrDiff (Mardani et al., 2023)

4

12/16/2025

5 of 12

Are Downscaling Functions Variant with Climate?

  • Can we apply historical downscaling functions to future?
  • Will future-data trained downscaling functions look different than historical downscaling?
  • When to retrain downscaling functions?

5

12/16/2025

Historical Low-Res

Future Low-Res

Historical High-Res

Future High-Res

6 of 12

99% Extreme Precip Days from 2 Climate Scenarios

6

12/16/2025

2001-2021

2041-2061

7 of 12

Eigenspectra as a spatial variability metric

7

12/16/2025

8 of 12

Dn-SPAL Framework

  • Gaussian Process
  • Upslope Model
  • GAN

8

12/16/2025

9 of 12

Downscale historical testing data�with downscaling function trained on historical data

9

12/16/2025

10 of 12

Downscale historical testing data�with downscaling function trained on historical data (mid-left), �vs. with downscaling function trained on future data (mid-right).

10

12/16/2025

11 of 12

Eigenspectra results:�same scenario downscaling functions (green-solid) may not be statistically different than cross-climate downscaling (green-dashed)

11

12/16/2025

12 of 12

Takeaway

  • Puerto Rican extreme precipitation events identified from climate models are re-simulated with MPAS with different resolutions.

​

  • Climate invariance is a general question to all downscaling functions trained on current scenario (ERA5). A physics-ML hybrid downscaling framework (Dn-SPAL) is used in this study.

​

  • Downscaling function changes are of an order smaller than the climate shift for extreme precipitation in Puerto Rico.

​

  • Email x_yan@mit.edu for questions and more!

​

12

12/16/2025