Onkar

Research Fellow at UWA | Scientific ML | Diffusion Models | Uncertainty Quantification

[Onkar @_Onkara]

@henrytdowling Yes, this is a challenge for most ML downscaling methods. The nice part about PODiff is that the POD modes are interpretable, which makes it easier to monitor mode drift and adapt the representation over time.

[Onkar @_Onkara]

Our paper “PODiff” has been accepted at ICML 2026.

PODiff introduces diffusion in Proper Orthogonal Decomposition space for efficient and uncertainty aware scientific super resolution.

Preprint: https://t.co/w6FcWg83xJ

#ICML2026 #MachineLearning #DiffusionModels #ScientificML