Onkar Jadhav

Scientific AI · Applied mathematics · Climate intelligence · Medical AI

About me

I am an applied mathematician, climate modeller and scientific machine-learning researcher specialising in uncertainty-aware generative AI for climate science and medical data. I am currently a Research Fellow at The University of Western Australia, with more than seven years of international research experience across Australia, Germany, Austria and Luxembourg.

My research includes PODiff, an uncertainty-aware generative-AI framework accepted at ICML 2026, and high-resolution sea-surface-temperature downscaling for marine heatwaves, coastal risk and climate resilience. I also develop Med-PODiff for brain-MRI reconstruction and hallucination-risk detection, with direct relevance to Australia’s health industries.

My Australian research engagement includes an invited seminar at the Australian Bureau of Meteorology and research collaboration with RMIT University. I was previously a Marie Skłodowska-Curie Fellow under EU Horizon 2020 and currently contribute to the machine-learning community as an ICML 2026 Gold Reviewer and invited reviewer for AAAI and NeurIPS. I also convened WIM2021, an international workshop on simulation and optimisation.

Selected recognition & engagement

ICML 2026
PODiff Paper accepted
ICML 2026 Gold Reviewer
Peer-review recognition
University of Western Australia
Research Fellow
Australian Bureau of Meteorology
Research engagement and invited seminar on SST downscaling
RMIT University
Climate-model downscaling research collaboration
Marie Skłodowska-Curie Fellowship
EU Horizon 2020

Current work

Structured generative modelling

I lead the development of PODiff, a conditional diffusion framework that works in a fixed Proper Orthogonal Decomposition space rather than directly on full-resolution pixels. The POD representation orders latent variables by explained variance and preserves a direct relationship between generated coefficients and physical spatial modes. This makes ensemble generation less computationally demanding and allows uncertainty to be examined in a structured, interpretable way. PODiff has been accepted at ICML 2026.

I am extending this work through Patch-PODiff-ViT. The method represents local image patches as compact orthonormal tokens and uses a Vision Transformer to model interactions between them. Because the decoder is fixed and linear, predictive uncertainty can be propagated analytically from the latent coefficients to the reconstructed physical field. The framework is being evaluated across sea-surface temperature, medical imaging and natural-image datasets.

Simplified PODiff workflow showing low-resolution conditioning, diffusion in POD coefficient space, reconstruction and uncertainty quantification
Simplified PODiff framework. Conditional diffusion operates in a compact, variance-ordered POD space before reconstruction and ensemble uncertainty quantification. View project details.
Explore PODiff Open live demo ↗ ICML paper ↗

Ocean forecasting and marine heatwaves

In parallel, I develop statistical-downscaling systems for coastal sea-surface temperature. These pipelines combine seasonal forecast ensembles, regional ocean simulations, climate-model projections and satellite observations. Residual-corrective neural networks learn the fine-scale coastal structure absent from coarse operational forecasts.

The aim is not simply to produce a sharper temperature map. Forecast ensembles are converted into daily exceedance probabilities, persistence measures and location-specific marine-heatwave information. These outputs are intended to support coastal planning, fisheries and reef management, marine operations and climate-risk assessment in Western Australia.

Comparison of sea-surface-temperature ground truth, predicted ensemble mean and spatial uncertainty
Coastal SST downscaling. Ground truth, ensemble prediction and spatial uncertainty for a Western Australian marine-heatwave case.
Comparison of analytical and empirical uncertainty estimates
Interpretable uncertainty. Analytical propagation closely follows empirical ensemble uncertainty in physical space.

Trustworthy AI across scientific domains

A further strand of the research asks how structured latent models can reveal when a reconstruction should not be trusted. Current work studies calibrated predictive intervals, scale-dependent uncertainty and indicators of hallucination or failure in medical-image super-resolution. This provides a bridge between environmental intelligence and reliable AI for other high-stakes scientific applications.

Coarse and fine decomposition of brain MRI reconstruction error and predictive uncertainty
Coarse–fine uncertainty decomposition. Total, coarse and fine components of reconstruction error and predictive uncertainty for a representative brain-MRI example. These are ongoing results rather than a published benchmark.

Research practice and collaboration

The work is conducted with collaborators across the UWA Oceans Institute, the School of Earth and Oceans, and computer science and applied mathematics. Experiments run on Setonix and NCI/Gadi using reproducible Python, PyTorch, Docker and Singularity workflows. Results are communicated through peer-reviewed papers, open preprints, research talks, visualisations and interactive demonstrations.

Explore my research