• Latent-variable graph neural network. A per-step Gaussian latent on the coarsest mesh level injects stochasticity into a hierarchical encode-process-decode backbone. The model only requires one forward pass per ensemble member.
• K-means cluster meshes. Latitude weighted spherical K-means produces a mesh that conforms better to ocean grids by construction compared to previously used quadrilateral or icosahedral meshes.
The ocean is inherently chaotic, yet existing data-driven ocean models produce deterministic forecasts. In our new preprint, we introduce Njord, a probabilistic graph neural network for ensemble ocean forecasting.
Link: arxiv.org/abs/2605.15470
A couple highlights below 🧵
• Sea ice is predicted alongside other physical state variables. Smooth invertible activation functions together with a binary density channel keep ice variables within realistic bounds.
• Lastly, animation of an extra long sea ice concentration rollout over the spring of 2024 for the Baltic Sea, with reanalysis atmospheric forcing. The regional model remains surprisingly stable while only trained with 2 autoregressive steps.
• Njord is trained globally at 0.25° and on the Baltic Sea at 2km resolution. In the regional setting Njord conditions on boundary data from an independent global ocean model, where previous emulators either lack boundary forcing or depend on the very system they aim to replace.
This award, supported by ICMBE2026, recognizes exceptional contributions to molecular beam epitaxy by researchers under 40. Prof. Ahmadi is a deserving honoree and we couldn't be prouder. 💛💙
#UCSB #MBE #WideBandgap #PowerElectronics #QuantumDevices #MaterialsScience
