Gabriele Accarino, Ph.D.
Hi there! My name is Gabriele Accarino. I’m a computer engineer working at the intersection of machine learning, weather, and climate. I am currently an Associate Research Scientist at Columbia University and the NSF’s Learning the Earth with AI and Physics (LEAP) Science and Technology Center. Working alongside Professor Viviana Acquaviva (CUNY City Tech & Lamont-Doherty Earth Observatory, Columbia University) at the TerraD2I Lab, I build generative models to emulate complex physical systems.
Physical processes span a wide range of spatial and temporal scales, and my research focuses on ways to effectively capture this multi-scale variability in data-driven emulators so they explicitly operate across scales. I explore signal processing tools, particularly wavelets, as a prescribed mathematical representation to access frequency information that is localized in space and scale. I designed WaveSim, a wavelet-based metric for comparing maps in weather and climate applications; it computes similarity scores scale by scale, providing diagnostic insights inaccessible through point-wise metrics.
Building on this, I am exploring whether wavelet space can serve as a principled domain for emulation. I formalized a new generative framework called Wavelet Flow Matching (WFM) (NeurIPS 2026), in which generative emulation is performed directly on wavelet representations of physical states. WFM demonstrates superior stability and spectral consistency compared to a wide range of generative baselines, including Neural Operators, positioning wavelet space as an effective, training-free representation for emulating complex physical dynamics.
In an ongoing project, I am also using tabular diffusion models to estimate surface-ocean ฮpCOโ distributions from sparse observations (NeurIPS 2026 RPS workshop).
Before moving to New York, I led the Machine Learning Research Unit at the CMCC Foundation in Italy and taught high-performance computing, machine learning, and information processing systems as an adjunct professor at the University of Salento. At LEAP, I continue to pursue my passion for teaching machine learning and generative modeling, helping train the next generation of climate data scientists through LEAP’s bootcamps and summer schools.
news
| Oct 2026 | New chapter: I am now an Associate Research Scientist at Columbia University, still at LEAP. |
|---|---|
| 2026 | Wavelet Flow Matching accepted at NeurIPS 2026 (main track). Our ฮpCOโ diffusion paper is in at the NeurIPS Workshop on Representations for the Physical Sciences. |
| Summer 2026 | Taught Week 2 of the LEAP Summer Bootcamp on machine learning for climate modeling: neural networks for images and sequences. |
| Dec 2025 | WaveSim is on arXiv and GitHub: a multi-scale similarity metric for weather and climate fields. Poster at AGU. |
| 2025 | Our PNAS paper Transferring climate change physical knowledge won the ADIA Lab Best Paper Award in Climate Data Sciences (1st place). |
| Apr 2025 | Talk at Climate Informatics 2025 in Rio de Janeiro on perceptual metrics for Earth system models. |
| Oct 2024 | Started a postdoc at Columbia University and LEAP in New York. |