Bio

Hello, my name is Francesco Pio Ramunno, and I am a PhD student at the University of Geneva (Switzerland), in collaboration with the University of Applied Sciences Northwestern Switzerland in Brugg/Windisch. My research bridges Deep Learning and Solar Physics, with a particular focus on diffusion-based generative models, video forecasting, and spatio-temporal prediction using real-world data from the Sun.

While my current work is rooted in solar physics—applying and benchmarking state-of-the-art architectures such as DDPMs and latent diffusion models—I am deeply interested in the theoretical foundations of these models. I enjoy exploring how their modular design, interpretability, and generative capacity can be extended to other scientific domains. My broader research interests include video understanding, semantic segmentation, scene understanding, image generation and restoration, and multi-modal AI. Ultimately, my goal is to develop AI tools that are not only effective for physical modeling but also broadly applicable and scientifically interpretable across disciplines.

In this blog, I share updates on my research, including slides, code, and publications that reflect both the scientific challenges and the methodological innovation behind my work. Feel free to reach out via email if you’re interested in connecting or collaborating!

Work Experience

Polymathic AI (Simons Foundation) — AI Foundation Models for Science Intern

At Polymathic AI, I’m working on building foundation models tailored to solar physics. I develop and train advanced diffusion-based models for image and video prediction tasks on solar data from the NASA SDO mission. My main focus is ensuring these models are not just accurate visually, but scientifically interpretable — transforming raw predictions back into the physical domain to aid real-world research on magnetic field evolution.

European Space Agency (ESA) — Data Science and Machine Learning Intern

During this internship, I developed a machine learning pipeline to study Near-Earth Objects (NEOs). The work included building a structured database and implementing models to predict the physical composition of asteroids, doubling the accuracy compared to classical methods — a satisfying example of how ML can enhance planetary science.

University of Rome Tor Vergata — Data Science Intern

I performed morphological analyses on X-ray data from the XMM-Newton telescope, using wavelet filtering techniques. This allowed for a >30% improvement in parameter measurement accuracy, highlighting the role of data preprocessing in astrophysical studies.