An FNRS fellowship for Sacha Lewin to develop continuous deep generative models for spatio-temporal data
Education path
From his secondary education, Sacha Lewin demonstrated a strong interest in science, mathematics, and computer science. After teaching himself programming, he decided to pursue formal studies in the field, enrolling in civil engineering studies in 2018 at the University of Liège.
Developing a deep curiosity for artificial intelligence during his studies, Sacha graduated five years later with the highest distinction as a civil engineer in data science. His curriculum was capped by a thesis on the use of neural radiance fields, an AI technique that reconstructs 3D environments from 2D images. This project was carried out during an internship at the Liège-based company EVS Broadcast Equipment and supervised by Professor Gilles Louppe, with the aim of reconstructing scenes from football matches. His work earned him the best thesis awards from both AIM (Association of Montefiore Engineers) and AILg (Association of Engineering Graduates from the University of Liège), as well as the Toyota award.
In 2023, Sacha Lewin began a PhD in artificial intelligence at the University of Liège, once again under the supervision of Professor Gilles Louppe, first as an assistant and later receiving an FNRS scholarship the following year to focus more fully on his research.
Continuous generative models for spatiotemporal data
The artificial intelligence lab of Professor Gilles Louppe focuses on applying machine learning to scientific problems, primarily in physics. In this context, Sacha's research aims to develop new AI methods to advance data-driven modeling of natural phenomena that occur in space and time. Traditional AI methods are poorly suited to modeling continuous processes and learning from irregular, multi-resolution, noisy, and incomplete observations—challenges frequently encountered in scientific applications.
Sacha’s thesis aims to address these limitations through the use of so-called continuous AI methods, similar to those used in his master's thesis, to better capture the properties and handle these complex, unique datasets. By solving these problems, he hopes to contribute to the development of models capable of processing much larger amounts of data, replicating the success of models in other domains, such as the famous large language models (e.g., ChatGPT), but for scientific data.
During his first year, Sacha conducted a literature review, completed his first publication based on his thesis results, and joined a project led by his colleague Omer Rochman Sharabi, which led to a second publication. Together with their advisor Gilles Louppe, they developed a new method for simulating fluids up to twenty-five times faster than traditional techniques. Sacha is now focusing on developing continuous AI methods hoping to apply them with his team to large-scale atmospheric data.
