Towards Scientific Foundation Models and How They Could Change ML In Survey Astronomy
Date: Monday, 23 March 2026
Time: 14:00 - 15:00
Location: AULA Auditorium, Astronomy Dept., Observatoire de Genève, Versoix
Speaker: François Lanusse (CEA Saclay)
Location: AULA Auditorium, Astronomy Dept., Observatoire de Genève, Versoix
Speaker: François Lanusse (CEA Saclay)
(Also transmitted on zoom)
Title: Towards Scientific Foundation Models and How They Could Change ML In Survey Astronomy
Deep Learning has seen a recent shift in paradigm, from training specialized models on dedicated datasets, so-called Foundation Models, trained in a self-supervised manner on vast amounts of data and then adapted to solve specific tasks with state-of-the-art performance. This new paradigm has been exceptionally successful not only for large language models (LLMs) but in other domains such as vision models. However applications of this new approach in scientific application domains are still very scarce, for reasons ranging from the need for new architectures to the (surprising) lack of availability of suitable large scale datasets.
In this talk, I will present the efforts of the Polymathic AI initiative to bring this new foundation model paradigm across broad scientific domains, from computational fluid dynamics to survey astronomy. I will discuss our efforts to bring together domain scientists to assemble large-scale and wide-ranging training datasets of scientific data, as well as building models and training strategies that can benefit from large-scale training compute.
I will specifically cover applications to survey astronomy, where we have deployed at scale multimodal self-supervised generative pretraining techniques. I will show how these approaches can be used to build models flexible to very diverse and inhomogeneous observations (e.g. different types of measurements such as spectra, time-series, or images, but also different instruments, etc...) and how they can then be leveraged by the end-user for a variety of downstream applications (e.g. redshift estimation, morphology classification, physical parameter inference, searching for rare objects) with very simple machine learning methods and still reach near-optimal performance.
In this talk, I will present the efforts of the Polymathic AI initiative to bring this new foundation model paradigm across broad scientific domains, from computational fluid dynamics to survey astronomy. I will discuss our efforts to bring together domain scientists to assemble large-scale and wide-ranging training datasets of scientific data, as well as building models and training strategies that can benefit from large-scale training compute.
I will specifically cover applications to survey astronomy, where we have deployed at scale multimodal self-supervised generative pretraining techniques. I will show how these approaches can be used to build models flexible to very diverse and inhomogeneous observations (e.g. different types of measurements such as spectra, time-series, or images, but also different instruments, etc...) and how they can then be leveraged by the end-user for a variety of downstream applications (e.g. redshift estimation, morphology classification, physical parameter inference, searching for rare objects) with very simple machine learning methods and still reach near-optimal performance.
This seminar is part of the AISC initiative to foster exchanges between researchers who are interested in, develop, use, and study AI technologies across different fields. The seminar will be followed by a coffee break, providing an opportunity for informal discussions and interactions.