- author = "Wassim Tenachi",
- title = "Symbolic Machine Learning for Physics \& Astrophysics",
- school = "Observatoire Astronomique de Strasbourg, Universite de Strasbourg",
- year = "2024",
- address = "France",
- month = "16 " # sep,
- keywords = "genetic algorithms, genetic programming, Physical Symbolic Optimization, physo, P-SO, Typhon, symbolic machine learning, deep reinforcement learning, symbolic regression, dark matter, stellar streams, Milky Way, apprentissage automatique symbolique, apprentissage profond par renforcement, regression symbolique, matiere noire, courants stellaires, Voie Lactee, SINDy, CDM, streamfinder, massfinder",
-
URL = "
https://publication-theses.unistra.fr/public/theses_doctorat/2024/TENACHI_Wassim_2024_ED182.pdf",
-
code_url = "
https://github.com/WassimTenachi/PhySO",
- size = "348 pages",
-
abstract = "We explore the transformative potential of symbolic
machine learning in physics and astrophysics, seeking
to overcome the interpretability challenges of
traditional methods in the era of data abundance. We
introduce ph-SO, a Physical Symbolic Optimization
framework that relies on deep reinforcement learning to
extract analytical symbolic expressions directly from
data. This symbolic regression (SR) framework achieves
state-of-the-art performance by integrating physical
dimensional analysis and enabling the exploitation of
diverse realisation of a singular class of phenomena;
an approach we dub Class SR.
Focusing on the dark matter challenges at the galactic scale, we uncover several new stellar streams from Gaia satellite data and perform follow-up observations using the INT and VLT telescopes. Notably, we discover a polar stream from the outer halo passing through the Solar neighborhood, which we dub Typhon. Finally, we propose a first observation-driven, unsupervised learning approach to agnostically constrain the dark matter distribution of the Milky Way from a snapshot of stellar coordinates using canonical transformations",
-
notes = "In english
Version: January 18, 2025, revision 1.2
(UMR 7550)
Supervisor: Rodrigo Ibata",
