On March 9, Sorbonne University hosted the launch day of the AI4Science initiative within SCAI (Sorbonne Cluster for Artificial Intelligence). The event brought together researchers from multiple disciplines, institutional representatives, and industrial partners to explore how artificial intelligence is transforming scientific discovery. Beyond a mere overview of projects, the day highlighted a deeper evolution: the emergence of a new research paradigm where AI and scientific reasoning are increasingly and inextricably intertwined.
Following the success of our inaugural launch event, SCAI is proud to announce its third seminar.
The topic of the thrird session is "Trustworthy AI for Cosmology: Auditable Inference and Hybrid Neural–Numerical Solvers" - by Tristan Hoellinger, PhD at Institut d'Astrophysique de Paris.
To participate at the seminar, the registration is mandatory.
Abstract:
Ongoing and future cosmological surveys map the large-scale structure of the Universe across immense volumes and with exquisite precision. Extracting robust information from these data requires fast, high-accuracy simulations and reliable statistical inference. Machine learning offers useful means of acceleration, but the resulting pipelines remain only as reliable as the models they contain. In this talk, I present two complementary approaches to controlling model error, with the aim of enabling trustworthy inference from high-accuracy forward models of galaxy surveys.
First, I introduce a framework for diagnosing model misspecification in implicit-likelihood cosmological inference. It accommodates arbitrarily complex black-box forward models, provided that a sufficiently informative latent function is available. Using a forward model of a spectroscopic galaxy survey, I quantify how modelling errors in galaxy bias, selection functions, survey masks, redshifts, and gravitational evolution distort the reconstructed initial matter power spectrum after recombination. I further show that percent-level misspecification can shift cosmological constraints by more than 2σ in the (Ωₘ, σ₈) plane; the framework detects this failure before cosmological parameters are inferred.
Second, I present tCOCA-P3M, a hybrid neural–numerical method for accelerating high-accuracy cosmological N-body simulations. The method supplements a perturbative moving frame with a machine-learnt momentum correction and uses a P3M solver to integrate the residual dynamics; equivalently, the numerical solver corrects the emulation error.
Together, these methods provide complementary tools for reliable simulation-based inference and AI-accelerated numerical modelling.
