Learning Landscapes

par Bruno De Luca (Lapth)

jeudi 5 nov. 2026, 11:00 → 12:00 Europe/Paris

Auditorium Vivargent (LAPTh)

Our universe is undergoing accelerated expansion, but describing it
within a quantum theory of gravity is a notoriously hard problem. In
string/M-theory, this amounts to finding a geometry for the extra
dimensions that solves the six- or seven-dimensional Einstein-like
equations of the theory. The common approach relies on very special
geometries, Calabi-Yau manifolds, where supersymmetry guarantees a
solution without needing to find the metric explicitly, at the price of
starting with zero cosmological constant. In the first part of this
talk, I will discuss a different paradigm: using machine learning to
deal directly with these equations and explore the landscape of
extra-dimensional geometries much more broadly, focusing on a concrete
proposal in M-theory based on negatively curved extra dimensions.

A crucial part of machine learning algorithms is the optimization
method. In the second part of the talk, I will turn the problem around,
and show how intuition from physical landscapes and dynamics allows us
to design more controllable optimization algorithms. In a physics
language, standard optimizers are friction dominated. If we conserve
energy, instead, Liouville’s theorem predicts where the dynamics spends
its time, making the outcome of the optimization more calculable and
tunable. This leads to a class of optimizers we named Energy Conserving
Descent (ECD), and to a related sampling algorithm, Microcanonical
Hamiltonian Monte Carlo (MCHMC). I will conclude by completing the loop,
discussing applications of these physics-inspired algorithms to a
variety of physics problems.