BEGIN:VCALENDAR
VERSION:2.0
PRODID:icalendar-ruby
CALSCALE:GREGORIAN
X-WR-TIMEZONE;VALUE=TEXT:Europe/Zurich
BEGIN:VEVENT
DTSTAMP:20261112T105331Z
UID:7865386a-82b6-45e0-9fe3-313202d779e5
DTSTART:20261113T140000
DTEND:20261113T150000
DESCRIPTION:Generative models such as normalizing flows and diffusion model
 s have emerged as promising tools for overcoming critical slowing down in 
 lattice field theory\, but scaling them to large lattice volumes while pre
 serving unbiased sampling remains an open problem. Annealing methods offer
  polynomial scaling and asymptotic exactness\, while diffusion models offe
 r efficient\, simulation-free training. In this talk\, I will present an a
 pproach that combines these strengths: we first train an energy-based diff
 usion model with score matching\, and then sample from it using either sto
 chastic normalizing flows (SNFs) or flow-transformed replica exchange (T-R
 EX). The choice of sampler is made after training\, with no retraining req
 uired\, and I will discuss the theoretical trade-offs between the two. I w
 ill also show how a thermodynamic view of the annealing path\, in which th
 e Fisher information defines a natural notion of distance\, leads to optim
 ized schedules that further improve sampling efficiency. Using local-footp
 rint energy models with efficient gradient and divergence calculations\, w
 e demonstrate results leveraging volume transfer from smaller to larger la
 ttices in scalar field theory and the O(3) sigma model that follow the pre
 dicted exactness and scaling behavior\, and significantly outperform sampl
 ing methods without a learned component.
LOCATION:ETH Zürich\, Hönggerberg HCP E 47.3
ORGANIZER:Octavio Vega
SUMMARY:Annealed transport and diffusion-based sampling in lattice field th
 eory
END:VEVENT
END:VCALENDAR