Title:	Scalable Differentiable Monte Carlo for Probabilistic Inference and Learning
DNr:	Berzelius-2025-434
Project Type:	LiU Berzelius
Principal Investigator:	Niklas Wahlström <niklas.wahlstrom@it.uu.se>
Affiliation:	Uppsala universitet
Duration:	2025-12-15 – 2026-07-01
Classification:	10210
Keywords:

Abstract

This proposal requests GPU resources on Berzelius to support research in a WASP-funded PhD project on scalable Monte Carlo methods for probabilistic inference and learning. In particular, the project targets (i) accelerating gradient-based MCMC for high-dimensional scientific inverse problems, with a focus on uncertainty quantification in stellar magnetic field reconstruction from spectropolarimetric data, (ii) developing a new pathwise-differentiable resampling method for sequential Monte Carlo (particle filtering) based on an ensemble score diffusion model, and (iii) differentiable Bayesian fusion for gradient-based optimization of expectations under fusion posteriors constructed from subposteriors in distributed or large-scale data settings. The Python library JAX will be used to improve performance in high dimensions via GPU acceleration. Due to the high dimensionality and computational demands of the models and algorithms, the requested GPU resources are a prerequisite for successful completion of the project.