Drifting Fields are not Conservative
Drift fields in generative models are generally non-conservative; sharp normalization makes them gradients of a scalar potential, bridging to Wasserstein gradient flows and denoising score matching.
Tim Weiland
PhD student working on scalable probabilistic ML for spatiotemporal inference and simulation, in the Methods of Machine Learning group at the University of Tübingen.
NEWS
Talk on structured Bayesian inference with Latte.jl at Compute! Paris on November 26.
Drifting Fields are not Conservative got accepted to NeurIPS 2026!
GP-FVM at ProbNum 2026 in Lappeenranta.
Hierarchical Bayesian Quadrature at UAI 2026 in Amsterdam.
Talk on Bayesian spatial modeling in Julia at JuliaCon 2026 in Mainz.
PUBLICATIONS
Scaling up Probabilistic PDE Simulators with Structured Volumetric Information
Introduces GP-FVM, a probabilistic counterpart to the finite volume method, enabling large-scale probabilistic PDE simulations.
Hierarchical Bayesian Quadrature
Bayesian quadrature for nonstationary integrands, through an adaptive tree of local stationary models.
Scalable Bayesian Inference for Nonlinear Conservation Laws
Numerically conservative Bayesian simulation of nonlinear conservation laws, scaled up with sparse approximations.
Flexible and Efficient Probabilistic PDE Solvers through Gaussian Markov Random Fields
Fast probabilistic counterparts to numerical PDE solvers, powered by SPDEs & sparse linear algebra.
PROJECTS
Latte.jl
Probabilistic programming for latent Gaussian models
SelectedInversion.jl
Quickly compute selected entries of the inverse of a sparse matrix
GaussianMarkovRandomFields.jl
Fast, flexible and user-centered Julia package for Bayesian inference with sparse Gaussians
Scaling GP-based PDE Solvers
MSc Research Project. A case study on the shallow-water equations.
AI Exhibition Object on Decision Trees
Teaching people about machine learning by predicting apartment rents with decision trees.