Uncertainty QuantificationBayesian InversionData AssimilationInverse Problems

Uncertainty Quantification & Bayesian Inversion

A collection on uncertainty quantification and inverse problems in physical settings — UQ methods, Bayesian inference and inversion, data assimilation and ensemble methods, Bayesian optimisation, plus a posterior-evaluation benchmark.

Updated Nov 3, 2026 25 resources
Introduction

Why this collection

Once physics AI enters engineering decisions, a single point prediction is not enough — you need to know how much to trust it and under which conditions it fails. Uncertainty quantification answers the first question; Bayesian inversion answers the second: given observations, recover the parameters.

The two are tightly coupled in practice. Inversion means running a surrogate through thousands of forward solves, each carrying its own error, and the Bayesian framework integrates those errors explicitly into the posterior.

The spectrum covered here:

  • UQ methods — from Gaussian processes and deep ensembles to flow matching and interpretable representative trees;
  • Bayesian inference and inversion — variational approaches, amortised posterior sampling, and large-scale engineering applications such as seismic full-waveform inversion;
  • Data assimilation and ensemble methods — fusing observations with simulations, the core technology behind weather and ocean forecasting;
  • Bayesian optimisation and design — black-box optimisation over expensive simulation, the most direct commercial outlet for surrogate models.

If this is new to you, enter through the "Benchmarks & Practical Guides" group — learn how to evaluate a posterior before comparing methods, otherwise attractive sample plots can easily mislead you.

Benchmarks & Practical Guides 3 resources

💬 Editor note: PosteriorBench is one of the few benchmarks dedicated to assessing posterior quality — filling the most glaring evaluation gap in this direction.

Search at the Cost of Sampling: Nearly-Instant Latent Space Bayesian Optimization

ScientificML
Proposes a nearly-instant latent space Bayesian optimization method that replaces expensive sequential decision-making with sampling overhead, making BO practical for generative model-driven de novo design pipelines with cheap virtual screens.
See LICENSE file Bayesian OptimizationLatent Space OptimizationGenerative Design
💬 Editor note: Rebuilds latent-space Bayesian optimisation around sampling cost, offering a direct answer to the practical "the posterior samples too slowly" bottleneck.
💬 Editor note: The tutorial demonstrates neural operators used as Bayesian-inversion surrogates — the fastest way to grasp the "surrogate plus inversion" combination.

Uncertainty Quantification Methods 8 resources

💬 Editor note: Test-time training plus flow matching for Bayesian filtering of physical systems, merging UQ with state estimation.

Beyond Point Prediction: Artificial Representative Trees with Uncertainty

ScientificML
Proposes combining Artificial Representative Trees with leaf-wise Mondrian Conformal Predictive Systems to enable a single decision tree to provide continuous predictions, prediction intervals, and probabilistic outputs as an interpretable ML approach.
See LICENSE file Decision TreesInterpretable MLUncertainty Quantification

Predictive Uncertainty for Neural CAE Surrogates

ScientificML
Investigates how uncertainty quantification approaches perform when adapted to geometry-conditioned neural surrogates for CAE, comparing closed-form and sampling-based methods across varying geometries and engineering quantities of interest.
See LICENSE file neural surrogatesuncertainty quantificationCAE simulation
💬 Editor note: A study of uncertainty quantification for neural CAE surrogates — aimed squarely at the industrial question of whether a surrogate can be trusted.

Triply-Scalable Equivariant Gaussian Process Modeling

ScientificML
Proposes a triply-scalable equivariant Gaussian process method that encodes symmetries while maintaining probabilistic predictions and uncertainty quantification for large-scale scientific computing.
See LICENSE file Gaussian ProcessEquivariant Neural NetworksUncertainty Quantification

IASC-ASCE Structural Health Monitoring Benchmark and Physics-Informed Damage Identification

Structural Health Monitoring / Damage Identification / Benchmark Dataset
The IASC-ASCE four-storey steel-frame SHM benchmark with its 12- and 120-DOF models and six damage patterns, reviewed alongside physics-informed damage identification methods, with corrections to earlier unverifiable claims.
New Various; paper copyright retained by the authors, data distributed under the task group terms Structural Health MonitoringDamage IdentificationBenchmark Dataset

Isotonic Surrogate Modeling for Computer Experiments with Many Input Variables

ScientificML
This paper develops isotonic Bayesian surrogate modeling for computer experiments involving many input variables and complex physical phenomena, including particle collisions and rocket propulsion. The method incorporates monotonicity or isotonicity information elicited from physical systems to improve surrogate modeling when the available sample size is limited, while retaining uncertainty quantification.
See LICENSE file Bayesian surrogate modelingIsotonic modelingMonotonicity constraints

Bayesian Inference & Inversion 8 resources

Emulation strategies for Bayesian inference of regional left ventricle material parameters

Biomechanics / Computational Medicine
This paper introduces a Bayesian surrogate-modeling framework for inferring regional Holzapfel-Ogden material parameters in a left ventricle partitioned into five physiological zones, using cardiac MRI data to overcome the limitation of treating the myocardium as mechanically homogeneous.
See LICENSE file Cardiac BiomechanicsBayesian InferenceSurrogate Modeling

OpenFWI: Multi-Structural Benchmark Datasets for Seismic Full Waveform Inversion

Geophysics / Inverse Problems
A seismic full waveform inversion benchmark from Los Alamos National Laboratory, with 12 datasets totalling 2.1 TB across interface, fault, style and CO2 storage structures, plus baselines for four deep learning methods and a cross-domain generalization study.
New Code is BSD 3-Clause (Copyright (c) 2022, Los Alamos National Laboratory); data is CC BY-NC-SA 4.0 GeophysicsInverse ProblemSeismic
💬 Editor note: OpenFWI is an open-source seismic full-waveform inversion implementation: large-scale and strongly coupled, a realistic yardstick for inversion stacks.
💬 Editor note: On the variational objective in amortised Bayesian inverse problems, with the focus on posterior conditioning — a key reference for understanding amortised inversion.

Data Assimilation & Ensemble Methods 5 resources

💬 Editor note: Combining data assimilation with reinforcement learning — a rare complete closed loop in the power-systems direction.
💬 Editor note: FREESIA achieves scalable data assimilation through covariance-aware posterior transport — the mainstream direction for high-dimensional problems.

PR-Smoother: Simulator-Preserving Non-Gaussian Smoothing for Data Assimilation

Geophysics
PR-Smoother is an amortized smoother for prescribed-simulator data assimilation, supporting non-Gaussian posterior estimation over physical state variables in high-dimensional simulators with calibration compatibility.
See LICENSE file Data AssimilationNon-Gaussian SmoothingAmortized Inference
💬 Editor note: A simulator-preserving non-Gaussian smoother, more reliable than Gaussian approximations under strongly nonlinear observations.

Bayesian Optimisation & Design 1 resources