pith. sign in

arxiv: 2509.19929 · v4 · pith:ZUJM75GHnew · submitted 2025-09-24 · 📊 stat.ML · cs.LG· physics.comp-ph· physics.data-an

Geometric Autoencoder Priors for Bayesian Inversion: Learn First Observe Later

Arnaud Vadeboncoeur , Gregory Duth\'e , Mark Girolami , Eleni Chatzi This is my paper

Pith reviewed 2026-05-18 14:19 UTC · model grok-4.3

classification 📊 stat.ML cs.LGphysics.comp-phphysics.data-an
keywords Bayesian inversiongeometric autoencodersuncertainty quantificationgeometry-conditioned priorsphysical response modelingapproximate Bayesian computationinverse problemslearn first observe later
0
0 comments X p. Extension
pith:ZUJM75GH Add to your LaTeX paper What is a Pith Number? 
\usepackage{pith}
\pithnumber{ZUJM75GH}

Prints a linked pith:ZUJM75GH badge after your title and writes the identifier into PDF metadata. Compiles on arXiv with no extra files. Learn more

The pith

Geometric autoencoders learn geometry-aware priors from diverse physical datasets to enable Bayesian inversion on new systems.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces GABI, a framework where geometric autoencoders are trained on large datasets of physical responses from systems with varying geometries. The resulting latent representation serves as a prior that captures shared information across geometries without any need for PDEs or observation models during learning. At inference, this prior combines with the specific likelihood from new noisy observations to produce a tailored posterior. A sympathetic reader would care because this 'learn first, observe later' strategy addresses the ill-posedness of recovering full-field information in engineering problems with complicated variable geometries. It shows comparable accuracy to supervised methods while providing well-calibrated uncertainty estimates.

Core claim

GABI distills information from large datasets of systems with varying geometries, without requiring knowledge of governing PDEs, boundary conditions, or observation processes, into a rich latent prior that is seamlessly combined with the likelihood of a specific observation process, yielding a geometry-adapted posterior distribution.

What carries the argument

The geometric autoencoder that encodes physical responses from varying geometries into a shared latent space to serve as a generative, geometry-conditioned prior for Bayesian inversion.

If this is right

  • Predictive accuracy becomes comparable to deterministic supervised learning in settings where the latter applies.
  • Uncertainty quantification stays well calibrated and robust on problems with complex geometries.
  • Information sharing across multiple distinct physical systems reduces ill-posedness without relying on standard multi-system Bayesian UQ techniques.
  • An architecture-agnostic design permits efficient GPU-based implementation through Approximate Bayesian Computation sampling.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same learned latent space could transfer information between related physical phenomena such as heat transfer and fluid flow if the training collection spans those domains.
  • Inverse problems could proceed in regimes where governing equations are unknown or intractable by relying entirely on the data-driven prior.
  • Online updating of the latent representation becomes feasible if new observations are folded back into the autoencoder after initial training.

Load-bearing premise

Large and representative datasets of physical responses across varying geometries exist and suffice for the autoencoder to learn priors that generalize to unseen observation processes and geometries.

What would settle it

Apply GABI to a held-out geometry and a new observation process, then check whether the posterior mean recovers the true field more accurately than a standard non-geometry-aware prior and whether the uncertainty intervals contain ground truth at the nominal rate.

Figures

Figures reproduced from arXiv: 2509.19929 by Arnaud Vadeboncoeur, Eleni Chatzi, Gregory Duth\'e, Mark Girolami.

Figure 1
Figure 1. Figure 1: Four out of 1k geometries and solutions in dataset for the steady-state heat problem. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (a) Four selected query locations for the sampled predictive solutions given data, the full [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of ground truth (GT), inferred mean, error, and standard deviation for pressure [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ground truth, inferred mean, stddev., and error for ampltitude [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Ground truth, inferred mean, error, and standard deviation for pressure and the magnitude [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: For the heat setup, we overlay in red the observation locations. (a) ground truth (b) decoded [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: (a) Lose functions during training – the total loss is the sum of these. (b) histogram of the [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Sampling schemes on the same autoencoder model with the same channel space and different [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Four samples from the Airfoil dataset for pressure, horizontal and vertical velocity fields. [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Comparison of ground truth (GT), inferred mean, error, and standard deviation for pressure [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Airfoil field samples drawn from the geometry conditioned joint prior over pressure, [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Three example meshes in dataset (a) u, i = 1 (b) u, i = 2 (c) u, i = 3 (d) f, i = 1 (e) f, i = 2 (f) f, i = 3 [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Given a geometry, we show 3 samples from the conditional prior. [PITH_FULL_IMAGE:figures/full_fig_p021_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Four examples of flows (p, vx, vy, vz) over complex terrain taken from the training set. geometry-conditional autoencoder in the form of a VAE θ ⋆ , ψ⋆ = arg min θ,ψ EDy KL  q θ z|yn [PITH_FULL_IMAGE:figures/full_fig_p022_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Flow field samples for a given test terrain drawn from the geometry conditioned joint prior [PITH_FULL_IMAGE:figures/full_fig_p023_15.png] view at source ↗
read the original abstract

Uncertainty Quantification (UQ) is paramount for inference in engineering. A common inference task is to recover full-field information of physical systems from a small number of noisy observations, a usually highly ill-posed problem. Sharing information from multiple distinct yet related physical systems can alleviate this ill-posedness. Critically, engineering systems often have complicated variable geometries prohibiting the use of standard multi-system Bayesian UQ. In this work, we introduce Geometric Autoencoders for Bayesian Inversion (GABI), a framework for learning geometry-aware generative models of physical responses that serve as highly informative geometry-conditioned priors for Bayesian inversion. Following a ''learn first, observe later'' paradigm, GABI distills information from large datasets of systems with varying geometries, without requiring knowledge of governing PDEs, boundary conditions, or observation processes, into a rich latent prior. At inference time, this prior is seamlessly combined with the likelihood of a specific observation process, yielding a geometry-adapted posterior distribution. Our proposed framework is architecture-agnostic. A creative use of Approximate Bayesian Computation (ABC) sampling yields an efficient implementation that utilizes modern GPU hardware. We test our method on: steady-state heat over rectangular domains; Reynolds-Averaged Navier-Stokes (RANS) flow around airfoils; Helmholtz resonance and source localization on 3D car bodies; RANS airflow over terrain. We find: the predictive accuracy to be comparable to deterministic supervised learning approaches in the restricted setting where supervised learning is applicable; UQ to be well calibrated and robust on challenging problems with complex geometries.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 3 minor

Summary. The manuscript introduces Geometric Autoencoders for Bayesian Inversion (GABI), a 'learn first, observe later' framework for uncertainty quantification in engineering systems with variable geometries. It trains a geometry-conditioned generative model on ensembles of simulations from systems with diverse geometries to create a latent prior, which is then combined with an arbitrary observation likelihood using Approximate Bayesian Computation (ABC) sampling to obtain a geometry-adapted posterior. The approach does not require knowledge of governing PDEs, boundary conditions, or specific observation processes during training. Experiments are presented on steady-state heat conduction over rectangular domains, RANS flow around airfoils, 3D Helmholtz resonance on car bodies, and RANS airflow over terrain, showing predictive accuracy comparable to deterministic supervised learning where applicable, and well-calibrated uncertainty quantification.

Significance. If the central claims hold, this work is significant for Bayesian inversion tasks involving complex, variable geometries where standard multi-system approaches fail. By distilling information from large simulation datasets into geometry-aware priors without PDE knowledge, it provides a practical way to share information across related systems. The architecture-agnostic nature and GPU-efficient ABC implementation are notable strengths. The empirical results on multiple challenging problems support the potential for broader application in engineering UQ, with explicit credit due for the reproducible experimental validation across four distinct physical regimes demonstrating calibrated posteriors.

minor comments (3)
  1. [§3.2] §3.2: The ABC sampling procedure would benefit from an explicit pseudocode listing of the GPU-parallelized steps, including how geometry conditioning is applied during proposal generation, to improve reproducibility.
  2. [Figure 4] Figure 4 (airfoil results): The legend and caption do not indicate the exact number of pressure observations used per test case, which is needed to evaluate the degree of ill-posedness addressed by the geometry-adapted prior.
  3. [Table 2] Table 2: The reported calibration metrics for the 3D Helmholtz case lack error bars or bootstrap estimates, making it difficult to assess whether the reported coverage probabilities are statistically distinguishable from the supervised baseline.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive and accurate summary of our work on Geometric Autoencoders for Bayesian Inversion (GABI). The assessment correctly captures the 'learn first, observe later' paradigm, the geometry-conditioned latent priors, the use of ABC sampling, and the empirical validation across multiple physical regimes. We appreciate the recognition of the method's significance for variable-geometry Bayesian inversion tasks and the strengths noted in the architecture-agnostic design and reproducible experiments. The recommendation for minor revision is welcome, and we will incorporate any specific suggestions in the revised manuscript.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper trains a geometry-conditioned generative model on simulation ensembles of full-field responses across varying geometries, without reference to the target observation processes or likelihoods. This prior is then combined at inference time with an arbitrary observation likelihood via ABC sampling. The training phase operates independently of the specific Bayesian inversion task, and the resulting posterior is not equivalent to any fitted quantity by construction. No load-bearing self-citations, self-definitional steps, or ansatzes that reduce the central claim to its inputs are present. The derivation remains self-contained and externally benchmarked against supervised baselines on multiple engineering problems.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The framework rests on availability of large multi-geometry datasets and the capacity of autoencoders to learn useful latent representations without explicit physics knowledge.

axioms (1)
  • domain assumption Large datasets of physical responses with varying geometries are available and representative for training.
    Method requires such data to distill the latent prior as stated in the abstract.

pith-pipeline@v0.9.0 · 5832 in / 1048 out tokens · 38508 ms · 2026-05-18T14:19:03.974923+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Distributional Inverse Homogenization

    physics.comp-ph 2026-04 unverdicted novelty 8.0

    Distributional inverse homogenization recovers microstructural statistics from macroscopic mechanical measurements by leveraging collections of bulk data in periodic and stochastic settings.

  2. Distributional Inverse Homogenization

    physics.comp-ph 2026-04 unverdicted novelty 7.0

    Distributional inverse homogenization learns microstructural statistics from bulk mechanical measurements by inverting the homogenization process statistically.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages · cited by 1 Pith paper · 7 internal anchors

  1. [1]

    Adam: A Method for Stochastic Optimization

    Kingma DP Ba J Adam et al. A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 1412(6),

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Analyzing Inverse Problems with Invertible Neural Networks

    Lynton Ardizzone, Jakob Kruse, Sebastian Wirkert, Daniel Rahner, Eric W Pellegrini, Ralf S Klessen, Lena Maier-Hein, Carsten Rother, and Ullrich Köthe. Analyzing inverse problems with invertible neural networks.arXiv preprint arXiv:1808.04730,

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges

    10 Michael M Bronstein, Joan Bruna, Taco Cohen, and Petar Veli ˇckovi´c. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges.arXiv preprint arXiv:2104.13478,

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    Graph transformers for inverse physics: recon- structing flows around arbitrary 2d airfoils.arXiv preprint arXiv:2501.17081,

    Gregory Duthé, Imad Abdallah, and Eleni Chatzi. Graph transformers for inverse physics: recon- structing flows around arbitrary 2d airfoils.arXiv preprint arXiv:2501.17081,

    work page arXiv

  5. [5]

    Variational Graph Auto-Encoders

    Thomas N Kipf and Max Welling. Variational graph auto-encoders.arXiv preprint arXiv:1611.07308,

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    Semi-Supervised Classification with Graph Convolutional Networks

    TN Kipf. Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907,

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    Nonlocality and nonlinearity implies universality in operator learning.URL https://arxiv

    11 Samuel Lanthaler, Zongyi Li, and Andrew M Stuart. Nonlocality and nonlinearity implies universality in operator learning.URL https://arxiv. org/abs/2304.13221,

    work page arXiv

  8. [8]

    Deepergcn: All you need to train deeper gcns.arXiv preprint arXiv:2006.07739,

    Guohao Li, Chenxin Xiong, Ali Thabet, and Bernard Ghanem. Deepergcn: All you need to train deeper gcns.arXiv preprint arXiv:2006.07739,

    work page arXiv 2006

  9. [9]

    Decoupled Weight Decay Regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101,

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    The geometrickernels package: Heat and matérn kernels for geometric learning on manifolds, meshes, and graphs

    Peter Mostowsky, Vincent Dutordoir, Iskander Azangulov, Noémie Jaquier, Michael John Hutchinson, Aditya Ravuri, Leonel Rozo, Alexander Terenin, and Viacheslav Borovitskiy. The geometrickernels package: Heat and matérn kernels for geometric learning on manifolds, meshes, and graphs. arXiv:2407.08086,

    work page arXiv

  11. [11]

    Relational vae: A continuous latent variable model for graph structured data.arXiv preprint arXiv:2106.16049,

    Charilaos Mylonas, Imad Abdallah, and Eleni Chatzi. Relational vae: A continuous latent variable model for graph structured data.arXiv preprint arXiv:2106.16049,

    work page arXiv

  12. [12]

    Geometric autoencoders–what you see is what you decode.arXiv preprint arXiv:2306.17638,

    Philipp Nazari, Sebastian Damrich, and Fred A Hamprecht. Geometric autoencoders–what you see is what you decode.arXiv preprint arXiv:2306.17638,

    work page arXiv

  13. [13]

    Differentiable Physics-informed Graph Networks

    Sungyong Seo and Yan Liu. Differentiable physics-informed graph networks.arXiv preprint arXiv:1902.02950,

    work page internal anchor Pith review Pith/arXiv arXiv 1902

  14. [14]

    doi:https://doi.org/10.1016/j.neunet.2021.04.018

    ISSN 0893-6080. doi:https://doi.org/10.1016/j.neunet.2021.04.018. URL https: //www.sciencedirect.com/science/article/pii/S0893608021001556. Nobuyuki Umetani and Bernd Bickel. Learning three-dimensional flow for interactive aerodynamic design.ACM Transactions on Graphics (TOG), 37(4):1–10,

    work page doi:10.1016/j.neunet.2021.04.018 2021

  15. [15]

    Learning incompressible fluid dynamics from scratch–towards fast, differentiable fluid models that generalize.arXiv preprint arXiv:2006.08762,

    Nils Wandel, Michael Weinmann, and Reinhard Klein. Learning incompressible fluid dynamics from scratch–towards fast, differentiable fluid models that generalize.arXiv preprint arXiv:2006.08762,

    work page arXiv 2006

  16. [16]

    A PROOFS A.1 PROOF OFLEMMA2.1 Proof.Remember Z(y) = Z U exp (−Φ(u;y)) dPu;P u :=g #Pz. Thus, assuming the relevant measurabilities, Z(y) = Z U exp (−Φ(u;y)) d(g#Pz) = Z Z exp (−Φ(g(u);y)) dPz, 13 where second equality is given by Theorem 3.6.1 in Bogachev & Ruas (2007). Applying the same equality,∀A∈ F, Pu|y(A) = 1 Z(y) Z A exp (−Φ(u;y)) d(g #Pz) = 1 Z(y)...

    work page 2007

  17. [17]

    The rectangle, airfoil, and car examples are run on a single RTX4090 GPU

    with maximum marginal likelihood estimation of the hyperparameters. The rectangle, airfoil, and car examples are run on a single RTX4090 GPU. The terrain example is run in a multi-GPU manner. B.2 GENERALIZEDAGGREGATIONNETWORK(GEN) For the graph-based autoencoder applied to the flow over complex terrain, we replace standard GCN layers with Generalized Aggr...

    work page 2020

  18. [18]

    Transformers can function as neural operators (Cao, 2021), achieving non-locality through attention mechanisms

    B.3.3 TRANSFORMERGABI As our framework is architecture-agnostic, we also test a Transformer-based variant. Transformers can function as neural operators (Cao, 2021), achieving non-locality through attention mechanisms. Unlike the graph-based approach, this variant treats the input as an unordered point set with coordi- nates, without exploiting the underl...

    work page 2021

  19. [19]

    We keep 100 samples as being drawn from the posterior

    Pred:Once trained, at inference time, in GABI-ABC we decode 50k samples in 100 batches of 500 samples. We keep 100 samples as being drawn from the posterior. B.5.2 ADDITIONALRESULTS In Figure 12 we show some example geometries along with the forcing function to and resonance field to reconstruct. In Figure 13 we show 3 random draws from the prior for a gi...

    work page 2020