pith. sign in

arxiv: 1907.00718 · v1 · pith:V6GSBDDLnew · submitted 2019-06-26 · 🧮 math.OC

Fast uncertainty quantification of reservoir simulation with variational U-Net

Larry Jin , Hannah Lu , Gege Wen This is my paper

Pith reviewed 2026-05-25 15:10 UTC · model grok-4.3

classification 🧮 math.OC
keywords uncertainty quantificationreservoir simulationvariational U-Netconvolutional neural networksurrogate modelwell controlsstochastic PDE
0
0 comments X

The pith

A variational U-Net learns hidden physical quantities to forecast reservoir production under many well controls without repeated PDE solves.

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

The paper sets out to replace conventional stochastic PDE solvers for uncertainty quantification in reservoir simulation with a convolutional encoder-decoder network. The goal is to handle arbitrary well control plans in a single trained model rather than repeating expensive Monte Carlo runs for each design. The network learns a backward step that extracts hidden physical quantities from simulation snapshots and a forward step that propagates those quantities to future production statistics under new controls. If successful, this would let engineers evaluate many production scenarios at far lower computational cost while retaining the essential uncertainty distributions of the original physics model.

Core claim

The central claim is that a variational U-Net architecture adapted from shape-guided image generation can serve as a control-guided surrogate for reservoir simulation. Backward propagation inside the network extracts hidden physical quantities; forward propagation then uses those quantities together with chosen well controls to predict future production. This learned mapping replaces repeated solves of the underlying stochastic PDE and yields uncertainty quantification for multiple control plans at substantially lower cost than Monte Carlo sampling.

What carries the argument

The variational U-Net that performs learned backward propagation to recover hidden physical quantities and learned forward propagation to generate production forecasts under varying well controls.

If this is right

  • Uncertainty quantification becomes feasible for large numbers of well-control scenarios without recomputing the underlying PDE for each one.
  • The same trained network can be queried for both mean forecasts and full uncertainty statistics under new control inputs.
  • Computational cost scales with the cost of a single network evaluation rather than the cost of repeated full-order stochastic solves.
  • The approach separates the extraction of hidden state from the application of control inputs, allowing reuse across different operating plans.

Where Pith is reading between the lines

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

  • The same encoder-decoder structure might be applied to other time-dependent physical systems where control parameters change frequently, such as groundwater flow or subsurface transport.
  • If the hidden variables truly encode the essential state, the network could be inserted into optimization loops that search for controls minimizing risk under uncertainty.
  • Extending the training set with snapshots from more heterogeneous permeability fields would test whether the learned mapping remains faithful when the underlying geology varies.

Load-bearing premise

Training the network on a finite set of simulation snapshots is enough to capture the mapping from arbitrary well controls to accurate future production statistics that match those from full stochastic PDE solves.

What would settle it

Generate uncertainty distributions from the trained network on a fresh set of well controls never seen during training, then compare them directly to distributions obtained from a large ensemble of independent Monte Carlo PDE simulations on the same controls; statistically significant mismatch in mean, variance, or tails would refute the claim.

Figures

Figures reproduced from arXiv: 1907.00718 by Gege Wen, Hannah Lu, Larry Jin.

Figure 2
Figure 2. Figure 2: Example of a data tuple: y, S, P, y 0 , S 0 , and P 0 3.3. Training, validation, and test sets For the training and the validation set, a total of 42 differ￾ent permeability map realizations k (under the same Gaus￾sian distribution) and 50 different well locations maps y were simulated in 42 × 50 = 2100 runs. Each simula￾tion lasts for 1,000 days and we took 4 time snapshots of the output, generating 2100 … view at source ↗
Figure 1
Figure 1. Figure 1: Example of the Gaussian permeability field [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: U-Net architecture variable. The following regularized MSE loss is employed as the training loss function: L(fθ(y, k), S) = 1 n Xn i=1 [fθ(yi , ki) − Si ] 2 + λkθk2, (5) where fθ is the neural net depending on the parameter θ and λ is the regularization strength. We refer to [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: (Modified) variational U-Net architecture [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Test simulation results with V-UNet for prediction [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Mean saturation (Column 1: True, Column 2: pre [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Mean pressure (Column 1: True, Column 2: pre [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
read the original abstract

Quantification of uncertainty in production/injection forecasting is an important aspect of reservoir simulation studies. Conventional approaches include intrusive Galerkin-based methods (e.g., generalized polynomial chaos (gPC) and stochastic collocation (SC) methods) and non-intrusive Monte Carlo (MC) based methods. Nevertheless, the quantification is conducted in reformulations of the underlying stochastic PDEs with fixed well controls. If one wants to take various well control plans into account, expensive computations need to be repeated for each well design independently. In this project, we take advantages of the equation-free spirit of convolutional neural network (CNN) to overcome this challenge and thus achieve the flexibility of efficient uncertainty quantification with various well controls. We are interested in the development of surrogate models for uncertainty quantification and propagation in reservoir simulations using a deep convolutional encoder-decoder network as an analogue to the image-to-image regression tasks in computer science. First, a U-Net architecture is applied to replace conventional expensive deterministic PDE solver. Then we adopt the idea from shape-guided image generation using variational U-Net and design a new variational U-Net architecture for "control-guided" reservoir simulation. Backward propagation is learned in the network to extract the hidden physical quantities and then predict the future production by the learned forward propagation using the hidden variable with various well controls. Comparisons in computational efficiency are made between our proposed CNN approach and conventional MC approach. Significant improvements in computational speed with reasonable accuracy loss are observed in the numerical tests.

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

2 major / 0 minor

Summary. The manuscript proposes a variational U-Net surrogate for reservoir simulation to enable fast uncertainty quantification under varying well controls. It adapts a convolutional encoder-decoder architecture to learn backward propagation for extracting hidden physical quantities from simulation snapshots and then uses learned forward propagation with these quantities and arbitrary controls to forecast production. The approach is positioned as an equation-free alternative to repeated Monte Carlo solves of stochastic PDEs, with claims of significant computational speed gains and only reasonable accuracy loss demonstrated in numerical tests.

Significance. If the variational U-Net accurately reproduces both mean forecasts and the full uncertainty distributions obtained from repeated stochastic PDE solves for out-of-distribution well controls, the method could provide a practical tool for efficient control-guided UQ in reservoir engineering. This would address the computational bottleneck of re-running expensive simulations for each new control plan. The work adapts established image-to-image regression techniques to a physical simulation setting, but its value hinges on whether the learned hidden variables preserve the necessary statistics without explicit physical constraints.

major comments (2)
  1. [Abstract] Abstract: The central claim that 'Significant improvements in computational speed with reasonable accuracy loss are observed in the numerical tests' is unsupported by any quantitative metrics, error bars, validation splits, training data generation details, or comparison tables. This absence makes it impossible to evaluate whether the speed-accuracy tradeoff holds or whether uncertainty statistics are preserved.
  2. [Method] Method description (variational U-Net architecture): The surrogate is trained directly on outputs from the conventional simulator. This creates a circular dependency in which the model parameters are fitted to the same class of simulation data it is meant to replace, with no mention of moment-preservation losses, divergence-free constraints, or out-of-distribution generalization tests that would be required to ensure the forward predictions under new well controls reproduce the target uncertainty distributions.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. We address the two major comments point by point below, indicating the changes we will make in revision.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that 'Significant improvements in computational speed with reasonable accuracy loss are observed in the numerical tests' is unsupported by any quantitative metrics, error bars, validation splits, training data generation details, or comparison tables. This absence makes it impossible to evaluate whether the speed-accuracy tradeoff holds or whether uncertainty statistics are preserved.

    Authors: We agree that the abstract statement would be stronger with explicit quantitative support. The manuscript body reports speed-up factors and accuracy comparisons from the numerical experiments, but these details are not summarized in the abstract. In the revised version we will update the abstract to include specific metrics (e.g., wall-clock speed-up ratios and relative L2 errors on mean and variance fields) together with a brief statement on the validation procedure and data-generation protocol. revision: yes

  2. Referee: [Method] Method description (variational U-Net architecture): The surrogate is trained directly on outputs from the conventional simulator. This creates a circular dependency in which the model parameters are fitted to the same class of simulation data it is meant to replace, with no mention of moment-preservation losses, divergence-free constraints, or out-of-distribution generalization tests that would be required to ensure the forward predictions under new well controls reproduce the target uncertainty distributions.

    Authors: Training a data-driven surrogate on simulator outputs is the standard supervised-learning approach for building fast approximations; the same data-generation step is required for any non-intrusive reduced-order model. The variational U-Net is designed precisely to learn a latent representation that can be paired with arbitrary controls at inference time. We acknowledge that the current text does not explicitly report out-of-distribution tests on well-control plans unseen during training. In revision we will add a dedicated subsection that evaluates the surrogate on held-out control schedules, comparing both first- and second-moment statistics against fresh Monte-Carlo runs, thereby demonstrating preservation of uncertainty distributions. revision: partial

Circularity Check

0 steps flagged

No circularity: surrogate trained on external simulator data for approximation

full rationale

The paper describes training a variational U-Net on snapshots from a conventional reservoir simulator to serve as a fast surrogate for PDE solves under varying well controls, then using the trained model for uncertainty quantification. This follows the standard supervised learning pattern of fitting a mapping from inputs (controls, states) to outputs (production forecasts) without any quoted step where a prediction is defined as or forced to equal its own training inputs by construction. No self-citations, uniqueness theorems, or ansatzes are invoked in the provided text to load-bear the central claim. The derivation chain is self-contained as an empirical approximation technique whose accuracy is assessed separately via numerical tests.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no explicit free parameters, axioms, or invented entities; the approach implicitly rests on standard supervised-learning assumptions (sufficient training data, appropriate loss, generalization to unseen controls) that are not stated or justified here.

pith-pipeline@v0.9.0 · 5788 in / 1160 out tokens · 52368 ms · 2026-05-25T15:10:45.843387+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages · 2 internal anchors

  1. [1]

    Babu ˇska, F

    I. Babu ˇska, F. Nobile, and R. Tempone. A stochastic collo- cation method for elliptic partial differential equations with 8 Layer Filters Output size Input ( Nx, Ny, nv) Nin lyr 8 of (1,1, nv), stride 1 ( Nx, Ny, 8) In res 8 of (3,3,8), stride 1 ( Nx, Ny, 8) Conv blk 16 of (3,3,8), stride 2 ( Nx/2, Ny/2, 16) In res 16 of (3,3,16), stride 1 ( Nx/2, Ny/2,...

    work page 2007

  2. [2]

    Baldi, P

    P. Baldi, P. Sadowski, and D. Whiteson. Searching for exotic particles in high-energy physics with deep learning. Nature communications, 5:4308, 2014

    work page 2014

  3. [3]

    Boucher, J

    A. Boucher, J. Wu, and N. Remy. Applied geostatistics with sgems. 2009

    work page 2009

  4. [4]

    Burhenne, D

    S. Burhenne, D. Jacob, and G. P. Henze. Sampling based on sobol?sequences for monte carlo techniques applied to build- ing simulations. In Building Simulation 2011: 12th Confer- ence of International Building Performance Simulation As- sociation, Sydney, Australia, Nov, pages 14–16, 2011

    work page 2011

  5. [5]

    Chan and A

    S. Chan and A. H. Elsheikh. A machine learning ap- proach for efficient uncertainty quantification using multi- scale methods. Journal of Computational Physics, 354:493– 511, 2018

    work page 2018

  6. [6]

    Esser, E

    P. Esser, E. Sutter, and B. Ommer. A variational u-net for conditional appearance and shape generation. In Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8857–8866, 2018

    work page 2018

  7. [7]

    M. B. Giles. Multilevel monte carlo methods. Acta Numer- ica, 24:259–328, 2015

    work page 2015

  8. [8]

    Goodfellow, J

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio. Gen- erative adversarial nets. In Advances in neural information processing systems, pages 2672–2680, 2014

    work page 2014

  9. [9]

    Climate change 2014: Synthesis report - summary chapter for policymakers

    IPCC. Climate change 2014: Synthesis report - summary chapter for policymakers. page 31, 2014

    work page 2014

  10. [10]

    S. Kalla. Use of orthogonal arrays, quasi-monte carlo sam- pling and kriging response models for reservoir simulation with many varying factors. 2005

    work page 2005

  11. [11]

    D. P. Kingma and M. Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  12. [12]

    J. N. Kutz. Deep learning in fluid dynamics. Journal of Fluid Mechanics, 814:1–4, 2017

    work page 2017

  13. [13]

    D. Lu, G. Zhang, C. Webster, and C. Barbier. An improved multilevel monte carlo method for estimating probability dis- tribution functions in stochastic oil reservoir simulations. Water resources research, 52(12):9642–9660, 2016

    work page 2016

  14. [14]

    Marc ¸ais and J.-R

    J. Marc ¸ais and J.-R. De Dreuzy. Prospective interest of deep learning for hydrological inference. Groundwater, 55(5):688–692, 2017

    work page 2017

  15. [15]

    S. Min, B. Lee, and S. Yoon. Deep learning in bioinformat- ics. Briefings in bioinformatics, 18(5):851–869, 2017

    work page 2017

  16. [16]

    S. Mo, Y . Zhu, N. Zabaras, X. Shi, and J. Wu. Deep convo- lutional encoder-decoder networks for uncertainty quantifi- cation of dynamic multiphase flow in heterogeneous media. arXiv preprint arXiv:1807.00882, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  17. [17]

    Raissi, P

    M. Raissi, P. Perdikaris, and G. E. Karniadakis. Physics- informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686–707, 2019

    work page 2019

  18. [18]

    Robert and G

    C. Robert and G. Casella. Monte Carlo statistical methods. Springer Science & Business Media, 2013

    work page 2013

  19. [19]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox. U-net: Convo- lutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention , pages 234–241. Springer, 2015

    work page 2015

  20. [20]

    Eclipse reference manual

    Schlumberger. Eclipse reference manual. 2014

    work page 2014

  21. [21]

    D. M. Tartakovsky and P. A. Gremaud. Method of distri- butions for uncertainty quantification. Handbook of Uncer- tainty Quantification, pages 1–22, 2016

    work page 2016

  22. [22]

    C. L. Winter, D. Tartakovsky, and A. Guadagnini. Mo- ment differential equations for flow in highly heterogeneous porous media. Surveys in Geophysics, 24(1):81–106, 2003. 9

    work page 2003

  23. [23]

    D. Xiu. Numerical methods for stochastic computations: a spectral method approach. Princeton university press, 2010

    work page 2010

  24. [24]

    Xiu and G

    D. Xiu and G. E. Karniadakis. Modeling uncertainty in flow simulations via generalized polynomial chaos. Journal of computational physics, 187(1):137–167, 2003

    work page 2003

  25. [25]

    Zhu and N

    Y . Zhu and N. Zabaras. Bayesian deep convolutional encoder–decoder networks for surrogate modeling and un- certainty quantification. Journal of Computational Physics, 366:415–447, 2018. 10

    work page 2018