Pith. sign in

REVIEW 3 major objections 5 minor 30 references

A Deep Generative Model for the Simulation of Discrete Karst Networks

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A two-stage graph generative pipeline trained on overlapping subgraphs of a single surveyed karst network reproduces the network's topology and most geometry metrics, with branch tortuosity consistently too high.

desk verdict An honest and reproducible first step toward learning karst network generation from a single survey, but the in-sample evaluation stops well short of the cross-formation claim. read the letter →

arxiv 2506.09832 v1 pith:L5PGQMUO submitted 2025-06-11 stat.ML cs.LG

Dany Lauzon , Julien Straubhaar , Philippe Renard This is my paper
classification stat.MLcs.LG
keywords karstnetworksdiscretenetworksimulationgraphgenerativemodelsRNNdenoisingdiffusionprobabilisticstochastichydrogeologygeometryandtopologymetrics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to show that discrete karst conduit networks can be simulated stochastically without modeling dissolution physics, by learning the statistics of an existing survey. It represents a karst network as a graph and splits generation into two learned steps: GraphRNN produces the topology node by node, and a graph denoising diffusion model (G-DDPM) places the nodes in space and attaches properties. Training data come from 500 overlapping breadth-first-search subgraphs cut from a single main network, which lets one survey stand in for a dataset. Evaluated on two real cave systems, one mostly two-dimensional and one fully three-dimensional, the generated ensembles match the surveyed topology well and match most geometry metrics, except that branch tortuosity is consistently too high. If the approach holds up, hydrogeologists could generate many plausible network realizations for flow and transport uncertainty studies from a single mapped cave.

What carries the argument

The load-bearing object is the two-step generative pipeline: GraphRNN, an autoregressive recurrent network that emits graph topology one node at a time under a breadth-first-search node ordering, and G-DDPM, a denoising diffusion model whose U-Net-style graph neural network (GraphSAGE layers) learns to remove noise from continuous node features while keeping the topology fixed. The argument is carried by the separation of topology from geometry, plus a data-augmentation step that extracts 500 connected BFS subgraphs of controlled size from a single main graph. The accompanying evaluation machinery is a set of ten statistical metrics—five geometric and five topological—computed on both training and generated graphs.

What would settle it

Train the pipeline on one surveyed karst network, then generate a large ensemble and compare its full-network statistics—branch-length distribution, tortuosity, degree correlations, and the joint distribution of cross-section width and height—against an independent survey of a different formation that was never used in training; the central claim fails if the same systematic tortuosity inflation and near-Gaussian feature distributions appear there, or if the generated ensembles cannot reproduce the statistics of a held-out portion of the same cave.

Watch

Extended reading notes

Core claim

The central claim is that a purely statistical, graph-based deep generative model can capture the distribution of discrete karst networks well enough for stochastic simulation. The paper demonstrates this with a two-step architecture: GraphRNN learns the sequential, BFS-ordered adjacency structure, and G-DDPM learns to denoise continuous node features—spatial coordinates plus, in the 3D case, log cross-section width and height—on the generated topology. On both test systems the topological metrics (average shortest path length, central point dominance, degree correlation, mean and coefficient of variation of degree) are reproduced with little difference, and the geometric metrics are mostly reproduced; the paper's own results show mean tortuosity as the one persistent failure, with generated branches more winding than surveyed ones, and the generated log cross-section features clinging to Gaussian shapes. The authors interpret this as evidence that the method learns the skeleton and connectivity of karst systems while struggling with the fine spatial placement of degree-2 nodes, and they argue the approach can transfer to other karstic environments because it does not depend on formation-specific physical processes.

Load-bearing premise

The whole approach rests on the assumption that 500 overlapping local chunks cut from one surveyed cave network represent the variability of that formation, and of other karst formations, so a model trained on those chunks can generate realistic networks elsewhere.

Editorial extensions

If this is right

  • Stochastic discrete karst networks can be generated from a single mapped survey, replacing chemistry-based or fast-marching simulators that need a geological grid and many parameters.
  • Ensembles of generated networks can feed groundwater flow and transport models, directly addressing the structural uncertainty that dominates prediction uncertainty in karst aquifers.
  • The same two-step recipe extends from a nearly 2D coastal cave system to a 3D alpine network with conduit cross-section data, so the method is not tied to one morphology.
  • Generated networks should be used with caution in applications sensitive to conduit sinuosity, because the paper's own results show mean tortuosity is systematically higher in generated graphs.
  • The authors propose the model could fill unmapped or inaccessible sections of a cave with plausible structures, which would require conditioning generated graphs on existing mapped fragments.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The 'various types of formations' claim is broader than the evidence: both test networks were used to build their own training sets, so cross-formation transfer is asserted, not demonstrated; a third, independent karst survey would be the decisive test.
  • The consistently high tortuosity is consistent with G-DDPM adding independent coordinate noise on topologies it never saw during training; a cheap test would be to smooth generated branches or condition on branch-angle distributions and see whether the other geometry metrics stay intact.
  • Because BFS subgraphs are overlapping neighborhoods of a single graph, the training distribution is not an i.i.d. sample of karst networks; comparing subgraph statistics with whole-network statistics would reveal how much local bias the augmentation injects.
  • The near-Gaussian generated distributions for log cross-section width and height suggest the diffusion model is not capturing feature-feature or feature-space correlations; reporting joint statistics, such as width versus local curvature, would make the limitation visible.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a two-step deep generative pipeline for stochastic simulation of discrete karst networks: GraphRNN generates graph topology, and a graph denoising diffusion probabilistic model (G-DDPM) generates node features (coordinates and additional attributes such as conduit width and height). To obtain training data from a single mapped network, the authors extract 500 connected subgraphs by breadth-first search with a prescribed Gaussian node-count distribution. The method is tested on two real karst systems, Ox Bel Ha (2D, 10,098 nodes) and Sakany (3D, 1,424 nodes). Evaluation compares ten geometry and topology metrics between the BFS-subgraph training set and the generated graphs. The authors report good matches for topology metrics and for most geometry metrics, but acknowledge a systematic overestimation of tortuosity in both case studies. The abstract claims the methodology allows stochastic simulation of discrete karst networks across various types of formations.

Significance. The paper is a novel application of graph generative models to a geoscience problem where stochastic simulation matters for uncertainty quantification in flow and transport. The two-step design, separating topology generation from feature generation, is sensible and the choice of metrics follows established karst-network characterization practice. The authors provide open code and data for the Sakany case, which is a concrete reproducibility asset. If the system-scale and cross-formation claims were established, the method would be a useful complement to process-based karst simulators. However, the evidence currently supports a narrower statement: the model reproduces local BFS-subgraph statistics of the two training networks reasonably well on most metrics, with a known geometry deficiency. The generalization of the trained model to full formation-scale networks and to other formations is not demonstrated.

major comments (3)
  1. [Sec. 5.1 and Figs. 10, 11, 15] The evaluation is entirely in-sample with respect to the BFS-subgraph training distribution. Generated graphs are compared only to subgraphs of the same type from which the training data were drawn, never to the full surveyed networks (Ox Bel Ha: 10,098 nodes; Sakany: 1,424 nodes) or to any independent karst survey. Because BFS subgraphs are overlapping local patches with node counts constrained to N(mu, sigma^2), their metric distributions reflect patch size and boundary effects; matching them is necessary but not sufficient for realism at the system scale relevant to flow and transport. The abstract's claim of stochastic simulation of discrete karst networks across various types of formations therefore goes beyond what the experiments demonstrate. Please add a comparison of generated full-size networks (or suitably assembled outputs) to the full-network metrics, or explicitly revise the central claim to the local-subgraph level.
  2. [Sec. 5.2, Fig. 10; Sec. 5.3, Fig. 15] The mean tortuosity is consistently higher for generated graphs in all Ox Bel Ha scenarios and in the Sakany case, and the coefficient of variation of branch length is also higher for Sakany. This is a geometry error in the direction that directly affects flow-path lengths and transport behavior, yet the paper does not quantify the discrepancy or test its impact on flow and transport. Since the stated motivation is that the generated networks be useful for studying flow and transport, the manuscript should either (i) reduce the tortuosity excess through the branch-smoothing pre-processing proposed in Sec. 6, (ii) quantify the magnitude of the bias and its practical consequences, or (iii) clearly delimit the claims to topology and coarse geometry.
  3. [Sec. 5.1 and Sec. 6] The data-augmentation premise is that N=500 BFS-sampled subgraphs extracted from one main graph represent the variability of the whole formation and, by extension, of other formations. No evidence is given that the local subgraph statistics match the global network statistics, nor that the Gaussian node-count distribution is appropriate. The text asserts that 'the extracted subgraphs best preserve the local topology' but does not show that local topology is representative of the system-scale connectivity. Please provide a diagnostic comparing the ten metrics computed on the full networks with the distribution of the same metrics on the subgraph training set, and discuss how overlapping, statistically dependent BFS patches support generalization to different karst environments.
minor comments (5)
  1. [Sec. 3.1] The statement that hyperparameters were 'set after some trial' is vague; please report the number of trials or at least the range considered, because the claimed flexibility of the architecture is otherwise difficult to assess.
  2. [Sec. 4] The Karstnet package is referenced only by URL; please add a formal citation or version identifier for the package in the references.
  3. [Sec. 5.3, Fig. 14 caption] The caption says 'for the four scenarios (each row)' but the Sakany study uses a single scenario; the caption should be corrected.
  4. [Throughout] There are several typographical errors, including 'betweeen' (Sec. 5.3), 'bottelneck' (Sec. 3.2), 'featuresis' (Sec. 5.1), and inconsistent notation for the Gaussian distribution (e.g., 'N(75,7.52)' should be 'N(75, 7.5^2)').
  5. [Sec. 6] The sentence 'this research area is entirely new' overstates novelty: stochastic karst network generators exist (Jaquet et al., 2004; Borghi et al., 2012; Fandel et al., 2022) and graph generative models exist (GraphRNN, G-DDPM); the novelty lies in the specific combination and application, which is already clear from the rest of the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the generative pipeline is trained with standard losses and evaluated post hoc; the in-sample subgraph comparison is a generalizability limitation, not a derivation that reduces to its inputs.

full rationale

The derivation chain is self-contained with respect to the two external generative methods it builds on. GraphRNN is trained by teacher forcing with a binary cross-entropy loss, and G-DDPM is trained with an MSE loss on the added noise; neither loss optimizes the ten evaluation metrics, which are computed post hoc on generated samples. There is no equation in the paper that defines a predicted quantity in terms of a fitted parameter or vice versa. The hyperparameters are acknowledged as chosen by trial (Sec. 3.1), and the BFS window size m is a data-derived preprocessing constant (Sec. 2.1.1); neither is renamed as a prediction. The main validity weakness is that generated graphs are compared only to the BFS-subgraph training distribution (Figs. 10, 11, 15), not to a held-out set or the full surveyed networks, which makes the system-scale and cross-formation claims under-supported. That is an in-sample evaluation and generalization concern, not circularity: the model is trained to reproduce a distribution, and matching the training distribution is a sanity check rather than an independent prediction. The paper itself flags the main failure (tortuosity is consistently higher for generated subgraphs, Sec. 5.2) and explicitly states that conditioning on existing paths and realistic extrapolation require further research (Sec. 6). Self-citations, such as Collon et al. (2017) for the statistical metrics and prior DKN modeling work, provide published external methodology and are not load-bearing justifications of the central result. No uniqueness theorem, ansatz, or known result is imported from the authors' own prior work to force the conclusion. The reported topology match is a genuine empirical result, though its generalization scope is narrower than the abstract claims.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The model introduces no new physical entities, forces, or particles. The free parameters are all either hyperparameters chosen by trial or statistics of the training data (m, subgraph size distributions, timesteps, epochs, hidden widths). The axioms are standard machine learning background (GRU, DDPM, GraphRNN training) plus two domain assumptions specific to this paper: that BFS subgraphs of one cave represent the full distribution of karst network variability, and that network regularity justifies a small BFS window m. The former is the load-bearing one and is not validated out-of-sample.

free parameters (5)
  • BFS neighborhood size m = 99.9th percentile over 100,000 BFS orderings of dataset graphs (not stated numerically)
    Controls the width of the adjacency encoding used by GraphRNN (Sec. 2.1.1); it is estimated from the training graphs rather than set by a theorem.
  • Subgraph node-count distribution (mean mu, variance sigma^2) = N(75,7.5^2), N(100,10^2), N(200,20^2), N(500,50^2) per scenario
    Chosen by hand to define the training subgraph sizes in the data augmentation scheme (Sec. 5.1); these distributions directly determine what sizes of generated graphs are possible.
  • Diffusion timesteps T and training epochs = T=2400, epochs=1201 for G-DDPM; 5*(mu+3*sigma) epochs for GraphRNN
    Set by trial; the epoch rule 5*(mu+3*sigma) is an ad hoc scaling 'reflecting the complexity of the topology' (Sec. 5.1).
  • G-DDPM U-Net hidden width and GraphSAGE depth = 8*d feature width, n=25 GraphSAGE operations
    Hand-chosen hyperparameters of the denoising network (Sec. 3.2).
  • GraphRNN hidden sizes and layer counts = L_G=L_E=4 layers, d_G,hidden=48, d_E,GRU(in)=24, etc.
    Selected 'after some trial, but other setups could be chosen' (Sec. 3.1).
assumptions (5)
  • domain assumption A karst conduit network is adequately represented as an undirected graph whose edges are conduits and whose node features are independent given the topology.
    Justified by the nature of cave surveys (Sec. 1), but the paper notes feature-spatial correlations are not modeled (Sec. 5.3), so the assumption is only partially tested.
  • ad hoc to paper BFS subgraphs of the first n nodes preserve the local topology and local statistics of the full network.
    Stated in Sec. 5.1: 'BFS ensures the selection of the nearest neighbors of the starting node... Therefore, the extracted subgraphs best preserve the local topology.' No quantitative validation against the global network statistics is provided.
  • standard math The DDPM forward process with scalar covariance beta_t*I converges to independent Gaussians, so the reverse process can be learned by MSE on the noise.
    Standard DDPM theory (Ho et al., 2020), Equations (2)-(3) in Sec. 2.2.1.
  • domain assumption Karst networks are 'particularly regular', so a small BFS window m suffices.
    Stated in Sec. 2.1.1; the same claim is made in You et al. (2018) for the original GraphRNN.
  • standard math Teacher forcing with binary cross-entropy on adjacency bits yields a valid graph topology distribution at inference.
    Standard GraphRNN training procedure (You et al., 2018), described in Sec. 2.1.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Deep Generative Model for the Simulation of Discrete Karst Networks." pith.science (2026). https://pith.science/paper/L5PGQMUO

@misc{pith2026250609832,
  author       = {Pith},
  title        = {Pith review of: A Deep Generative Model for the Simulation of Discrete Karst Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L5PGQMUO}},
  note         = {Machine review of arXiv:2506.09832}
}
read the original abstract

The simulation of discrete karst networks presents a significant challenge due to the complexity of the physicochemical processes occurring within various geological and hydrogeological contexts over extended periods. This complex interplay leads to a wide variety of karst network patterns, each intricately linked to specific hydrogeological conditions. We explore a novel approach that represents karst networks as graphs and applies graph generative models (deep learning techniques) to capture the intricate nature of karst environments. In this representation, nodes retain spatial information and properties, while edges signify connections between nodes. Our generative process consists of two main steps. First, we utilize graph recurrent neural networks (GraphRNN) to learn the topological distribution of karst networks. GraphRNN decomposes the graph simulation into a sequential generation of nodes and edges, informed by previously generated structures. Second, we employ denoising diffusion probabilistic models on graphs (G-DDPM) to learn node features (spatial coordinates and other properties). G-DDPMs enable the generation of nodes features on the graphs produced by the GraphRNN that adhere to the learned statistical properties by sampling from the derived probability distribution, ensuring that the generated graphs are realistic and capture the essential features of the original data. We test our approach using real-world karst networks and compare generated subgraphs with actual subgraphs from the database, by using geometry and topology metrics. Our methodology allows stochastic simulation of discrete karst networks across various types of formations, a useful tool for studying the behavior of physical processes such as flow and transport.

Figures

Figures reproduced from arXiv: 2506.09832 by the authors.

Figure 1
Figure 1. Illustration of GraphRNN (x: input vector; h: hidden state vector; y: output vector; p: vector of predicted probabilities; SOS: start-of-sequence, a vector of ones; B(p): Bernoulli distribution of parameter p). 2.1.1 Node Ordering and Adjacency Matrix Encoding One significant challenge with a sequential approach is that every node must be visited to ensure all possible connections, potentially leading to time-consum… view at source ↗
Figure 2
Figure 2. Different node orderings π on a 5 × 4 grid graph. Left: Numbering from bottom to top and left to right. Middle: Random ordering. Right: one BFS ordering. of dimension n−1×m, is defined as ci,j = ai−j,i (if i−j ≥ 0 and 0 otherwise. It is important to note that this embedding is reversible and corresponds to a unique graph or adjacency matrix, which means that the transformation preserves a one-to-one correspondence b… view at source ↗
Figure 3
Figure 3. ) correspond to the vectors x given as inputs to the Graph-Level RNN in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Design of the RNN model (GRU: Gated Recurrent Unit, MLP: Multi-Layer Percep￾tron, ReLU: Rectified Linear Unit). The Graph-Level RNN employs LG = 4 layers of GRU cells, with a dG,hidden = 48- dimensional hidden state in each layer (vectors h in [PITH_FULL_IMAGE:figures…
Figure 5
Figure 5. Figure 5: Design of the GDDPM: U-net architecture, based on GraphSAGE operations and time embedding. The network takes as input the set of input features xin = xt defined on a connected graph, and the corresponding timestep t, and predicts the noise ϵˆt through a sequence of ope…
Figure 6
Figure 6. Figure 6: Illustration of GraphSAGE operation: ui is the input vector of features attached to node i, the green nodes are the neighbors N (i) of the node i involved in a GraphSAGE operation: u ′ i = W1 · ui + W2 · meanj∈N(i) uj ; input and output vectors of features are of dimen…
Figure 7
Figure 7. Figure 7: Main graph of the Ox Bel Ha karst system, in two dimension (the vertical dimension is omitted, because only very low variations are observed), 10 098 nodes, 10 098 edges. 5.2 Coastal–Karst : Ox Bel Ha Ox Bel Ha, one of the longest cave systems in Mexico, is also among …
Figure 8
Figure 8. Figure 8: First graphs of datasets (left column) and first generated graphs (right column), where the dataset is built with a number of nodes following the distribution (a) N (75, 7.5 2 ), (b) N (100, 102 ), (c) N (200, 202 ), and (d) N (500, 502 ). distributions from the datase…
Figure 9
Figure 9. Figure 9: Histogram of number of nodes (left), and x coordinate (middle column) and y coordi￾nate (right column) of nodes for all graphs in the dataset (orange) and all generated graphs (blue) for the four scenarios (each row). The ten statistical metrics presented in Sec. 4 are…
Figure 10
Figure 10. Figure 10: Geometry metrics (from left to right: ‘mean length, ‘cv length’, ‘length entropy’, ‘mean tortuosity’, ‘orientation entropy’, see [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]
Figure 11
Figure 11. Figure 11: Topology metrics (from left to right: ‘aspl’, ‘cpd’, ‘node degree corr.’, ‘mean degree’, ‘cv degree’, see [PITH_FULL_IMAGE:figures/full_fig_p019_11.png]
Figure 12
Figure 12. Figure 12: Main graph of the Sakany karst system, 1424 nodes, 1463 edges; node features are represented: logarithm in base 10 of cross-section width (‘log10 cw’, left) and height (‘log10 ch’, right). The dataset used to train the deep neural networks is built by randomly extract…
Figure 13
Figure 13. Figure 13: First graphs of the dataset (left column) and first generated graphs (right column), node feature ‘log10 cw’ (top row), and ‘log10 ch’ (bottom row). betweeen the features and spatial statistics are not considered in this work. It is important to note that for this cas…
Figure 14
Figure 14. Figure 14: Histogram of number of nodes (left), and each node features (x, y, z coordinates and ‘log10 cw’, ‘log10 ch’, right) for all graphs in the dataset (orange) and all generated graphs (blue) for the four scenarios (each row). (a) Geometry metrics (b) Topology metrics [PI…
Figure 15
Figure 15. Figure 15: Geometry (a) and topology (b) metrics (see [PITH_FULL_IMAGE:figures/full_fig_p022_15.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 18 canonical work pages

  1. [1]

    , Renard, P

    Borghi2012 APACrefauthors Borghi, A. , Renard, P. \ Jenni, S. APACrefauthors \ 2012 . A pseudo-genetic stochastic model to generate karstic networks A pseudo-genetic stochastic model to generate karstic networks . Journal of Hydrology 414–415 516--529 . APACrefDOI doi:10.1016/j.jhydrol.2011.11.032 APACrefDOI

    work page doi:10.1016/j.jhydrol.2011.11.032 2012

  2. [2]

    \ Bigot, J Y

    cassou2007 APACrefauthors Cassou, J P. \ Bigot, J Y. APACrefauthors \ 2007 . Le labyrinthe de la grotte de Sakany (Qui \'e , Ari \`e ge) Le labyrinthe de la grotte de sakany (qui \'e , ari \`e ge) . Actes de la 17e Rencontre d'Octobre - Orgnac 2007 17 29--36

    work page 2007

  3. [3]

    \ Delfiner, P

    Chiles2012 APACrefauthors Chilès, J. \ Delfiner, P. APACrefauthors \ 2012 . Geostatistics: Modeling spatial uncertainty Geostatistics: Modeling spatial uncertainty . Wiley . APACrefDOI doi:10.1002/9781118136188 APACrefDOI

    doi:10.1002/9781118136188 2012

  4. [4]

    , van Merrienboer, B

    Cho2014 APACrefauthors Cho, K. , van Merrienboer, B. , Gulcehre, C. , Bahdanau, D. , Bougares, F. , Schwenk, H. \ Bengio, Y. APACrefauthors \ 2014 . Learning phrase representations using RNN encoder-decoder for statistical machine translation. Learning phrase representations using RNN encoder-decoder for statistical machine translation. arXiv . APACrefDOI...

    arXiv doi:10.48550/arxiv.1406.1078 2014

  5. [5]

    APACrefauthors \ 2019

    Coke2019 APACrefauthors Coke, J G. APACrefauthors \ 2019 . Underwater Caves of the Yucatan Peninsula Underwater caves of the yucatan peninsula . Encyclopedia of Caves Encyclopedia of caves \ ( 1089–1095). Elsevier . APACrefURL http://dx.doi.org/10.1016/b978-0-12-814124-3.00127-8 APACrefURL APACrefDOI doi:10.1016/b978-0-12-814124-3.00127-8 APACrefDOI

    work page doi:10.1016/b978-0-12-814124-3.00127-8 2019

  6. [6]

    , Bernasconi, D

    Collon2017 APACrefauthors Collon, P. , Bernasconi, D. , Vuilleumier, C. \ Renard, P. APACrefauthors \ 2017 . Statistical metrics for the characterization of karst network geometry and topology Statistical metrics for the characterization of karst network geometry and topology . Geomorphology 283 122--142 . APACrefDOI doi:10.1016/j.geomorph.2017.01.034 APACrefDOI

    work page doi:10.1016/j.geomorph.2017.01.034 2017

  7. [7]

    , Ferré, T

    Fandel2021 APACrefauthors Fandel, C. , Ferré, T. , Chen, Z. , Renard, P. \ Goldscheider, N. APACrefauthors \ 2021 . A model ensemble generator to explore structural uncertainty in karst systems with unmapped conduits A model ensemble generator to explore structural uncertainty in karst systems with unmapped conduits . Hydrogeology Journal 29 1 229--248 . ...

    work page doi:10.1007/s10040-020-02227-6 2021

  8. [8]

    , Miville, F

    Fandel2022 APACrefauthors Fandel, C. , Miville, F. , Ferré, T. , Goldscheider, N. \ Renard, P. APACrefauthors \ 2022 . The stochastic simulation of karst conduit network structure using anisotropic fast marching, and its application to a geologically complex alpine karst system The stochastic simulation of karst conduit network structure using anisotropic...

    work page 2022

Show all 30 references
  1. [9]

    , Jeannin, P Y

    Filipponi2009 APACrefauthors Filipponi, M. , Jeannin, P Y. \ Tacher, L. APACrefauthors \ 2009 . Evidence of inception horizons in karst conduit networks Evidence of inception horizons in karst conduit networks . Geomorphology 106 1–2 86--99 . APACrefDOI doi:10.1016/j.geomorph....

    2009 doi

  2. [10]

    APACrefauthors \ 1978

    freeman1978 APACrefauthors Freeman, L C. APACrefauthors \ 1978 . Centrality in social networks conceptual clarification Centrality in social networks conceptual clarification . Social Networks 1 3 215-239 . APACrefDOI doi:10.1016/0378-8733(78)90021-7 APACrefDOI

    1978 doi

  3. [11]

    , Miller, J

    Goodman1991 APACrefauthors Goodman, R. , Miller, J. \ Smyth, P. APACrefauthors \ 1991 . Objective Functions For Neural Network Classifier Design Objective functions for neural network classifier design . Proceedings. 1991 IEEE International Symposium on Information Theory Proc...

    1991

  4. [12]

    , Collon, P

    Gouy2024 APACrefauthors Gouy, A. , Collon, P. , Bailly-Comte, V. , Galin, E. , Antoine, C. , Thebault, B. \ Landrein, P. APACrefauthors \ 2024 . Karst NS im: A graph-based method for 3 D geologically-driven simulation of karst networks Karst NS im: A graph-based method for 3 D...

    2024

  5. [13]

    APACrefauthors \ 2024

    groupe_de_recherches_et_dactivites_speleologiques_de_lourdes_base_2024 APACrefauthors Groupe de Recherches et d'Activités Spéléologiques de Lourdes . APACrefauthors \ 2024 . Base de données topographiques de la grotte Sakany . Base de données topographiques de la grotte Sakany...

    2024

  6. [14]

    , Jain, A

    Ho2020 APACrefauthors Ho, J. , Jain, A. \ Abbeel, P. APACrefauthors \ 2020 . Denoising diffusion probabilistic models. Denoising diffusion probabilistic models. arXiv . APACrefDOI doi:10.48550/ARXIV.2006.11239 APACrefDOI

    2020 doi arXiv

  7. [15]

    , Satorras, V G

    Hoogeboom2022 APACrefauthors Hoogeboom, E. , Satorras, V G. , Vignac, C. \ Welling, M. APACrefauthors \ 2022 . Equivariant Diffusion for Molecule Generation in 3D. Equivariant diffusion for molecule generation in 3d. arXiv . APACrefDOI doi:10.48550/ARXIV.2203.17003 APACrefDOI

    2022 doi arXiv

  8. [16]

    , St\" a rk, H

    Igashov2024 APACrefauthors Igashov, I. , St\" a rk, H. , Vignac, C. , Schneuing, A. , Satorras, V G. , Frossard, P. Correia, B. APACrefauthors \ 2024 . Equivariant 3D-conditional diffusion model for molecular linker design Equivariant 3d-conditional diffusion model for molecul...

    2024 doi

  9. [17]

    , Siegel, P

    Jaquet2004 APACrefauthors Jaquet, O. , Siegel, P. , Klubertanz, G. \ Benabderrhamane, H. APACrefauthors \ 2004 . Stochastic discrete model of karstic networks Stochastic discrete model of karstic networks . Advances in Water Resources 27 7 751--760 . APACrefDOI doi:10.1016/j.a...

    2004 doi

  10. [18]

    Liao2019 APACrefauthors Liao, R. , Li, Y. , Song, Y. , Wang, S. , Hamilton, W., L. , Duvenaud, D. Zemel, R. APACrefauthors \ 2019 . Efficient graph generation with graph recurrent attention networks Efficient graph generation with graph recurrent attention networks . Proceedin...

    2019

  11. [19]

    , Qamar, M

    Zhang2023 APACrefauthors Mengchun, Z. , Qamar, M. , Taegoo, K. , Jung, Y. , Chenshuang, Z. , Bae, S H. \ Chaoning, Z. APACrefauthors \ 2023 . A survey on graph diffusion models: Generative AI in science for molecule, protein and material. A survey on graph diffusion models: Ge...

    2023

  12. [20]

    , Renard, P

    Miville2025 APACrefauthors Miville, F. , Renard, P. , Fandel, C. \ Filipponi, M. APACrefauthors \ 2025 . pyKasso: An open-source three-dimensional discrete karst network generator pykasso: An open-source three-dimensional discrete karst network generator . Environmental Modell...

    2025

  13. [21]

    APACrefauthors \ 1991

    Palmer1991 APACrefauthors Palmer, A N. APACrefauthors \ 1991 . Origin and morphology of limestone caves Origin and morphology of limestone caves . Geological Society of America Bulletin 103 1 1--21 . APACrefDOI doi:10.1130/0016-7606(1991)103<0001:oamolc>2.3.co;2 APACrefDOI

    1991 doi

  14. [22]

    , Durán-Valsero, J J

    PardoIguzquiza2011 APACrefauthors Pardo-Iguzquiza, E. , Durán-Valsero, J J. \ Rodríguez-Galiano, V. APACrefauthors \ 2011 . Morphometric analysis of three-dimensional networks of karst conduits Morphometric analysis of three-dimensional networks of karst conduits . Geomorpholo...

    2011 doi

  15. [23]

    APACrefauthors \ 2024

    qrss_quintana_roo_speleological_survey_ox_2024 APACrefauthors QRSS (Quintana Roo Speleological Survey) , GEO (Grupo de Exploración Ox Bel Ha) , MCEP (Mexico Cave Exploration Project) \ CINDAQ (El Centro Investigador del Sistema Acuífero de Quintana Roo) . APACrefauthors \ 2024...

    2024

  16. [24]

    , van der Sluijs, J P

    Refsgaard2006 APACrefauthors Refsgaard, J C. , van der Sluijs, J P. , Brown, J. \ van der Keur, P. APACrefauthors \ 2006 . A framework for dealing with uncertainty due to model structure error A framework for dealing with uncertainty due to model structure error . Advances in ...

    2006 doi

  17. [25]

    , Fischer, P

    Ronneberger2015 APACrefauthors Ronneberger, O. , Fischer, P. \ Brox, T. APACrefauthors \ 2015 . U-Net: Convolutional Networks for Biomedical Image Segmentation. U-net: Convolutional networks for biomedical image segmentation. arXiv . APACrefURL https://arxiv.org/abs/1505.04597...

    2015 doi arXiv

  18. [26]

    \ Komodakis, N

    Simonovsky2018 APACrefauthors Simonovsky, M. \ Komodakis, N. APACrefauthors \ 2018 . GraphVAE : Towards generation of small graphs using variational autoencoders GraphVAE : Towards generation of small graphs using variational autoencoders . Artificial Neural Networks and Machi...

    2018 doi

  19. [27]

    , Shazeer, N

    Vaswani2023 APACrefauthors Vaswani, A. , Shazeer, N. , Parmar, N. , Uszkoreit, J. , Jones, L. , Gomez, A N. Polosukhin, I. APACrefauthors \ 2023 . Attention Is All You Need. Attention is all you need. arXiv . APACrefDOI doi:10.48550/arXiv.1706.03762 APACrefDOI

    2023 doi arXiv

  20. [28]

    , Krawczuk, I

    Vignac2022 APACrefauthors Vignac, C. , Krawczuk, I. , Siraudin, A. , Wang, B. , Cevher, V. \ Frossard, P. APACrefauthors \ 2022 . Di G ress: Discrete denoising diffusion for graph generation. Di G ress: Discrete denoising diffusion for graph generation. arXiv . APACrefDOI doi:...

    2022 doi arXiv

  21. [29]

    , Ying, R

    You2018 APACrefauthors You, J. , Ying, R. , Ren, X. , Hamilton, W L. \ Leskovec, J. APACrefauthors \ 2018 . Graph RNN : G enerating realistic graphs with deep auto-regressive models. Graph RNN : G enerating realistic graphs with deep auto-regressive models. arXiv . APACrefDOI ...

    2018 doi arXiv

  22. [30]

    Zhu2022 APACrefauthors Zhu, Y. , Du, Y. , Wang, Y. , Xu, Y. , Zhang, J. , Liu, Q. \ Wu, S. APACrefauthors \ 2022 . A survey on deep graph generation: Methods and applications. A survey on deep graph generation: Methods and applications. arXiv . APACrefDOI doi:10.48550/ARXIV.22...

    2022 doi arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.