acceptodds
Under review as a conference paper at ICLR 2027

ECHO: A Hybrid Neural Operator For Coupled Thermo-Hydro-Mechanical Modelling In Enhanced Geothermal Systems

Abstract

Coupled thermo-hydro-mechanical (THM) modelling is central to geothermal reservoir simulation, but resolving these interactions with high-fidelity numerical simulators is computationally expensive. Existing geomechanical surrogates reduce this cost but do not explicitly account for the distinct spatial and temporal behaviour of discrete fracture networks. We propose ECHO (Embedded Coupled Hybrid Operator), a hybrid surrogate that combines a Fourier neural operator (FNO) for reservoir-scale geomechanical prediction, a residual convolutional network for local near-well correction, and a graph-convolutional recurrent network (GCN–GRU) for fracture-specific dynamics. Evaluated on the Milford coupled reservoir–geomechanics dataset against a 3D U-Net, a full-grid GCN, DeepONet, a standard FNO, and U-FNO, ECHO obtains the lowest free-running fracture-response error in every evaluated operating and temporal regime, reducing elastic-dilation RMSE by 55.5–82.4% and fracture-normal-stress RMSE by 49.0–64.3% relative to the strongest competing architecture, while reaching on both targets. The evaluation considers held-out operating and material-parameter settings within a single fixed reservoir geometry and fracture network, and therefore measures interpolation within this geological configuration rather than generalisation to unseen geometries. Controlled ablations show that temporal recurrence is the most consistently beneficial architectural component, whereas graph connectivity and message passing contribute more conditionally across regimes. The complete surrogate runs in 6.6 ms per coupling transition, corresponding to a measured replacement-runtime ratio relative to the reference finite-element geomechanics solve, and is integrated with a commercial reservoir simulator through a two-way coupling interface. Within the evaluated setting, these results support representation-specific neural surrogates as an effective approach for accelerating coupled physical simulation when a domain contains sub-regions with qualitatively different spatial structure.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural operators, Fourier neural operators, graph neural networks, spatiotemporal modelling, enhanced geothermal systems, thermo-hydro-mechanical modelling

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_mW6VOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Thermo-hydro-mechanical simulations couple the heat transfer, fluid flow, and rock deformation changes in geothermal reservoirs. This paper introduces a thermo-hydro-mechanical emulator for coupled INTERSECT–VISAGE simulation. The paper proposes a Fourier neural operator to operate on the grid, then a CNN to correct the prediction near to well regions, and finally a standalone GNN to operate on the object-level in the reservoir to predict additional information. The method is evaluated on a dataset generated from INTERSECT–VISAGE simulation and compared to several implemented neural networks as baselines.

Critical strengths
  • The model achieved good performance in emulating with a significant gain in speed against the reference finite-element solver.
Critical weaknesses

This paper should be clearly rejected.

  • The method (to be honest, the whole paper) is very difficult to understand with many missing details e.g. the paper refers to the supplementary materials which do not exist, and it assumes that the reader has a perfect knowledge about thermo-hydro-mechanical without giving background or conceptual explanation about what the paper is aiming for. In summary, the paper is unpolished and imprecise and thus results are not reproducible.
  • The proposed method does not support a sufficient contribution to the field. Where is the novelty aspect of this work?
Questions for authors
  • Why did you choose Fourier neural operator?
  • Do you think the results presented are significant in comparison to FNO/U-FNO? And why grid GCN performs so badly? Given the model capacity, I do not think the comparison between the models is fair.
  • Can you clarify the novelty aspect of this paper? Is it only about that ECHO differs adds a fracture-specific prediction instead of having a one join model?
Reviewer_kd2YOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes ECHO, a hybrid neural surrogate for thermo-hydro-mechanical simulations in geothermal systems. It combines an FNO for reservoir-scale prediction, a residual CNN for near-well correction, and a GCN–GRU for fracture dynamics. Experiments on a fixed reservoir geometry with 16 simulation realizations report improvements over five neural-network baselines and substantial computational acceleration.

Critical strengths
  1. Meaningful application and promising results. The paper addresses the computational cost of coupled geothermal simulations and reports substantial improvements in fracture prediction accuracy over several standard neural architectures.

  2. Physically motivated architecture. The separation of reservoir-scale, near-well, and fracture-specific predictions is intuitively appealing. The ablations also provide useful evidence about the importance of temporal recurrence.

Critical weaknesses
  1. Limited novelty and insufficient methodological justification. The proposed architecture primarily combines three established networks without introducing a fundamentally new learning mechanism. Although component-wise ablations are provided, it remains unclear whether the three-network decomposition is necessary or whether a simpler architecture could achieve comparable performance. More importantly, the paper does not fully substantiate the central motivation for distinct spatial and temporal behavior between the reservoir and fracture network. The graph uses Cartesian-grid connectivity rather than physical fracture relationships, and graph message passing does not consistently improve performance. Similarly, the benefits of temporal recurrence do not establish that fracture dynamics fundamentally differ from reservoir dynamics.

  2. Insufficient experimental validation and comparisons. The baselines are restricted to several standard neural architectures, without comparisons to previous domain-specific surrogate methods or relevant state-of-the-art approaches. All experiments use a fixed geological configuration, and each operating-control family has only one held-out realization, substantially limiting the generalization claims. Moreover, the paper does not demonstrate the quantitative accuracy of the fully coupled simulator. The authors describe several experimental findings only briefly, without sufficient quantitative detail or visualizations. Furthermore, Table 2 selectively highlights ECHO's 4,107× runtime ratio, despite Vanilla FNO (4,753×) and DeepONet (4,164×) achieving better results. This presentation is misleading.

Questions for authors
  1. What evidence supports the claimed distinct spatial and temporal behavior of fracture networks? Why are three separate architectures necessary instead of a simpler unified model?

  2. How does ECHO compare with previous domain-specific or state-of-the-art geomechanical surrogate methods, beyond standard neural architectures?

  3. Where are the results of the claimed "software-level" evaluation, which is described in your contribution: "integrate the complete surrogate with a commercial reservoir simulator through a gRPC-based two-way coupling interface"?

Reviewer_CJ2GOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

ECHO is a surrogate for the VISAGE finite-element geomechanics step in a sequentially coupled INTERSECT–VISAGE thermo-hydro-mechanical workflow for enhanced geothermal systems. It has three parts:

An FNO predicts reservoir-wide increments of strain, effective stress and plastic strain from pressure/temperature increments and the current mechanical state. It keeps an XGBoost plasticity gate from a predecessor surrogate.A residual 3D CNN corrects the FNO inside a fixed near-well mask. A GCN–GRU over the 235 fracture-associated cells (six-face Cartesian adjacency) autoregressively predicts fracture elastic dilation, plastic dilation and normal stress. Monthly and yearly transitions are handled by separately trained models.

The data are 16 coupled simulations of a single 180×60×25 reservoir with one fixed fracture network, split 6/1/1 within each control family (BHP vs. rate). Each of the four evaluation regimes is therefore tested on one held-out trajectory.

Critical strengths

Within its narrow setting, the paper shows reasonable experimental hygiene. the paper has an explicit limitations section covering the fixed geometry, the single held-out realisation per family, the absence of closed-loop validation, and the non-hardware-matched timing. Integration with a commercial simulator has practical value for the authors’ workflow.

Critical weaknesses

1, very Limited methodological novelty and limited relevance to the ICLR audience. Every component is an existing method, used largely as published. FNO (Li et al., 2021). A residual CNN correction applied inside a hand-specified near-well mask. An XGBoost activation gate carried over from the authors’ predecessor surrogate. A GCN–GRU, which is a standard graph-convolutional recurrent architecture.

  1. The contribution is a domain decomposition: separate off-the-shelf models for the matrix and for the fracture cells. It is not a new learning principle, architecture or training method.

  2. The representational argument is weak. The fracture “graph” is built from Cartesian index offsets on the same grid. It carries no fracture geometry, orientation or transmissibility, so in effect it is a sparse lattice neighbourhood.

  3. The paper’s own ablations undercut the graph rationale. Removing message passing improves both targets under monthly rate control. The 18-direction graph is worse in every regime. Sparser stencils help only in some cases. The component that helps consistently is recurrence, which is generic rather than representation-specific.

  4. Nothing is released. Data, code, weights and simulators are all proprietary, and the evaluation uses a single proprietary configuration. The community cannot reproduce the work, compare against it or build on it. As submitted, it reads as an engineering case study better suited to a geothermal or computational-geosciences venue

  5. The evaluation is too narrow, and the benchmark design favours ECHO, so the results do not support the claims.

  6. Selective headline reporting. for example, Table 1 reports only ranges of relative reduction against the per-regime strongest baseline. The persistence, training-mean and ridge baselines are not quantified in the main text, and neither is the capacity-controlled CNN comparison.Additionally, the work compare A100 inference against a VISAGE solve on unrecorded hardware. The paper’s own end-to-end estimate is about 2.3×, meaning each surrogate-coupled run saves roughly 57% of a full run. That is a modest gain for a model that must be retrained for every new geometry.

Questions for authors
  1. Were the regime-specific choices selected on the validation realisation alone, or with test results in view?

  2. These are the rollout window, graph stencil, log-ratio target and cadence-specific models. Did the baselines get any comparable per-regime tuning?

  3. Please report baselines with recurrent memory, for example FNO or U-FNO plus a GRU or ConvLSTM, or the existing baselines trained with a fracture-weighted loss. If these close most of the gap, the representation-specific claim does not hold.

  4. give absolute per-regime errors for all models. This should include persistence, training-mean and ridge, and numeric results for the capacity-controlled fracture-masked CNN.

5 Each realisation has 22 states, but the monthly and yearly regimes contain only 11 + 8 transitions. Which transitions are excluded (for example, the cadence-change transition), and why?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.