acceptodds
Withdrawn from ICLR 2027

When Is Source Information Worth Retaining? Evidence-Gated Selective Transfer for Sparse Geothermal Profiles

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Sparse geothermal temperature logs often leave long depth intervals unobserved, making profile reconstruction difficult when only a few target measurements are available. Target-only methods avoid cross-well mismatch, but sparse anchors often constrain the large-scale trend more strongly than the nonlinear structure inside a wide gap. Source-assisted methods can recover some of that structure from other wells, but only when source and target profiles are sufficiently compatible. With few target anchors, transfer can be selected too optimistically if the same observations are used both to calibrate a source correction and to judge its reliability. We introduce Evidence-Gated Selective Transfer (ESGT), which separates these two roles. Geometry-matched inner pseudo-gaps estimate the correction magnitude, outer held-out anchor blocks evaluate the calibrated correction without reusing their labels, and a detrended source-target shape statistic provides an additional compatibility check. The source correction is retained only when these signals support transfer; otherwise, ESGT uses a target-only reconstruction. We evaluate ESGT under controlled structural shifts and on measured geothermal profiles, with additional report-derived profiles prepared through lightweight multimodal indexing. Across the controlled scenarios, ESGT reduces RMSE from 0.876 for the strongest selective baseline to 0.822 while substantially reducing harmful transfer under source-target mismatch. On the Montana field data, it achieves an RMSE of C. Separating calibration from retention is most useful when source relevance varies across targets.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
multimodal indexing, contrastive hashing, geothermal profile reconstruction, selective transfer learning, multi-source regression, negative transfer, Gaussian processes

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_WCYVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript developed a transfer learning method, to reconstruct target geothermal profiles that have masked intervals, with the correction of source geothermal profiles. This method selectively retains source correction using evidence based on outer pseudo-gap blocks, and refit calibration on all anchors to predict target profiles. This method was compared with baselines including a shape-preserving interpolation method (PCHIP), several Gaussian-process regression methods without transferring (e.g., Target GP, Trend GP), calibration / gating (e.g., TCMSStack, gain calibration, shape-gated calibration), the proposed ESGT achieved best overall performance in a controlled simulation comparison, and two real-world datasets. Experiments also show that ESGT has less negative transfer compared with gain calibration. Ablation studies demonstrate the effectiveness of geometry weighting, shrinkage, and shape similarity in ESGT.

Critical strengths
  1. The proposed ESGT addresses a practical problem of reconstructing missing intervals in sparsely observed geothermal profiles. The architecture is easy to understand, have answered the question of when source information remains useful under limited target supervision, and demonstrated to be effective across regimes (i.e., shared structure, independent structure, and near-linear) by empirical experiments.
Critical weaknesses
  1. Lack Montana well-level performance disclosure. The Montana evaluation contains 540 feasible episodes. Did these episodes build evenly on target wells, or skew toward some target wells? Does ESGT performs better in most target wells, or concentrate on some wells that have more episode? Because multiple gap and anchor configurations can be constructed from the same well, episode-level metrics are likely correlated. Thus, reporting well-level performance and analysis uncertainty is beneficial.

  2. The rational of selecting baselines is not explained enough. For example, why did the authors use TCMSStack with fixed kernel hyperparameters instead of optimized hyperparameters, like the original TCMSStack did? Is this a fair comparison? Why not select the transfer-kernel learning method (Wei et al., 2023) mentioned in Related Work? Why not include geothermal and well-log imputation and reconstruction models? If logs with missing intervals can be well reconstructed by those models without transfer learning, it would be important to examine whether the proposed ESGT provides additional benefits over existing domain-specific approaches.

Questions for authors

For Montana, how the target wells and source wells were selected? For example, according to line 161, the four source wells were selected by “the similarity of their observed physical depth spans to the target span”. What’s the similarity here?

Reviewer_1iF7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Authors are looking at a problem with extremely limited amount of data points, well temperature estimation using only the depth as a feature. In particular they are interested in temperature data gap filling for a particular well.

The basic method would rely on a sort of regression/forecasting model, authors are looking into methods to increase the prediction power by looking at similar wells.

They build 2 models: (1) using the wells own data; (2) using the similar wells data. They propose an ensemble model that selectively decide whether to use the second model or not. The selection is mainly based on a heldout dataset for each well, where they assess the positivity of using the external well prediction model.

Overall it seems a sort of nested cross-validation idea. I would say its borderline, and here are a couple of comments for the authors:

  1. rA, relying on fixed thresholds is extremely unreliable given that geological data are clearly diverse; A better way would be to rely on a learned weighted model;
  2. They are using Gaussian processes, I would report those uncertainty;
  3. (minor) Very simple baselines (all the representation learning models were discarded);
  4. Given the amount of data, and the nature of the problem relying only on target and similar having similar shapes might produce models that can't generalize ;
  5. Evaluation rely on fixed missing gaps, and fixed center; I would recommend here to analyze a large set of wells missing data, look at how they are distributed and rely on that to build a strong experimental framework. It would be also great to take an external wells data (different than Montana) and verify the model effectiveness.

Overall it seems a promising idea, but it would need a little bit of additional work to be ready.

Critical strengths
  1. A creative way to build an ensemble model that can "decide" when to use the other source of data vs not use them.
Critical weaknesses

The main weaknesses are in the evaluation methods, they overly here on fixed centered missing points and grids, basic baselines. The other is on the model generalization.

Questions for authors

(related to rA) Why did you choose to use a hard filter here?

Reviewer_enbvOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes Evidence-Gated Selective Transfer (ESGT) for reconstructing missing intervals in sparse geothermal temperature profiles. It combines a target fallback with a calibrated source correction. Inner pseudo-gaps estimate the correction magnitude, while outer validation blocks and detrended shape similarity determine whether to retain it. Experiments on simulated and measured profiles show improvements in reconstruction accuracy and reductions in harmful transfer.

Critical strengths

The method separates correction calibration from source retention and maintains an explicit target fallback. This provides a clear formulation for exploiting useful source information without requiring transfer in every episode.

Critical weaknesses
  1. In my opinion, this manuscript does not yet meet the standards for submission; every figure shows clear, unremoved traces of AI assistance. These AI-generated artifacts are both unnecessary and glaringly obvious.

  2. There are many meaningless subsections and passages, and the symbols used in the formulas are left unexplained. These parts require careful review.

  3. Algorithm 1 selects the fallback before the outer loop and subsequently states that it is refitted. It does not explicitly distinguish refitting parameters from repeating model-family selection. If the family remains selected using all anchors, the outer validation labels could still influence the evaluated predictor.

  4. With only five target anchors, nested removal can leave very few observations for inner calibration. The pseudo-gaps also overlap. Sections 4.3–4.4 do not report feasible-block counts or variability in the resulting gains, leaving the stability of the retention decision under the sparsest setting unclear.

  5. The literature comparisons use a fixed-kernel TCM S Stack adapter and a MAP plug-in version of RECaST. These are explicitly restricted implementations, and their kernel or fitting sensitivity is not established. Their results support comparisons with these particular configurations more strongly than broad conclusions about competing transfer methods.

Questions for authors

Please refer to weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.