- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies recurrent sparse-field reconstruction when sensor locations are uncertain and only some vector components are observed. It introduces a cross-fitted, calibrated auxiliary location-weight generator, inserts its probabilities consistently into a shared kernel correction, and studies stage assignments in which the original generator A or new generator B is used for model/configuration selection and final fitting.
Critical strengths
- The distinction between location-probability quality and final-field quality is conceptually important.
- Cross-fitting seems to be useful in addressing the reuse bias.
Critical weaknesses
- The public neural comparisons assess complete systems and cannot isolate the contribution of auxiliary location inference, since the adapters do not share AB’s kernel correction and selection machinery. If I understand it correctly, the paper partly addresses this with matched reference-weight and GP-moment comparisons, plus a limited conditional-mixture diagnostic and a NIGP-inspired adaptation. I would like a fuller recurrent comparison against averaging per-location reconstructions, and a comparison that changes only the uncertain-input treatment while holding the downstream solver and selection procedure fixed.
- The evidence for selecting AB over BB is weak. AB was chosen using development-exposed cylinder cases, and its small advantage over BB in the original window reverses in the later, nonoverlapping window: BB wins five of six case–backbone groups, and AB fails the paper’s preregistered later-window criterion.
- The probability-to-field analysis validates how changes in location weights propagate through the reconstruction, including tests of frozen response-accuracy predictions. It would be more useful if the authors can validate a prospective prediction of reconstruction benefit. The failure explanations rely on target-assisted diagnostics, and the paper offers no demonstrated observation-only criterion for deciding when AB will outperform the alternatives. This limits the analysis’s value for choosing a method on a new task.
Questions for authors
- Where is the matched baseline that exactly or by Monte Carlo marginalizes over the supplied discrete location support while sharing AB’s references, kernel, candidate grid, selector, closure, cap, and recurrence?
- What auxiliary-inference claim, if any, is supported by the gaps to RecFNO, FLRONet, Senseiver, and PalSB? These are complete-system comparisons with different machinery and training assets, and the public adapters’ convergence sensitivity remains unresolved.
- What are confidence intervals when physical cases or trajectories are the sampling units, without treating horizons, observation seeds, backbones, conditions, or overlapping windows as independent?
- How much selection optimism is introduced by choosing among 109 configurations from sparse held-sensor scores, and where is a nested-selection or multiplicity-matched control?
- On a physical panel untouched during method development, how does frozen AB compare with its matched alternatives? Separately, how does it behave, or detect failure, when the true sensor site lies outside the supplied candidate support?