Reviewer3 (AI review)

September 28, 2026

AI Review (reviewer3.com)

Comments (8)

1. Inconsistent country name 'israel' vs '.' — Editorial

seven entities contained at least one record and 42 contained none (Figures 5 and 6, Table 3). South Africa led with 40 si tes, followed by Zimbabwe with 31, Algeria and Egypt with 19 each, the Democratic Republic of the Congo with 12, Namibia with 10, Zambia and Eritrea with 8 each, Morocco and Sierra Leone with 6 each, Saudi Arabia with 5, . with 2 and Jordan with 1. Every s

There is an internal contradiction and placeholder error regarding the country name for the two sites 'King Solomon's Mine' and 'Timna Mine'. Table 3 lists these under the country name 'israel' (with 2 documented sites), whereas Appendix A lists them under the country name '.'. This placeholder '.' propagates into several text passages and figures throughout the manuscript, such as in Section 4.1 ('Saudi Arabia with 5, . with 2 and Jordan with 1'), the Abstract ('Africa, the Arab States and .'), Section 2.5 ('the member states of the Arab League and.'), Figure 5's caption ('Africa, the Arab States and ..'), and Appendix A's introduction ('Africa, the Arab States and .').

2. Missing regional entity name in scope definition — Minor

The regional module covers the sovereign states of the African continent, the member states of the Arab League and.

The definition of the regional module's scope is incomplete and contains a recurring typo/omission. In multiple places, the text refers to "Africa, the Arab States and ." or "the Arab League and.", leaving a blank space or a period where a third regional category should be. Given that Israel is included in Table 3 and Appendix A but is neither in Africa nor a member of the Arab League, it is highly likely that "Israel" (or a similar term) was intended but omitted during editing. The authors should complete this definition and correct the text throughout the manuscript.

3. Trivial decision algorithm output by design — Major

validated. In the present evidence state the gate returned no site with a permissible cooling estimate, a result about evidence readiness rat her than about whether the sites hold water. The principal barrier to

The workflow is designed such that any site without manual validation (E3) or physical measurements (E4) is automatically classified as Class C (unscorable). Because the authors did not perform or report any manual validation or physical measurements for the global dataset, it is mathematically guaranteed that every single site would return Class C. This makes the "gate test" result—that zero sites were eligible for cooling estimates—a trivial consequence of the study's incomplete execution rather than an empirical finding about the database or the presence of water. The authors should clarify that this outcome is a direct result of the validation step not being performed in this study.

4. Omission of automated satellite water detection counts — Major

The decision algorithm was run across all 115,216 sites in their current evidence state. Every site returned Class C, retained in the inventory but unscorable because validated water evidence was absent, and no cooling value was issued

Although Section 2.6 describes the satellite water detection workflow using the JRC Global Surface Water dataset in detail, the manuscript does not report any results from this automated screening stage (such as the number of sites with automated detections, or Grade E2). Simply stating that all sites returned Class C because they lacked manual validation (E3/E4) omits the valuable preliminary insights that the automated screening stage is designed to provide. The authors should report the counts and geographic distribution of automated water detections (E2) to demonstrate the functionality and output of this stage of the pipeline.

5. Unexplained missing sites in continental table — Major

The continental table covers 115,214 of the 115,216 sites and 101 of the 103 countries

Table 2 (Continental distribution) contains 115,214 sites across 101 countries, which is 2 sites and 2 countries short of the global total of 115,216 unique sites and 103 countries reported elsewhere in the manuscript. While the text notes that the continental table covers a subset of the sites and countries, it does not explain why these 2 sites and countries were excluded or could not be assigned to a continent. The authors should clarify the reason for this omission and, if possible, resolve the attribution of these remaining sites.

6. Corrupted country name in regional module — Minor

Leone with 6 each, Saudi Arabia with 5, . with 2 and Jordan with 1. Every s

The regional module contains a corrupted country/entity name represented as a period (".") in the text, figures, and tables (e.g., "Africa, the Arab States and .", "Saudi Arabia with 5, . with 2 and Jordan with 1", and in Appendix A). This appears to be a processing or spatial join error where the country name was not resolved or was corrupted to a period, which was then treated as a valid entity name. The authors should correct this geocoding/attribution error and replace the period with the correct country name (e.g., Israel, based on the coordinates of Timna Mine and King Solomon's Mine) throughout the manuscript, figures, and Appendix A.

7. Missing clustering sensitivity analysis results — Major

Sensitivity will be assessed at neighbourhood radii of 2, 5 and 10 km and minimum cluster sizes of 3, 5 and 10, reporting cluster count, clustered sites, isolated sites, screening units and whether any regional site enters a cluster. Until that analysis is reported, the regional isolation result is p

In Section 2.9, the authors state that DBSCAN clustering sensitivity will be assessed across multiple neighborhood radii (2, 5, and 10 km) and minimum cluster sizes (3, 5, and 10). However, the results of this sensitivity analysis are not presented anywhere in the manuscript. Instead, Section 7 lists "report the clustering sensitivity analysis" as a future step. The authors should either remove the description of the sensitivity analysis from the Methods section to avoid misleading the reader, or actually perform and report the sensitivity analysis results in the Results section.

8. Incomplete and inconsistent regional prose description — Minor

seven entities contained at least one record and 42 contained none (Figures 5 and 6, Table 3). South Africa led with 40 si tes, followed by Zimbabwe with 31, Algeria and Egypt with 19 each, the Democratic Republic of the Congo with 12, Namibia with 10, Zambia and Eritrea with 8 each, Morocco and Sierra Leone with 6 each, Saudi Arabia with 5, . with 2 and Jordan with 1. Every s

The prose description of the regional module in Section 4.1 is highly incomplete and inconsistent with Table 3. It lists a few countries (South Africa, Zimbabwe, Algeria, Egypt, DRC, Namibia, Zambia, Eritrea, Morocco, Sierra Leone, Saudi Arabia, ".", and Jordan) but completely omits 15 other countries that have documented sites in Table 3 (such as Madagascar, eSwatini, Tanzania, Ethiopia, Uganda, Tunisia, Sudan, Rwanda, Somaliland, Mauritania, Angola, Burkina Faso, Kenya, and Liberia). Furthermore, it refers to Israel as "." in the prose and Appendix A, but as "israel" in Table 3. The authors should revise the prose to accurately and completely summarize the regional findings and ensure consistency across the text, tables, and appendices.

xPeerd (AI review)

September 28, 2026

AI Review (xPeerd — DAReview)

Summary

This manuscript, titled Mining the archive: a global audit of inactive-mine records for siting water-cooled AI data centers, presents a data-audit and screening-workflow study rather than a validated engineering feasibility study. Its main contribution is to characterize the documentary coverage of the open WAPHA inactive-metal-mine dataset and to propose a reproducible evidence-gated workflow for later mine-water screening and cooling assessment. The paper is strongest where it is cautious: it repeatedly states that documentary coverage is not equivalent to physical resource availability, that no site has yet passed the persistent-water validation gate, and that no cooling capacity should be reported before benchmark calibration and manual validation

The manuscript’s central weakness is that its framing sometimes approaches an application paper on siting AI data centers, while its actual evidentiary contribution remains descriptive and procedural. The study reports 115,216 unique coordinates from the WAPHA deposit, strong geographic concentration in a few countries, regional counts for Africa and Arab states, and clustering-based reduction of screening units, but it does not yet validate water presence, legal usability, or cooling performance at any site The work is method-aware and generally honest about limitations, but several textual, formatting, and referencing defects reduce credibility. I do not see decisive evidence of fabrication; however, there are presentation anomalies, unresolved placeholders, inconsistent references, and some promotional wording that should be corrected before the paper can be considered mature.

Potential Major Revisions

1. Align the title, abstract, and contribution claims more tightly with what the study actually demonstrates.

The manuscript is careful in parts, but the title and framing may still over-signal practical siting utility relative to the evidence produced. The title promises “a global audit of inactive-mine records for siting water-cooled AI data centers” Yet the paper later states that “no site in this manuscript has passed the full persistent-water validation gate, and no cooling capacity is reported” and that “this is a screening and data-audit study; it does not recommend any site for construction” The conclusion is equally restrictive: “Until these steps are complete, no site should receive a numerical cooling estimate or a construction recommendation”

This internal consistency is commendable, but it creates a scope mismatch. The manuscript should be reframed more explicitly as a documentary-coverage audit and workflow specification, not as a siting study in the practical engineering sense. The abstract sentence “We specify an evidence-graded workflow that links mine records to satellite water persistence” is accurate, whereas the title’s “for siting” risks implying operational readiness that the results do not support I recommend revising the title, abstract opening, and highlights to foreground data auditing, inventory bias, and screening-governance rather than site selection.

2. Strengthen the novelty argument and distinguish the contribution from prior mapping and feasibility studies.

The manuscript does a reasonable job distinguishing itself from He et al. and from mining-footprint studies. It states: “This study does not claim a complete global map of cooling-suitable mine lakes. It asks a narrower and reproducible question” It also explains that He et al. is the benchmark for future model calibration rather than a directly transferable capacity factor, and Table 4 contrasts the present work with that prior feasibility case

However, the novelty remains somewhat procedural unless the authors sharpen what is genuinely new. Is the novelty the data audit of WAPHA for cooling-oriented reuse? The evidence-grade framework? The exclusion-first pipeline? The regional queue for Africa and Arab states? At present, several elements are assembled rather than empirically validated. The manuscript would benefit from a more explicit contribution statement, ideally in the Introduction and Discussion, separating: a) novel dataset auditing outputs, b) novel workflow design choices, c) non-novel but adapted components from existing hydrologic, geospatial, and mine-footprint literature.

Without that clarification, the paper may be read as an ambitious framing with limited validated findings.

3. Provide fuller methodological specification for reproducibility of the geospatial workflow.

The manuscript makes a strong reproducibility claim: “The full workflow was re-executed from a clean environment before submission” It also states that the repository records checksums, software environment, reverse-geocoding method, and model parameters This is a positive feature.

However, the manuscript text still lacks sufficient detail at several points for a reader to reproduce or critically assess the workflow without consulting external code. For example, Section 2.6 says the script extracts “water occurrence, recurrence, seasonality, persistent-core area, maximum water area and transition information,” but the threshold values that define “persistent,” “seasonal,” or “uncertain” are not given in the retrieved text The decision algorithm is conceptually described, but the precise operational thresholds, buffer sizes, and exclusion logic should be summarized in the manuscript itself, not only in code repositories. Similarly, Section 2.4 reports fallback geocoding for 519 sites but does not explain error propagation from nearest-populated-place assignment into later clustering or country totals

The authors should add a compact parameter table in the Methods covering:

• point-buffer size where polygon geometry is unavailable,
• persistence thresholds,
• exclusion intersection rules,
• reverse-geocoding fallback details,
• sensitivity-analysis outputs promised for DBSCAN but not yet reported.

4. Report the promised sensitivity analyses or moderate conclusions that depend on unreported parameter choices.

The manuscript openly acknowledges parameter dependence in clustering: “These are screening parameters, not natural definitions of mining districts” and “Sensitivity will be assessed at neighbourhood radii of 2, 5 and 10 km and minimum cluster sizes of 3, 5 and 10” Yet the Results section still foregrounds the “5.9-fold reduction in screening units” and the finding that “No regional site entered a cluster under the primary parameters”

Because the authors explicitly state that the regional isolation result is parameter-dependent, these findings should either be accompanied by the sensitivity analysis now or described more cautiously as preliminary. Otherwise, readers may overinterpret a screening convenience result as a meaningful spatial structure. The paper itself recognizes this danger: “it may instead reflect sparse sampling in the underlying record” That caveat is appropriate, but the manuscript should do more than caveat; it should show the sensitivity outputs or defer emphasis.

5. Clarify the status of the cooling model and avoid giving the impression of latent quantitative readiness.

Section 2.10 presents explicit heat-flux equations and allowable warming scenarios, including “6 K for the air-cooled case and 15 K for the liquid-cooled case” This can give the reader the sense that engineering estimation is a near-term output of the current study. Yet the same section correctly states: “The model will be calibrated against the published Morea case before any new site receives a numerical cooling estimate. Until calibration is documented, the tool is treated as uncalibrated and its outputs are suppressed”

This is prudent, but the manuscript should explain why the equations are included in full at this stage and what aspects remain uncertain. Are bulk transfer coefficients fixed or site-specific? What climate data source and temporal resolution are intended for future runs? How will uncertainty be handled when using satellite-derived water area but no bathymetry, water chemistry, or withdrawal constraints? As written, the engineering branch appears conceptually serious but empirically unready. The authors should state more clearly that this component is a provisional architecture for future validated use rather than a tested model in the present paper.

6. Address anti-fraud and coherence concerns raised by residual placeholders, promotional wording, and incomplete metadata.

I do not find strong signs of fabricated science in the core analytical logic. In fact, the manuscript is unusually explicit about what it does not establish, which argues against classic overclaiming. However, several anomalies weaken trust and should be corrected:

• The abstract includes a broken phrase: “Africa, the Arab States and . covered 69 entities”
• Section 2.5 repeats the unresolved scope phrase “the member states of the Arab League and.”
• The highlights include promotional and grammatically defective lines: “There is website for as engine for access easily for any mine at any country: www.mine2ai.com” and “For location of any mine copy the google earth link and paste at google earth website”
• Administrative statements appear as “Funding NO” and “Declaration of competing interests NO,” which is not journal-ready phrasing

These are not by themselves evidence of fraud, but they are characteristic of insufficiently controlled manuscript preparation and could trigger editorial skepticism. The authors need a thorough document audit before submission.

7. Expand the discussion of legal, environmental, and practical suitability beyond acknowledgement-level treatment.

The manuscript correctly states that “Mine-water reuse must account for acidity, dissolved metals, suspended solids, corrosion, fouling, discharge regulation and long-term liability” and that these concerns do not establish suitability for any particular site This is an important caution. But given the paper’s application framing toward AI data-center cooling, the discussion should do more to explain how these liabilities constrain the practical value of the proposed screening workflow. A site can be documented, geolocated, water-positive, and still be unusable because of hydrochemistry, permitting, remediation obligations, or social license.

The conclusions already distinguish five separate questions, including whether water is “legally usable” and whether a “validated engineering model supports the required cooling load” The manuscript should integrate these constraints more centrally in the Introduction and Discussion so readers understand that the current workflow is only the first gate in a much harder decision chain.

Potential Minor Revisions

1. Typographic, grammatical, and wording issues

The manuscript contains numerous language problems that should be corrected carefully:

• “There is website for as engine for access easily for any mine at any country” is ungrammatical and unsuitable for a scholarly paper, especially in the Highlights on page 1
• “For location of any mine copy the google earth link and paste at google earth website” is also ungrammatical, informal, and promotional rather than scientific
• “Africa, the Arab States and . covered 69 entities” contains an unresolved placeholder or deletion artifact in the abstract
• “The regional module covers the sovereign states of the African continent, the member states of the Arab League and.” repeats the same unresolved truncation in Methods
• “Where none is available, the workflow uses a prespecified point buffer, treated as lower-confidence geometry” is acceptable, but the next line in the manuscript excerpts shows occasional punctuation and spacing irregularities, suggesting copyediting is needed throughout
• “spatially compatible..” contains a doubled period in Section 2.6
• “A site is excluded when its geometry intersects an excluding polygon Each excluded site records...” is missing punctuation after “polygon”
• “The model will be calibrated against the published Morea case before any new site receives a numerical cooling estimate. Until calibration is documented, the tool is treated as uncalibrated and its outputs are suppressed.8” shows footnote or pagination collision formatting
• “The water-screening workflow uses the Joint Research Centre Global Surface Water dataset described by Pekel et al. [9], Earth Engine asset” appears incomplete, with the asset identifier omitted
• “The analysed file was the deposited inactive-mine spreadsheet its checksum is recorded in the analysis log” needs punctuation after “spreadsheet”
• “Each unique coordinate was assigned to a country by point-in-polygon join against the Natural Earth 110m boundary layer Points left unresolved...” is missing punctuation after “layer”

2. Formatting and presentation issues

• The manuscript duplicates “Highlights” on page 1, first as a numbered list and again as bullet points, creating redundancy and inconsistency
• The running title “Buried assets for artificial intelligence” feels rhetorically stylized and less precise than the main title; a more descriptive running title may better match the paper’s tone
• Page-flow formatting appears corrupted in several places, with page numbers intruding into sentences, for example “The principal barrier to 2 global mine-reuse screening...” and “supports the required cooling load.15”
• “Funding NO” and “Declaration of competing interests NO” should be replaced by standard declarative journal language such as “The authors received no specific funding...” and “The authors declare no competing interests”
• The appendix inventory is extensive, but the manuscript should state clearly whether coordinates are rounded consistently and whether all names are exactly as deposited or lightly cleaned; currently it says they are “reproduced without harmonisation,” which is helpful but may need stronger formatting guidance in the appendix header

3. Reference-list problems

The references contain inconsistencies and likely errors that need full cleanup:

• Reference numbering is inconsistent, with both “[6)” and “7.” forms appearing around the Tang and Werner citations
• A DOI appears malformed: “https://doi.org/10.1038/s43247-023-00805-618” likely contains an extra “18” at the end
• Placeholder-like entries such as “10- Natural Earth boundary layer,” “11- Google Earth Engine dataset page,” “12- WDPA,” and “-13UNESCO World Heritage List” are incomplete for scholarly referencing
• The manuscript appears to cite Tang and Werner twice in different formats, which may be legitimate if one is the paper and one the dataset, but the list should be normalized and checked carefully

4. AI content analysis

The paper discloses that “Artificial-intelligence tools were used to assist with code scaffolding, literature triage and language editing. The authors reviewed and edited the manuscript and take full responsibility for its content, accuracy and integrity” That disclosure is appropriate and preferable to non-disclosure.

Estimated percentage of AI-generated content: moderate, approximately 20–35%.

Rationale:

• The prose is highly uniform, cautious, and template-like in many methodological and discussion passages.
• Repetitive sentence structures such as “This study does not...”, “The present workflow...”, and “These results describe...” recur across sections
• The acknowledgments explicitly admit AI-assisted language editing and code scaffolding

Highlighted likely AI-assisted or AI-polished sections:

• The abstract and introduction contain highly standardized framing and repeated cautionary hedging, for example “This study does not claim a complete global map...”
• The limitations and conclusion lists are concise, orderly, and strongly formulaic, which is not problematic by itself but is consistent with AI-assisted polishing
• The numbered promotional highlights on page 1, by contrast, look less like polished AI output and more like manually inserted, poorly edited additions

Assessed epistemic impact: The likely AI involvement does not, by itself, undermine the manuscript’s core contribution. The larger concern is not authorship purity but editorial control. AI assistance appears to have produced relatively consistent expository text, yet human oversight seems weaker in the document-preparation layer, as shown by broken placeholders, malformed references, and promotional insertions. The epistemic risk is therefore moderate: not because AI seems to have invented results, but because manuscript assembly and quality control appear incomplete.

Recommendations

1. Recast the manuscript explicitly as a geospatial data-audit and evidence-gating paper, not a siting or feasibility paper.
2. Revise the title, abstract, and highlights to avoid implying that candidate sites for cooling have been validated.
3. Add a compact methods parameter table that includes all thresholds, buffers, fallback rules, and exclusion logic.
4. Either report the DBSCAN sensitivity analysis now or reduce emphasis on parameter-dependent clustering claims.
5. Clarify the engineering model’s provisional status and specify what remains to be calibrated before any meaningful cooling estimate is possible.
6. Remove promotional website language from the highlights or move any platform description to a neutral data-availability statement if truly necessary.
7. Conduct a full technical edit to resolve placeholders, punctuation failures, broken sentences, duplicate headings, and pagination artifacts.
8. Rebuild the reference list in a single journal-compliant format and verify every DOI, URL, and numbering sequence.
9. Replace nonstandard declarations such as “Funding NO” with conventional scholarly statements.
10. Add a short subsection on practical decision constraints beyond water detection, especially hydrochemistry, liability, permitting, and infrastructure compatibility.

Fraud / paper-mill warning: I do not find strong evidence of fabrication or paper-mill style pseudo-science in the analytic core. The manuscript’s logic is coherent and often conservatively self-limiting, which argues in its favor. However, the unresolved placeholders, reference corruption, and promotional inserts are serious credibility defects and should be corrected before submission because they could otherwise mimic low-quality or machine-assembled manuscript traits.

XSci (AI review)

September 28, 2026

Referee Report

Manuscript: "Mining the archive: a global audit of inactive-mine records for siting water-cooled AI data centers"

Recommendation: Major revision required.

Summary and Recommendation

The manuscript asks how much georeferenced information about inactive metal mines exists in an openly deposited global database, how unevenly it is distributed, and what evidence chain would be required before a legacy mine site could be credited with a data-centre cooling capacity. Motivated by a recent Pennsylvania study that modelled an abandoned mine pool as a cooling resource, the author analyses the Water and Planetary Health Analytics global metal mines deposit, reducing 150,309 inactive-mine records to 115,216 unique coordinate pairs across 103 countries. Coverage proves extremely concentrated: the United States and Australia hold 86.6% of records, four countries hold 95.8%, ten hold 98.0%, and 54 countries hold fewer than ten each, while Africa contributes 184 sites, or 0.16%. A regional module spanning 69 entities in Africa and the Arab States documents 192 sites, with 42 entities holding none. Density-based clustering yields 2,546 groups and 16,962 isolated sites. A six-stage evidence-graded pipeline is specified, and a gate test across the full record returns the unscorable Class C for every site, so no cooling estimate is issued. The author concludes that documentary coverage, rather than physical geography, is the binding constraint on global mine-reuse screening.

There is real merit here. The refusal to publish capacities ahead of validated evidence, the repeated insistence that grey map cells denote missing records rather than mine-free ground, the candid disclosure that the deposit omits several major national and proprietary sources, and the publication of the full 192-record regional inventory all reflect an author who takes seriously the difference between a dataset and the world it describes. The reported arithmetic is internally consistent: country shares, continental totals, cluster counts and the 5.9-fold reduction all reconcile against the tabulated values. Against this, three groups of problems stand in the way of publication. The manuscript delivers considerably less than its methods and title promise, and its central conclusion is not tested against the alternative it most needs to exclude. The source records appear, on inspection of the author's own appendix, to lack the positional accuracy, entity resolution and status validity the intended application requires. And the screening instrument itself contains design gaps that would persist even once the pipeline is run. Major revision is required.

Major Concern Category 1: What Is Specified, What Is Delivered, and What Is Published

1.1 Most of the described methodology has not been executed

Of the six stages set out in Section 2, only ingestion with deduplication and country attribution appear to have been run. The water screening of Section 2.6 produced no reported detection; the exclusions of Section 2.8 produced no reported count; the cooling model of Section 2.10 is explicitly suppressed pending a calibration that has not occurred; and the clustering sensitivity analysis is deferred. The tense makes this plain, with the methods stating that results "will be reported," that sensitivity "will be assessed," and that every regional detection "will be reviewed," while Figure 1 refers to a stratified validation of 300 detections that appears nowhere in the results. Section 4.3's gate test is accordingly a test of the gate rather than of the sites, which the author states honourably, but the consequence is that the evidentiary apparatus claimed as the paper's methodological contribution has never been exercised against data. What remains is a deduplication and a country join — useful curation, but not a global audit for siting water-cooled data centres.

1.2 The coverage conclusion is not tested against the alternative it most needs to exclude

Because "documented" is operationalised throughout as "present in this deposit," the conclusion that documentary coverage is the principal barrier is close to definitional. The contestable claim — that documentation is genuinely absent rather than merely unconsolidated — goes untested, and Table 1 supplies internal evidence pointing the other way. The United Kingdom appears with 126 records and Canada with 327, figures difficult to reconcile with jurisdictions maintaining extensive public abandoned-mine registers, and Section 2.2 itself names the BGS BritPits database as a source the deposit does not reproduce. The distinction matters because the remedies diverge: a consolidation gap calls for harmonisation across existing sources, whereas a documentation gap calls for national inventory construction, which Section 5.4 declines to cost. Ingesting the openly available USGS Mineral Resources Data System and reporting how far the country distribution shifts would settle this directly, and relating national counts to mapped mining area from the footprint datasets already cited would identify which countries are genuinely under-documented relative to their physical endowment.

1.3 The link from a mine coordinate to a viable data-centre site is assumed rather than analysed

The introduction frames AI computing as a planning problem involving electricity supply, land, grid connection and heat rejection, and the analysis then screens on heat rejection alone. No variable relating to interconnection capacity, transmission proximity, fibre availability, latency, land tenure or ambient wet-bulb temperature enters the workflow, and the suitability index sketched in Figure 1 comprises only water and exclusion terms. The caveat that water presence does not establish suitability is stated but never operationalised, leaving the relationship between the candidate list and deployable capacity unknown — a concern sharpened by the regional module's concentration on jurisdictions where grid reliability rather than cooling architecture is likely to dominate siting, and where hyper-arid conditions make evaporative rejection least appropriate. Intersecting the documented sites against open transmission and fibre layers, and reporting what fraction lie within a stated distance of each, would let a reader judge whether cooling-water screening is a binding or a slack constraint.

1.4 Provenance commitments are undischarged and the published queue raises unaddressed governance questions

Section 2.11 undertakes to record the source file name and checksum, the boundary-layer version, the Earth Engine asset, exclusion-layer versions and the software environment. None appears: the checksum is said to be logged but not given, the asset identifier is missing mid-sentence, no layer versions are supplied, the Data Availability statement ends without naming a repository, and the associated web tool is unversioned, unlicensed and unarchived — while Table 4 identifies a reproducible evidence workflow as a primary contribution. A related question concerns what the paper publishes. Appendix A describes itself as a screening queue and lists 192 named, geolocated sites in low-capacity and water-stressed jurisdictions, among them Kabwe, Shinkolobwe and Manono-Kitotolo, each carrying well-known contamination, radiological or tenure histories. The exclusion architecture screens conservation values only, with no consent, tenure, livelihood or benefit-sharing dimension, and the scope rule takes implicit positions on contested sovereignty on purely cartographic grounds.

Major Concern Category 2: The Measurement Properties of the Source Records

2.1 Positional accuracy appears incompatible with the primary detection threshold

The primary threshold of 0.01 km² corresponds to a feature 100 m on a side, which Section 2.6 concedes sits near the detection floor of a 30 m product, and for most sites the interrogated geometry will be a point buffer whose radius is never stated. Appendix A suggests the coordinates cannot support that precision: fractional parts cluster strongly on values such as .001, .016, .501 and .999 — Mangula at −16.997, 30.001, Pickstone at −17.997, 30.001, Aouli-Mibladen at 32.999, −4.665, Asmaradebara at 15.000, 38.751 — a pattern consistent with capture in whole or half arc-minutes and therefore with positional granularity of roughly one to two kilometres. A buffer robust to that error would routinely enclose unrelated drainage and impoundments; a buffer tight enough to be specific would routinely miss the target. Section 4.4 notes that the Timna record lies roughly 180 km from its named valley, and the record "Aberketieb Red Sea Hills" is placed near Khartoum. Plotting the distribution of coordinate fractional parts, and georeferencing a stratified sample against high-resolution imagery, would yield the error distribution from which a defensible buffer radius and an honest minimum detectable feature size could be derived.

2.2 The site unit is not well defined

Deduplication by exact coordinate is acknowledged not to merge distinct names referring to one working, but the limitation is treated as residual when the appendix suggests it is systematic, and the resulting count of 115,216 propagates into every percentage, every cluster statistic and the headline workload reduction. Broken Hill District and Kabwe share a longitude and lie under two kilometres apart, Broken Hill being the historical name of the Kabwe mine; Empress and Empress Mine lie about five kilometres apart and share a commodity pair; Matchless and Matchless Mine, Shinkolobwe and Shinkolobwe Signal, Caesar 2 and Caesar 3, and the Iron Sides, Iron Ton and Magazine Hill triplet present the same pattern. A prior difficulty is that the records are not one kind of object: the appendix mixes discrete workings, undeveloped prospects, district-scale entities, alluvial stream localities and two entries named simply "Unknown," which are not interchangeable for a screen requiring a centroid and a void. A record typology over the regional set, and a proximity-and-name-similarity rule yielding a sensitivity band around the site count, would both be tractable.

2.3 The sole inclusion criterion is unvalidated

The inactive status flag defines the study population, underwrites evidence grade E1 and serves as one of four controls against false-positive detections, yet no validation is reported and no procedure for confirming inactivity is specified. Appendix A raises a concern: the Kansanshi record corresponds to a Zambian copper-gold mine that produced 181 kt of copper in 2025 and has just completed a major expansion, and the Jabal Sayid record corresponds to a Saudi copper mine in commercial production since July 2016. This reflects no fault in the depositors, whose database was assembled to assess river-system impacts, where operational status is secondary; it bears on the present study, which repurposes the flag as its inclusion criterion. If producing mines are present, detected water at such sites is more likely process or tailings water than a mine pool, and the sites are in any case unavailable for redevelopment. Stating how the depositors defined inactivity, and checking the 192 regional records against national cadastres or recent imagery, would let a misclassification rate be reported.

Major Concern Category 3: The Design of the Screening Instrument

3.1 The energy balance is incomplete and no water balance is closed

The cooling model comprises sensible, latent and net longwave terms summed over pool area. Absorbed shortwave radiation, typically the dominant input to an open water surface and very large in the arid settings dominating the regional module, does not appear, so the balance is not closed and the unloaded equilibrium temperature cannot be recovered from the stated equations. More consequentially, the formulation computes total surface loss at the allowable temperature rather than the increment above the unloaded state, crediting the facility with heat the pool sheds regardless; in dry air the baseline latent flux driven by the vapour-pressure deficit can plausibly exceed the load-driven increment. The latent term also implies consumptive evaporative loss, nowhere converted into a consumption rate or tested against site inflow, although a shrinking pool falls below the very area threshold the model depends on. Water quality is likewise absent: Section 5.5 correctly names acidity, dissolved metals, fouling and discharge liability, but none enters the model, the grades or the exclusions, and grade E4 is defined by flow quantity rather than quality.

3.2 The coupling between facility and pool, and the governing coefficients, are unspecified

Two allowable warming scenarios of 6 K and 15 K are evaluated, but these are coolant-side values and no relation is given between the coolant circuit and the pool surface temperature the equations take as input. There is no heat-exchanger effectiveness, approach temperature, distinction between open-loop abstraction and an immersed closed loop, or pumping term, so the controlling input is undefined; the label "air-cooled case" compounds this, since a conventionally air-cooled facility rejects heat to the atmosphere. The pool appears only as an area, with no depth, volume, residence time or stratification, although most pit lakes are strongly stratified and only the epilimnion participates in surface exchange on operational timescales. The bulk transfer coefficients, emissivity, wind reference height, effective sky temperature and meteorological source are all unstated, and no sheltering correction is applied despite a pit lake sitting where near-surface wind is substantially reduced. The underground branch, meanwhile, is unreachable for every record, since the source file contains no hydrological field.

3.3 The evidence framework is under-specified and has produced a degenerate result

Only the area threshold is reported; the occurrence and recurrence values determining every automated classification are not, and no confirmation protocol, reviewer count, imagery requirement or agreement statistic is given for grade E3, while the stratified validation referenced in Figure 1 is not performed. More structurally, Class C merges three distinct epistemic states — never screened, screened and null, and detected but unvalidated — precisely the conflation a framework designed to separate knowledge from inference ought to avoid, and one that prevents the manuscript from reporting how many sites occupy each. The validation design is also asymmetric: four controls guard against false positives while nothing validates a non-detection, so sites missed by a screen biased toward small and shaded water bodies are assigned Class C and never revisited. Tabulating the thresholds, splitting Class C, specifying the validation protocol and adding a negative-control review of non-detections would let the framework demonstrate discriminating power on the regional subset.

3.4 The clustering claim does not follow, and the exclusion logic is geometrically under-defined

Screening cost is incurred per geometry, since each site requires its own buffer, extraction and manual review, none of which is shared across cluster members; the reduction from 115,216 to 19,508 units therefore records a change of counting convention rather than of analytical effort, and a cluster of 4,851 sites cannot be screened once. Applying a constant 5 km radius to a point process the manuscript itself argues is shaped by documentation effort makes the finding that no regional site clusters close to inevitable, and the Figure 7 legend defines a smallest class of three to ten sites although the stated minimum is five. On exclusions, every outcome depends on the unreported buffer radius; a binary intersection test misses the proximity pathway that published evidence indicates exceeds direct overlap; the justifying statistic of 9.5% is areal yet applied to points; and applying exclusions before grading renders unobservable the quantity an audit most wants, namely how much apparent resource is lost to designation. Converting exclusions from filter to flag, with distance bands, would preserve it.

Minor Issues

1. The abstract terminates mid-sentence at "The principal barrier to," and a bare full stop replaces an entity name there, in Section 4.1, in the Figure 5 caption and in two Appendix A records, although Table 3 names it. The abstract also calls the regional inventory "complete" in a paper arguing that the record is incomplete; presumably completeness of scope is meant, and saying so would remove an apparent contradiction on first reading.

2. The Figure 5 legend reports 37 in-scope entities without a record, whereas Section 4.1 and the list beneath Table 3 give 42, which reconciles correctly against a 69-entity scope holding 27 entities with records. If the difference arises from the five states lacking a boundary polygon the caption should say so; otherwise one count is in error.

3. The reference list needs a full pass. Tang and Werner appears three times under three numbers in two formats, with the text citing two of them for the same work; entries ten through thirteen are name-only stubs; no entry exists for the Ramsar List despite its use as an exclusion layer; and a digital object identifier fragment is fused into the Appendix A heading.

4. The title page names one author while the Author Contributions statement credits a second individual with validation, data curation, and review and editing. Authorship and affiliations should be reconciled before an editorial decision, since the statement implies a collaboration the byline does not reflect.

5. Two non-identical Highlights blocks appear, one carrying a garbled promotional line for a web tool that is not described in the Methods, not covered by the availability statements and not archived; another highlight states a design policy rather than a result. Any interest in that tool should be declared alongside the competing-interests statement, which presently reads only "NO."

6. The cooling model's governing parameters are absent — bulk transfer coefficients, emissivity, wind reference height, effective sky temperature and meteorological source are all unstated, air density is rendered as "pa" in the variable list, the reporting unit is not fixed between watts and megawatts, and no uncertainty propagation is described despite the promised paired confidence.

7. The third evaluation scenario, "the hottest month in the climate record," identifies neither dataset, period, resolution nor variable, though wet-bulb rather than dry-bulb temperature governs evaporative rejection. The 6 K and 15 K values are unsourced, and thermodynamic allowability is not distinguished from ecological or regulatory permissibility.

8. Section 3.4 reports that the continental table covers 115,214 of 115,216 sites and 101 of 103 countries without naming the omissions, which matters where 54 countries hold fewer than ten records. The convention admitting non-United-Nations-member entities should be stated for the global totals, and the promised sensitivity analysis excluding the 519 fallback-assigned sites reported.

9. Country attribution rests on a boundary layer generalised for very small-scale display, a weak basis for assignment near coasts and borders, and the dependency is flagged without being bounded. The one-degree binning is not equal-area, so cell ground area varies roughly twofold across the latitudes holding the densest clusters, and Figure 4 is captioned but never described.

10. A production pass would help. Tables 1 and 3 are fragmented across blocks with duplicated headers, and Appendix A is split across five blocks in an order neither alphabetical nor by count, impairing its use as the stated queue. Section 6 is a bare list restating methods text, and the keywords and running title promise cooling capacities and Earth Engine outputs that are not reported.

Missing and Inadequately Integrated Literature

Cited but Inadequately Integrated

Four cited works are used more narrowly than they might be. He et al. [2] supply the 0.01 km² threshold and the capacities of at least 9.5 MW under air cooling and 30.6 MW under liquid cooling, but the manuscript passes over their statewide screen, in which only a small fraction of an inventory of 5,586 abandoned mine lands cleared the area threshold — an empirical base rate that would let the author state an expected yield in advance — and over their finding that wind speed and allowable coolant temperature dominate capacity, precisely the parameters the present model leaves unspecified. Macklin et al. [4] are cited for the extent of river and floodplain impact, yet their modelling of where mine-derived contamination propagates would give the exclusion architecture the environmental dimension it lacks, and their assembly of a compilation substantially larger than the inactive subset analysed here demonstrates that cross-source consolidation is achievable at global scale.

Maus et al. [5] and Tang and Werner [6,7,8] are correctly distinguished as different objects from point records, but that caution forecloses analyses it does not rule out: intersecting the mine points against these polygon layers would furnish an independent positional check, identify which records correspond to features large enough to hold a screenable water body, and supply the polygon geometry Section 2.6 contemplates substituting for buffers, while their national breakdowns are the yardstick needed to separate under-documentation from under-ingestion. Finally, Pekel et al. [9] are cited as the water source without discussion of the product's documented behaviour over small, turbid, shaded and seasonal bodies, which describes mine pools closely; their transition layers, distinguishing new from lost permanent water, also offer a data-driven substitute for the closure dates whose absence is given as the reason an emergent-water classification cannot be applied.

Truly Missing Papers

Several directly relevant works are absent. On the demand side, Mytton (2021) quantifies data-centre water consumption and documents that fewer than a third of operators measure it at all, supplying both the accounting framework for the consumptive loss implied by the model's latent term and the baseline against which mine-pool cooling would have to be judged water-advantaged; Li, Yang, Islam and Ren (2023) quantify the operational and embodied water footprint of artificial-intelligence workloads and show that identical computation carries very different water costs depending on where and when it runs, which is exactly the spatial-substitution logic the manuscript invokes. On the supply side, Carcamo-Medel, Fuentes and Narsilio (2024) is the closest published analogue outside the Pennsylvania case, assessing decommissioned open pits as geothermal heating and cooling sources for nearby population centres, treating pit temperature as the controlling economic variable and considering exchangers embedded in existing mine infrastructure — the interface the present model leaves undefined. Loredo, Roqueñí and Ordóñez (2016) is the standard review of flow and heat-transfer modelling in flooded mines, identifying finite recharge and finite heat capacity as the principal constraint on sustained exchange.

Three further works bear on the data and screening design. Blanchette and Lund (2016) establish pit lakes as a global post-mining legacy, document their stratification and water quality, and identify which post-mining configurations give rise to persistent lakes — the domain knowledge needed to justify persistence thresholds and to decide which record types belong in a screening queue. Werner and colleagues (2020) delineate 3,736 features across 295 mines, separating pits, waste rock dumps, tailings facilities and water storage ponds, a typology bearing directly on entity resolution and on interpreting any detection within a mine buffer. Durán, Rauch and Gaston (2013) compute protected-area coincidence at the level of individual mine locations, reporting roughly 7% direct overlap and a further 27% within 10 km, which is the appropriate reference expectation for point-based exclusions and argues for distance bands rather than a binary test. The clustering algorithm at the centre of Section 2.9 is also applied without citation to Ester and colleagues (1996), whose treatment of parameter selection under non-uniform density speaks to the constant-radius concern. All of the above were verified via WebSearch except the last, which is foundational.

Blanchette, M.L., and M.A. Lund. 2016. Current Opinion in Environmental Sustainability 23:28-34.
Carcamo-Medel, M., R. Fuentes, and G. Narsilio. 2024. Communications Earth & Environment 5:292.
Duran, A.P., J. Rauch, and K.J. Gaston. 2013. Biological Conservation 160:272-278.
Ester, M., H.-P. Kriegel, J. Sander, and X. Xu. 1996. KDD-96 Proceedings:226-231.
Li, P., J. Yang, M.A. Islam, and S. Ren. 2023. arXiv:2304.03271; Commun. ACM, doi:10.1145/3724499.
Loredo, C., N. Roqueni, and A. Ordonez. 2016. International Journal of Coal Geology 164:115-122.
Mytton, D. 2021. npj Clean Water 4:11.
Werner, T.T., G.M. Mudd, A.M. Schipper, M.A.J. Huijbregts, L. Taneja, and S.A. Northey. 2020.
  Global Environmental Change 60:102007.

Conclusion and Path Forward

The question this manuscript asks is timely and worth asking, and the author's instinct for epistemic restraint is the right foundation for answering it. Suppressing capacity estimates until evidence is validated, insisting that empty map cells denote missing records rather than mine-free ground, disclosing the deposit's ingestion limits, and publishing the regional inventory in full are practices that should survive into any revision. The difficulty is that the manuscript promises an audit and delivers a tabulation; that its central conclusion about documentary coverage restates its own scope rather than testing it; that the source records appear to lack the positional accuracy, entity resolution and status validity the intended application requires; and that the screening instrument contains physical and procedural gaps that would remain even once the pipeline is run.

A coherent revision is available and requires no new data collection. Executing the workflow on the 192 regional sites would convert the evidence framework from proposal to demonstrated instrument and yield a non-degenerate grade distribution; ingesting one additional open source and reporting how far the country distribution shifts would turn the coverage claim into a finding; and a positional-accuracy assessment, a record typology, a status check against national cadastres and the deferred clustering sensitivity grid would together produce the measurement-led audit the title promises. On the model, closing the energy balance, defining capacity as an increment above the unloaded equilibrium, tabulating every coefficient with a source, specifying the exchanger interface, and reporting evaporative consumption and water chemistry alongside any capacity would make the instrument implementable and checkable. Depositing the derived outputs under a citable identifier, adding a governance and equity subsection stating plainly that the regional list is a research prioritisation rather than a development recommendation, and reconciling the authorship would complete the picture. Should the full programme prove impractical in one revision, reframing the work explicitly as a data descriptor and pre-registered protocol would also yield a publishable and genuinely useful contribution.