A Large Open Multi-Energy Corpus of Soil Compaction Tests, with Machine-Learning Baselines
Sompote Youwai, Chana Phutthananon, Warat Kongkitkul
cs.LG
Quick verdict
Open 2,854-test compaction corpus with provenance-aware ML baselines; evaluation is refreshingly honest, but 96.5% standard-effort data thins the energy claim.
Summary
The paper assembles and releases an open corpus of 2,854 laboratory Proctor compaction tests from six public sources, spanning 162 provenance groups and four compaction energy levels, each record audited to the Proctor method its source names. Screening against the zero-air-voids condition removes 11.8% of 8,242 harmonised records (5.7% where specific gravity was measured), yielding the finding that a material share of published compaction data is physically impossible; the corpus-wide optimum degree of saturation is 0.815 (CV 11%). Predicting maximum dry density and optimum moisture content from six index inputs, TabPFN reaches R2 0.824/0.784 under random folds, 0.727/0.696 under provenance-grouped folds, and 0.520/0.614 with a whole source held out. Compactive energy is marginal under pooled ablation yet decisive on the 66 modified-Proctor records (R2 0.740 with, -0.651 without). Symbolic regression yields closed forms coupled through the phase relation, guaranteeing no predicted pair exceeds the zero-air-voids line at a cost of about 0.11 R2. Main limitations, largely self-declared: laboratory-only data, LTPP contributing 69% of records, all 101 non-standard records from a single deposit, no independent test set, and closed-form structure selected with sight of the evaluation partition.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- An open corpus of 2,854 laboratory compaction tests from six public sources, 162 provenance groups and four Proctor energy levels is released under CC BY 4.0 with per-record provenance, an order of magnitude larger than the datasets behind published correlations.
- Screening against the zero-air-voids condition removes 11.8% of 8,242 harmonised published records (5.7% of those with measured specific gravity), so a material share of published compaction data is physically impossible.
- The optimum degree of saturation over 2,854 tests has a mean of 0.815 (CV 11%), a corpus-level baseline that varies systematically by soil group (0.706-0.853) rather than a soil-independent constant.
- Provenance-respecting folds cost about 0.10 R2 versus random folds at matched training size, and compactive energy — negligible in pooled ablation — is decisive on modified-Proctor records (R2 0.740 with, -0.651 without).
- Closed forms coupled through the phase relation cannot violate the zero-air-voids bound (0 of 2,720 predicted pairs across all designs) at a measured cost of about 0.11 R2 under random folds.
Strengths
Corpus with per-record provenance enabling transfer-facing evaluation
2,854 tests from six audited sources (Table 3), 162 groups, four energies taken only from named standards and never inferred; the release includes query strings, screening decisions and source adapters, making the grouped (0.727/0.696 R2) and source-held-out (0.520/0.614) designs of Table 9 possible at all.
Leakage is measured, not asserted
Table 7 quantifies random-fold leakage (96.6% of records share a provenance group with training, 40.1% share an exact input vector); the duplicate-collapse control (2,126 records, -0.005 R2 MDD, +0.002 OMC) cleanly attributes the ~0.10 R2 random-to-grouped gap to shared provenance rather than duplication.
Physics screening with honest sensitivity analysis
Structurally admissible closed forms with a measured price
Exemplary limitation and reproducibility reporting
Weaknesses
The multi-energy claim rests on a sliver of data
Energy effect is confounded with source
Closed-form structure is data-selected
Pooled results are LTPP-weighted
Substituted specific gravity touches everything
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper makes three linked contributions: (1) an open corpus of 2,854 laboratory Proctor compaction tests pooled from six public sources under one harmonisation protocol, released under CC BY 4.0 with per-record provenance (162 groups) and four stated compaction energy levels (355/593/1,347/2,693 kJ/m3); (2) predictive baselines for maximum dry density (MDD) and optimum moisture content (OMC) from six index inputs (wP, IP, fines, sand, log-energy, Gs) under three fold designs; (3) closed-form expressions coupled through the phase relation so that no predicted pair can cross the zero-air-voids (ZAV) line. A side result — that 11.8% of 8,242 harmonised published records are physically inadmissible (94% above the ZAV line) — is arguably the most consequential claim for the wider literature. Scope is explicitly laboratory-only; field densities and curve points are outside the corpus, and this is restated in Table 19.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What data does this paper release?
A corpus of 2,854 laboratory Proctor compaction tests pooled from six public sources (the FHWA LTPP programme, two figshare deposits, three Zenodo deposits). Each record carries maximum dry density, optimum moisture content, index properties (plastic limit, plasticity index, fines, sand, specific gravity), the named compaction standard and a provenance identifier. It is released under CC BY 4.0 on GitHub together with all analysis code and per-record out-of-fold predictions.
How accurate are the predictions of maximum dry density and optimum moisture content?
Under the random cross-validation used in the prior literature, R2 is 0.824 for density (MAE 0.066 Mg/m3) and 0.784 for water content (MAE 1.87 percentage points) using the TabPFN foundation model. Under folds grouped by provenance the figures drop to 0.727/0.696, and with a whole data source held out to 0.520/0.614. The authors present the random figures as interpolation within the corpus, not transfer to new sources.
Why do the authors say many published compaction records are physically impossible?
A compaction state above the zero-air-voids line cannot exist. Screening 8,242 harmonised records against this bound removed 11.8% (94% of them above the line). Among records with a measured specific gravity the rate is 5.7%, and the overall rate ranges from 8.2% to 19.8% depending on the substituted Gs value, so the qualitative finding is robust even though the exact figure is convention-dependent.
Can I use these predictions to accept or reject construction fill?
No. The authors explicitly restrict the intended use to screening: ranking borrow sources, deciding which material warrants a Proctor test first, and setting the water content of the first specimen. A mean absolute error of 0.066 Mg/m3 is the same order as the margin a 95%-of-MDD specification leaves, so an acceptance decision would spend that entire margin on prediction error.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The six figures are present only as captions in the extracted text; their visual content (MDD-OMC plane, saturation distributions, Pareto fronts) could not be assessed, so figure-based judgments rely on the surrounding text and tables.
- The released GitHub repository (Sompote/open_compaction) was not inspected; availability and completeness claims are taken from the paper's Data Availability section.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.