kth.sePublications KTH
Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Compaction quality assessment based on machine learning with small datasets
KTH, School of Architecture and the Built Environment (ABE), Civil and Architectural Engineering, Soil and Rock Mechanics.ORCID iD: 0000-0003-1927-6034
KTH, School of Architecture and the Built Environment (ABE), Civil and Architectural Engineering, Soil and Rock Mechanics. Kerberos Geoteknik.ORCID iD: 0000-0002-7361-0729
KTH, School of Architecture and the Built Environment (ABE), Civil and Architectural Engineering, Soil and Rock Mechanics.ORCID iD: 0000-0001-9615-4861
(English)Manuscript (preprint) (Other academic)
Abstract [en]

Current practices in compaction quality assessment rely heavily on traditional spot tests, such as the static plate load test (PLT), which are labor-intensive, time-consuming, and limited in spatial coverage. This study develops a machine-learning framework to estimate deformation moduli from continuous compaction measurements. A small dataset of 60 samples from full-scale trials was used to train multi-output regression models, with deformation moduli from the first and second loading cycles (Ev1 and Ev2) as the targets. Two feature representations were compared: a five-feature set of selected intelligent compaction measurement values (ICMVs) and an eight-feature set of curated vibration response characteristics extracted from acceleration signals and contact force–displacement curves. Six algorithms were evaluated, including multiple linear regression, random forest, XGBoost, support vector regression, k-nearest neighbors, and the tabular foundation model TabPFN. The results show that curated response characteristics improved Ev2 prediction relative to the selected ICMV feature set. For Feature Set 1, SVR achieved the best Ev2 prediction with R2 = 0.34 and RMSE = 7.1 MPa. For Feature Set 2, TabPFN achieved the best Ev2 prediction with R2 = 0.45 and RMSE = 6.5 MPa, despite requiring no hyperparameter tuning. The simultaneous prediction of Ev1 and Ev2 enabled estimation of the Ev2/Ev1 ratio, with the lowest RMSE = 0.34 achieved by SVR on Feature Set 2. SHAP analysis indicated that second-harmonic acceleration amplitude, loading and unloading stiffness, and displacement amplitude were key features, providing physical interpretability for the data-driven assessment.

Keywords [en]
Vibratory compaction, Intelligent compaction measurement value, Plate load test, Deformation modulus, TabPFN, SHAP analysis
National Category
Geotechnical Engineering and Engineering Geology
Identifiers
URN: urn:nbn:se:kth:diva-381412OAI: oai:DiVA.org:kth-381412DiVA, id: diva2:2060254
Projects
BIG – Branschsamverkan I Grunden
Note

QC 20260604

Revised manuscript submitted to the 6th International Conference on Geotechnics for Sustainable Infrastructure Development

Available from: 2026-05-15 Created: 2026-05-15 Last updated: 2026-06-04Bibliographically approved
In thesis
1. Toward reliable vibratory compaction control: Integrating full-scale testing and advanced finite element analysis
Open this publication in new window or tab >>Toward reliable vibratory compaction control: Integrating full-scale testing and advanced finite element analysis
2026 (English)Licentiate thesis, comprehensive summary (Other academic)
Abstract [en]

Vibratory compaction is widely used for improving the bearing capacity and stiffness of earthworks. The quality of a compacted layer is traditionally verified by spot tests, such as the plate load test (PLT), which provide limited spatial coverage. Continuous compaction control (CCC) addresses this limitation by deriving intelligent compaction measurement values (ICMVs) from the roller response. However, the use of CCC for quality assurance is still limited by uncertain correlations between ICMVs and PLT moduli, by incomplete physical interpretation of the roller response, and by the small number of measurements normally available for calibration.

This thesis investigates the reliability of CCC-based quality assurance for vibratory compaction by combining full-scale testing, finite element analysis, and machine learning. Five full-scale trials were conducted at the Dynapac compaction laboratory in Karlskrona on a 1 m thick layer of well-graded gravel. Two single-drum rollers, two PLT plate diameters, and different compaction states were included, and sixteen ICMVs were compared with the deformation modulus from PLTs. A two-dimensional plane-strain finite element model of the roller–soil system was implemented with four constitutive descriptions, ranging from linear elasticity to hypoplasticity with intergranular strain. Finally, six regression algorithms were evaluated on 60 samples in a multioutput framework for estimating 𝐸𝑣1 and 𝐸𝑣2 simultaneously.

The full-scale trials show that mechanics-based ICMVs, in particular the loading-phase vibration modulus, provide the strongest correlation with 𝐸𝑣2. The 600 mm PLT plate gives stronger and less scattered correlations than the 300 mm plate, and the lighter roller gives stronger correlations than the heavier roller. The numerical study shows that hypoplasticity with intergranular strain gives the most consistent description of cyclic densification, stiffness evolution, and loading–unloading response. It also shows that the influence depth depends on the selected response quantity, with density-related criteria giving larger depths than settlement-based criteria. In the simulations, the vibration modulus is more closely related to the final density state than CMV or OMEGA, while normalized void-ratio reduction helps describe the consumed part of the available compaction potential. The machine-learning study shows that curated response characteristics improve the prediction of 𝐸𝑣2 compared with predefined ICMVs. Among the tested algorithms, TabPFN gives the best overall estimates of 𝐸𝑣2, while the multi-output framework also enables the 𝐸𝑣2/𝐸𝑣1 ratio to be estimated. The results demonstrate both the potential and the remaining limitations of CCC-based quality assurance for vibratory compaction.

Abstract [sv]

Vibrationspackning används i stor omfattning för att förbättra bärförmågan och styvheten hos jordkonstruktioner. Kvaliteten hos ett packat lager kontrolleras traditionellt med punktvisa försök, till exempel statiska plattbelastningsförsök (PLT), som ger begränsad yttäckning. Yttäckande packningskontroll (eng. continuous compaction control, CCC) hanterar denna begränsning genom att härleda intelligenta packningsmätvärden (eng. intelligent compaction measurement values, ICMV) från responsen hos en instrumenterad vält. Den praktiska användningen av CCC för kvalitetssäkring begränsas dock fortfarande av osäkra samband mellan ICMV och deformationsmoduler från PLT, av en ofullständig fysikalisk tolkning av vältresponsen och av det begränsade antal mätningar som normalt finns tillgängliga för kalibrering.

I denna avhandling undersöks tillförlitligheten hos CCC-baserad kvalitetssäkring av vibrationspackning genom en kombination av fullskaleförsök, finita elementanalys och maskininlärning. Fem fullskaleförsök utfördes i Dynapacs packningslaboratorium i Karlskrona på ett 1 m tjockt lager av välgraderat grus. Två enkelvalsvältar, två plattdiametrar för PLT och olika packningstillstånd undersöktes, och sexton ICMV jämfördes med deformationsmodulen från PLT. En tvådimensionell finita elementmodell av systemet vält–jord under plant töjningstillstånd implementerades med fyra konstitutiva modeller, från linjär elasticitet till hypoplasticitet med intergranulär töjning. Slutligen utvärderades sex regressionsalgoritmer på 60 prover i en flermålsformulering för samtidig skattning av 𝐸𝑣1 och 𝐸𝑣2.

Fullskaleförsöken visar att mekanikbaserade ICMV, särskilt vibrationsmodulen under lastfasen, ger starkast korrelation med 𝐸𝑣2. PLT-plattan med 600 mm diameter ger starkare och mindre spridda korrelationer än 300 mm-plattan, och den lättare välten ger starkare korrelationer än den tyngre. Den numeriska studien visar att hypoplasticitet med intergranulär töjning ger den mest konsekventa beskrivningen av cyklisk förtätning, styvhetsutveckling och lastnings–avlastningsrespons. Den visar också att influensdjupet beror på vilken responsstorhet som används, där densitetsrelaterade kriterier ger större djup än kriterier baserade på sättning. I simuleringarna har vibrationsmodulen ett tydligare samband med det slutliga densitetstillståndet än CMV eller OMEGA, medan den normaliserade reduktionen av portalet bidrar till att beskriva den utnyttjade delen av den tillgängliga packningspotentialen. Maskininlärningsstudien visar att utvalda responsegenskaper ger bättre prediktion av 𝐸𝑣2 än fördefinierade ICMV. Bland de testade algoritmerna ger TabPFN de bästa övergripande skattningarna av 𝐸𝑣2, medan multi-output-ramverket också gör det möjligt att uppskatta kvoten 𝐸𝑣2/𝐸𝑣1. Resultaten visar både möjligheten och de kvarstående begränsningarna med CCC-baserad kvalitetssäkring av vibrationspackning.

Place, publisher, year, edition, pages
Stockholm, Sweden: KTH Royal Institute of Technology, 2026. p. 46
Series
TRITA-ABE-DLT ; 2620
Keywords
Vibratory compaction, intelligent compaction, continuous compaction control, plate load test, vibration modulus, hypoplasticity, machine learning, quality assurance, Vibrationspackning, intelligent packningskontroll, yttäckande packningskontroll, plattbelastningsförsök, vibrationsmodul, hypoplasticitet, maskininlärning, kvalitetssäkring
National Category
Geotechnical Engineering and Engineering Geology
Research subject
Civil and Architectural Engineering, Soil and Rock Mechanics
Identifiers
urn:nbn:se:kth:diva-382940 (URN)978-91-8106-647-0 (ISBN)
Presentation
2026-08-19, M24, Brinellvägen 64A, KTH Campus, Public video conference link: https://kth-se.zoom.us/j/62254984784, Stockholm, 10:00 (English)
Opponent
Supervisors
Note

QC 260812

Available from: 2026-08-12 Created: 2026-06-04 Last updated: 2026-08-18Bibliographically approved

Open Access in DiVA

No full text in DiVA

Authority records

Hua, WenjunWersäll, CarlLarsson, Stefan

Search in DiVA

By author/editor
Hua, WenjunWersäll, CarlLarsson, Stefan
By organisation
Soil and Rock Mechanics
Geotechnical Engineering and Engineering Geology

Search outside of DiVA

GoogleGoogle Scholar

urn-nbn

Altmetric score

urn-nbn
Total: 22 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf