Integrated Numerical and Digital Twin Framework for Scour Assessment of Offshore Wind Turbine with Tripod Suction Bucket Foundations

Kyeong Sun Kim

Seoul National University · Department of Civil and Environmental Engineering

2026-05-19

Table of Contents

1. Introduction · context · problem · research gap · objective
2. Methodology · experimental · numerical · software · decision
3. Results · sensitivity · validation · field · transferability
4. Discussion · engineering impact · before and after
5. Conclusion · findings · limitations · future work

The talk follows the standard research-paper structure. Introduction sets the context and states the research gap. Methodology presents the four evidence layers — experimental, numerical, software, and decision. Results quantify the findings. Discussion connects the results to engineering practice. Conclusion summarises the contributions, states the limitations, and outlines future work.

Introduction · Context

Offshore wind capacity exceeds 75 GW worldwide and is projected to quadruple by 2035. As turbines move into deeper water, foundation integrity under long-term environmental loading becomes a governing design consideration.

Scour — the progressive erosion of seabed sediment around the foundation — reduces the effective embedment depth and alters the soil–structure interaction that controls the dynamic response. For offshore wind turbines, even modest changes in foundation stiffness can shift the natural frequency toward the 1P resonance boundary and accelerate fatigue accumulation.

Offshore wind is moving into deeper water with larger turbines. Foundation integrity under long-term loading is the governing concern. Scour is the progressive erosion of seabed sediment around the foundation. It reduces embedment and alters soil-structure interaction. For offshore wind turbines, even modest changes in foundation stiffness can shift the natural frequency toward the 1P rotor-passing resonance boundary and accelerate fatigue accumulation.

Introduction · Problem

Monopile foundations dominate the existing offshore wind fleet, and scour detection on monopiles is mature. Tripod suction bucket foundations — already deployed at gigawatt scale in Korean and European waters — are structurally different and their scour response is poorly understood.

A tripod distributes the overturning load through three discrete footings connected by a push–pull couple. The geometry suppresses the natural-frequency change caused by scour by a factor of two to five compared with an equivalent monopile. Scour-induced capacity loss may therefore become structurally significant before any frequency-based alarm triggers.

Monopile foundations dominate the existing fleet and their scour detection is mature. Tripod suction bucket foundations are already deployed at gigawatt scale but their scour response is poorly understood. A tripod distributes overturning load through three footings connected by a push-pull couple. That geometry suppresses the frequency change caused by scour by two to five times compared to an equivalent monopile. Scour-induced capacity loss therefore becomes structurally significant before any frequency-based alarm triggers.

Introduction · Literature Review

Prendergast and co-workers (2013, 2015, 2020) established the laboratory evidence that natural frequency decreases monotonically with scour depth on single piles and bridge foundations. Kallehave et al. (2015) surveyed approximately 400 monopile-supported turbines and reported systematic discrepancies between design and operational frequencies.

Two parallel research tracks exist for suction bucket foundations. Houlsby and Byrne (2005, 2014, 2017) characterised the capacity of suction caissons under combined loading. Jalbi and Bhattacharya (2018, 2020) derived closed-form frequency expressions for multi-footing foundations. Neither track has produced a validated numerical model for scour-induced frequency degradation on tripod geometries.

Prendergast and co-workers established the laboratory evidence that natural frequency decreases monotonically with scour depth on single piles and bridge foundations. Kallehave et al surveyed about four hundred monopile-supported turbines and reported systematic discrepancies between design and operational frequencies. Two parallel research tracks exist for suction bucket foundations. Houlsby and Byrne characterised the capacity of suction caissons under combined loading. Jalbi and Bhattacharya derived closed-form frequency expressions for multi-footing foundations. Neither track has produced a validated numerical model for scour-induced frequency degradation on tripod geometries.

Introduction · Research Gap

Existing beam-on-Winkler-foundation models for offshore wind turbines are calibrated against monopile data. No distributed Winkler formulation exists for a suction bucket of any geometry, and operational field records of tripod scour response have not been published.

The combined gap produces a practical consequence. An asset owner cannot currently answer the question “should I inspect this foundation now, or can it wait another year?” with a defensible cost–benefit calculation. Design tools, sensor technology, and decision frameworks exist in isolation but do not connect.

Existing beam-on-Winkler models for offshore wind turbines are calibrated against monopile data. No distributed Winkler formulation exists for a suction bucket of any geometry. Operational field records of tripod scour response have not been published. The combined gap produces a practical consequence. An asset owner cannot currently answer the question should I inspect this foundation now or can it wait another year with a defensible cost-benefit calculation. Design tools, sensor technology, and decision frameworks exist in isolation but do not connect.

Introduction · Objective

This thesis develops a validated numerical framework, a field-verified detection pipeline, and a Bayesian decision layer that together enable cost-optimal scour-related maintenance decisions for tripod suction bucket foundations.

Four sub-objectives structure the work: (1) quantify the scour–frequency sensitivity across soil conditions, (2) bridge centrifuge measurements to the operational turbine, (3) integrate three-dimensional soil analysis with structural dynamics in one open-source pipeline, and (4) fuse multi-channel monitoring evidence into a single actionable recommendation.

This thesis develops a validated numerical framework, a field-verified detection pipeline, and a Bayesian decision layer that together enable cost-optimal scour-related maintenance decisions for tripod suction bucket foundations. Four sub-objectives structure the work. First, quantify the scour-frequency sensitivity across soil conditions. Second, bridge centrifuge measurements to the operational turbine. Third, integrate three-dimensional soil analysis with structural dynamics in one open-source pipeline. Fourth, fuse multi-channel monitoring evidence into a single actionable recommendation.

Introduction · Research Questions

Physics · numerical · experimental

J11   Does cavity expansion explain tripod spring shapes?

J2   Can calibrated Winkler replace underwater inspection?

J3   Does saturation change scour–frequency sensitivity?

J5   How does soil uncertainty amplify scour damage?

Op³   Can the assessment chain be automated?

Monitoring · decision · transfer

V   Can scour be separated from environmental variability?

V2   Why does cointegration fail on SHM data?

B   Which feature carries the scour signal?

A   How does evidence become a cost-optimal action?

E   Can the encoder transfer to unseen soils via CPT?

Ten research questions, grouped into two balanced themes of five each. Left column — physics, numerical, experimental: the theoretical backbone J11, the numerical Winkler J2, the centrifuge saturation question J3, the probabilistic capacity question J5, and the software-integration question Op3. Right column — monitoring, decision, transfer: the field-monitoring question V, the cointegration-failure question V2, the feature-selection question B, the decision-fusion question A, and the cross-soil-transfer question E. Each question maps to one paper; together they are the spine of the thesis.

Introduction · Research Framework

Three evidence tiers connect the ten first-author papers:

Tier 1 · Physics

J11 · theoretical backbone

J2 · 3D-FE-calibrated Winkler

J1, J3 · centrifuge programme

J5 · Monte Carlo capacity

Op³ · open-source integration

Tier 2 · Field & Features

V · 32-month Gunsan record

V2 · state-function theory

B · dimensionless features

Tier 3 · Decision

A · Bayesian fusion

E · cross-soil encoder

Each tier’s output is the next tier’s input. The framework is the spine the committee should remember.

The ten papers form three evidence tiers. Tier 1 — physics and numerical foundation — contains the theoretical backbone, the Winkler model, the centrifuge programme, the Monte Carlo capacity, and the software integration. Tier 2 — field and features — contains the 32-month field record, the state-function theoretical result, and the dimensionless feature selection. Tier 3 — decision — contains the Bayesian fusion and the cross-soil encoder. Each tier’s output is the next tier’s input. The framework is the spine the committee should remember as I walk through methodology, results, and discussion.

Methodology · Experimental Programme

Facility. Centrifuge tests were conducted at the KOCED geotechnical centrifuge at KAIST, at seventy-g acceleration, on a 1:70-scale model of the 4.2 MW Gunsan tripod bucket foundation.

Programme. Twenty-two test cases spanned five soil conditions — dense dry, loose dry, silty sand, dense saturated, and loose saturated. Four progressive scour stages up to S/D ≈ 0.6 were applied, with two of the series including a backfill stage.

Centrifuge tests were conducted at the KOCED geotechnical centrifuge at KAIST at seventy-g acceleration on a 1:70-scale model of the 4.2 MW Gunsan tripod bucket foundation. Twenty-two test cases spanned five soil conditions — dense dry, loose dry, silty sand, dense saturated, loose saturated. Four progressive scour stages up to scour depth ratio zero-point-six were applied, with two of the series including a backfill stage.

Methodology · Numerical Framework

Three-phase pipeline. Phase 1 is a three-dimensional finite-element limit analysis in OptumGX. Phase 2 extracts nonlinear spring parameters via Fourier decomposition of contact pressure fields. Phase 3 performs an eigenvalue analysis on the calibrated one-dimensional beam-on-nonlinear-Winkler-foundation model in OpenSeesPy.

Theoretical foundation. The spring shape is derived from Paper J11’s dissipation-weighted generalisation of Vesic’s cavity-expansion theory. The method uses a single calibration constant in place of the three separate empirical curves required by the conventional approach.

The numerical framework comprises three phases. Phase 1 is three-dimensional finite-element limit analysis in OptumGX. Phase 2 extracts nonlinear spring parameters via Fourier decomposition of the contact pressure field. Phase 3 performs eigenvalue analysis on the calibrated one-dimensional beam-on-nonlinear-Winkler-foundation model in OpenSeesPy. The spring shape is derived from Paper J11’s dissipation-weighted generalisation of Vesic’s 1972 cavity expansion theory. A single calibration constant replaces the three separate empirical curves required by the conventional approach.

Methodology · Software Framework

Op³. The Op³ framework couples three analysis tools — OptumGX, OpenSeesPy, and OpenFAST — through a single Python interface, enabling automated per-realisation calibration and population-scale Monte Carlo studies.

Verification. The framework ships with thirty-nine cross-validation benchmarks drawn from twenty-five published sources. Continuous integration runs on every commit. Distribution is through the Python Package Index and Zenodo, with a concept DOI that auto-resolves to the latest version.

The Op3 framework couples three analysis tools, OptumGX, OpenSeesPy, and OpenFAST, through a single Python interface, enabling automated per-realisation calibration and population-scale Monte Carlo studies. The framework ships with thirty-nine cross-validation benchmarks drawn from twenty-five published sources. Continuous integration runs on every commit. Distribution is through PyPI and Zenodo, with a concept DOI that auto-resolves to the latest version.

Methodology · Decision Framework

Bayesian fusion. Three evidence channels — a frequency channel from Paper J2’s Winkler model, a capacity channel from Paper J5’s Monte Carlo ensemble, and a statistical-detection channel from Paper V’s compensation residual — are combined through a hierarchical Bayesian posterior over scour depth.

Value of information. A classical engineering decision tree selects the cost-optimal action. Preposterior value-of-information analysis ranks the decision-worth of each sensor channel and supports sensor-budget allocation.

Three evidence channels, frequency from J2’s Winkler model, capacity from J5’s Monte Carlo, and statistical detection from V’s compensation residual, are combined through a hierarchical Bayesian posterior over scour depth. A classical engineering decision tree selects the cost-optimal action. Preposterior value-of-information analysis ranks the decision-worth of each sensor channel and supports sensor-budget allocation.

Results · Centrifuge Sensitivity

7× soil-dependent variation

Scour sensitivity α ranges from 0.015 in dense saturated sand to 0.103 in loose dry sand — a seven-fold spread.

1.94× saturation correction

Dry-sand centrifuge calibrations overestimate saturated-site sensitivity by approximately a factor of two.

Implication: a single-calibration detector misses alarms on stiff soils and cries wolf on soft ones. Multi-evidence fusion is not optional.

Scour sensitivity varies by a factor of seven across soil types, from 0.015 in dense saturated sand to 0.103 in loose dry sand. Dry-sand centrifuge calibrations overestimate saturated-site sensitivity by approximately two times. A single-calibration detector misses alarms on stiff soils and cries wolf on soft ones. Multi-evidence fusion is not optional.

Results · Numerical Validation

Baseline match. The 3D-FE-calibrated Winkler model predicts the first natural frequency of the 4.2 MW Gunsan tripod within minus 3.9 percent of 32 months of field measurements.

Power-law relationship. A power-law of the form \(|\Delta f / f_0| = 0.167 (S/D)^{1.47}\) holds across all eleven scour depths tested, enabling scour depth inference from routine accelerometer data. Computational cost is five orders of magnitude below direct 3D analysis.

The numerical model predicts the first natural frequency within minus 3.9 percent of 32 months of field measurements. A power-law relationship holds across all eleven scour depths tested, enabling scour depth inference from accelerometer data. Computational cost is five orders of magnitude below direct 3D analysis.

Results · Field Record

Monitoring campaign. A thirty-two-month continuous vibration record was collected on the 4.2 MW Gunsan tripod. Twenty-two thousand six hundred seventeen parked-state windows were processed through a three-stage compensation pipeline.

Performance. The pipeline produces zero false alarms over the full record. Detection probability reaches 95 percent at a synthetic frequency shift of 0.2 percent, which maps to a projected scour depth ratio of 0.39 through the J2 model.

A thirty-two-month continuous vibration record was collected on the 4.2 MW Gunsan tripod. Twenty-two thousand six hundred seventeen parked-state windows were processed through a three-stage compensation pipeline. The pipeline produces zero false alarms over the full record. Detection probability reaches 95 percent at a synthetic frequency shift of 0.2 percent, which maps to a projected scour depth ratio of 0.39 through the J2 model.

Results · Theoretical Result

Johansen cointegration is the most widely recommended statistical method for environmental compensation in structural health monitoring. Its integrated-of-order-one assumption fails on parked-state offshore wind turbine features because the features are stationary around a reversible environmental manifold.

Equivalence theorem. Under Gaussian-linear conditions, the state-function residual coincides with the Johansen cointegrating residual. Three explicit counterexamples — regime-shift boundaries, heavy-tailed innovations, and rank-deficient environmental covariates — show where state-function strictly outperforms.

Johansen cointegration is the most widely recommended statistical method for environmental compensation in structural health monitoring. Its integrated-of-order-one assumption fails on parked-state offshore wind turbine features because features are stationary around a reversible environmental manifold. Under Gaussian-linear conditions, the state-function residual coincides with the Johansen cointegrating residual. Three explicit counterexamples — regime-shift boundaries, heavy-tailed innovations, rank-deficient environmental covariates — show where state-function strictly outperforms. Monte Carlo verification confirms 0.019 percent agreement at sample size ten thousand.

Results · Feature and Transfer

Feature selection. Across sixty-four candidate vibration features evaluated via Buckingham-Pi dimensional analysis, the mid-elevation strain–acceleration coherence at the first natural frequency is the only sign-consistent cross-soil discriminator.

Cross-soil transfer. A physics-informed encoder with a cone-penetration-test-conditioned subject layer achieves correlation r ≈ 0.76 on seventeen unseen centrifuge sand series. Removing the conditioning drops cross-soil correlation by at least 0.10, establishing the conditioning mechanism as load-bearing.

Across sixty-four candidate vibration features evaluated via Buckingham-Pi dimensional analysis, the mid-elevation strain-acceleration coherence at the first natural frequency is the only sign-consistent cross-soil discriminator. A physics-informed encoder with a cone-penetration-test-conditioned subject layer achieves correlation 0.76 on seventeen unseen centrifuge sand series. Removing the conditioning drops cross-soil correlation by at least 0.10, establishing the conditioning mechanism as load-bearing.

Discussion · The Five-Fold Asymmetry

At a scour depth ratio of 0.5, the horizontal capacity drops by approximately sixteen percent while the natural frequency drops by only three percent. This five-fold asymmetry is the physical basis for the entire engineering recommendation.

Three consequences follow. Frequency-only monitoring is a blind spot on tripod foundations. Capacity-linked features such as bending strain detect scour earlier than acceleration-based features. Multi-channel fusion improves detection by approximately an order of magnitude over any single channel.

At a scour depth ratio of zero-point-five, horizontal capacity drops by approximately sixteen percent while natural frequency drops by only three percent. This five-fold asymmetry is the physical basis for the entire engineering recommendation. Frequency-only monitoring is a blind spot on tripod foundations. Capacity-linked features such as bending strain detect scour earlier than acceleration-based features. Multi-channel fusion improves detection by approximately an order of magnitude over any single channel.

Discussion · What Changes for Engineers

AS-IS

• Monopile-calibrated Winkler model
• Accelerometer only
• Fixed frequency-threshold alarm
• Inspection every two years, calendar-based
• Deterministic capacity estimate

TO-BE

• Tripod-validated, dissipation-weighted model
• Three channels fused via Bayesian posterior
• Condition-based alarm from posterior
• Monte Carlo capacity distribution
• Open-source, CPT-transferable to new sites

Outcome: misclassification below two percent · lifecycle cost reduction of forty to seventy percent.

Current practice applies a monopile-calibrated Winkler model with a single accelerometer and a fixed frequency-threshold alarm, inspects every two years regardless of condition, and estimates capacity deterministically. The proposed framework uses a tripod-validated dissipation-weighted model, fuses three sensor channels via a Bayesian posterior, sets condition-based alarms, estimates capacity as a Monte Carlo distribution, and is open-source and CPT-transferable to new sites. Misclassification drops below two percent and lifecycle cost reduces by forty to seventy percent.

Discussion · Bayesian Decision Performance

< 2 %

misclassification with three-channel fusion

40–70 %

lifecycle cost reduction vs. fixed-interval inspection

2×

capacity-sensor value-of-information vs. frequency sensor

The engineer’s question — “inspect now, or wait another year?” — is priced.

Three headline numbers from the Bayesian decision framework. Three-channel fusion drops misclassification from fifteen-to-twenty-five percent on a single channel to below two percent. Lifecycle cost reduces by forty to seventy percent compared with fixed-interval inspection. Value-of-information analysis ranks the capacity sensor at approximately twice the decision-worth of the frequency sensor, supporting concrete sensor-budget allocation. The engineer’s question, inspect now or wait another year, is priced.

Conclusion · Main Findings

Summary. A validated numerical framework, a field-verified detection pipeline, and a Bayesian decision layer together enable cost-optimal scour-related maintenance decisions for tripod suction bucket foundations.

• Seven-fold soil-dependent sensitivity variation
• Minus 3.9 percent baseline field match on a 32-month record
• Zero false alarms on a 32-month operational record
• Three-channel fusion reduces misclassification to below two percent

A validated numerical framework, a field-verified detection pipeline, and a Bayesian decision layer together enable cost-optimal scour-related maintenance decisions for tripod suction bucket foundations. Four main findings. Seven-fold soil-dependent sensitivity variation in the centrifuge programme. Minus three-point-nine percent baseline field match on the 32-month Gunsan record. Zero false alarms on the operational record. Three-channel Bayesian fusion reduces misclassification to below two percent.

Conclusion · Limitations

• Single site. Gunsan 4.2 MW tripod in normally consolidated marine clay. Generalisation relies on the CPT-conditioned transfer mechanism — demonstrated, not yet deployed at fleet scale.
• Axisymmetric scour only. Real scour is upstream-biased; the asymmetric-scour module is a follow-up.
• Parked-state detection only. Power-production regime compensation requires non-stationary forcing models and is reserved for future work.

Three principal limitations. Single site Gunsan tripod in normally consolidated marine clay; fleet deployment relies on the CPT-conditioned transfer mechanism which is demonstrated but not yet at fleet scale. Axisymmetric scour only; real scour is upstream-biased and the asymmetric-scour module is the follow-up. Parked-state detection only; power-production regime requires non-stationary forcing models.

Conclusion · Future Work

• Asymmetric scour and fleet deployment. Validation at a second site with per-bucket bathymetric survey; deployment through the CPT-conditioned encoder.
• Population-based structural health monitoring. Shared prior and per-turbine posterior across a wind farm; expected order-of-magnitude reduction in per-turbine calibration overhead.
• Decision theory under deeper uncertainty. Per-asset cost-matrix calibration and time-dependent extensions for lifecycle optimisation.

Three future directions. Asymmetric scour and fleet deployment through second-site validation and the CPT-conditioned encoder. Population-based structural health monitoring with shared prior and per-turbine posterior, expected to reduce per-turbine calibration overhead by an order of magnitude. Decision theory under deeper uncertainty through per-asset cost-matrix calibration and time-dependent extensions for lifecycle optimisation.

One Thing to Remember

For tripod suction bucket foundations under scour,
capacity degrades five times faster than frequency.

Frequency-only monitoring is a blind spot.
Three-channel Bayesian fusion closes it.

Ten papers · one spine · one result: monitor what degrades first, and fuse the evidence.

For tripod suction bucket foundations under scour, capacity degrades five times faster than frequency. Frequency-only monitoring is a blind spot. Three-channel Bayesian fusion closes it. Ten papers, one spine, one result — monitor what degrades first, and fuse the evidence. Pause deliberately. Breathe between lines.

감사합니다 · Thank You

Acknowledgements. KEPCO · MMB · Unison for the industrial partnership · KAIST 70 g centrifuge team · Advisor Prof. Sung-Ryul Kim · SNU CEE community.

Open science.

• pip install op3-framework
• https://doi.org/10.5281/zenodo.19476542
• https://github.com/ksk5429/numerical_model
• https://ksk5429.github.io/research-notes

Thank you. Acknowledgements to KEPCO, MMB, Unison; KAIST 70g centrifuge team; advisor Professor Sung-Ryul Kim; SNU Civil and Environmental Engineering community. Everything open. I am ready for questions.

Q&A · Methodology

Why eleven scour depths in J2? Twenty-four minutes per OptumGX scenario at 30,000 elements. Eleven depths resolve the power-law curvature within the field-measurement noise envelope.

Why is the equivalence proof Gaussian-linear only? Outside Gaussian-linear conditions the counterexamples demonstrate empirical superiority of state-function. Monte Carlo verifies 0.019 percent agreement within conditions at sample size ten thousand.

Why dissipation-weighted cavity expansion? Dissipation is a free output of any 3D limit analysis. The weighting mechanism is the contribution, not the underlying cavity-expansion kernel.

Q&A · Scope

How does the Gunsan result generalise? The CPT-conditioned encoder achieves r ≈ 0.76 on seventeen unseen centrifuge sand series. Op³ applies unchanged to monopile and jacket geometries; only the site-specific coefficients differ.

What if real asymmetric scour breaks the power-law? The axisymmetric power-law is replaced by per-bucket independent scour depths in the asymmetric-scour module. End-to-end validation awaits a second-site bathymetric survey.

Why 40–70 percent cost reduction rather than one number? The KEPCO cost matrix is order-of-magnitude. Action selection is stable for at least 80 percent of ±50 percent perturbations; specific cost reduction moves within the 40–70 percent band with the cost weights.

Q&A · Impact

How does this integrate with SCADA and CMMS systems? The Op³ operator interface consumes standard SCADA streams. The Bayesian-fusion output is compatible with Maximo and SAP PM.

What is the deployment cost? The three-channel sensor package breaks even within two years on a single 4 MW turbine. Fleet-scale payback is sub-year.

Industry-audience headline? Detection of 95 percent of damage events at a 0.2 percent frequency shift, zero false alarms over 32 months of operational data, open-source decision framework that cuts lifecycle inspection cost by 40 to 70 percent.

Questions?

• https://ksk5429.github.io/research-notes
• pip install op3-framework
• https://doi.org/10.5281/zenodo.19476542

Ready for questions.