Integrated numerical + digital-twin framework for offshore wind turbine foundations. Now covers fixed-bottom scour assessment (tripod suction buckets, monopiles, jackets) and floating-platform mooring anchors (suction caissons) in a single Python stack, with a production web GUI and an LLM chat front-end.
-brightgreen.svg){kind=icon}
PhD dissertation · Seoul National University (2026) · Kyeong Sun Kim · Civil & Environmental Engineering
Op³ bridges three otherwise-disconnected solvers — OptumGX (3D FE limit analysis, commercial), OpenSeesPy (structural dynamics, BSD-3-Clause), and OpenFAST v5 (aero-hydro-servo-elastic, Apache 2.0) — into a single V&V'd Python pipeline. A Bayesian decision layer, a digital-twin encoder, and a new web GUI turn it into a production tool for offshore-wind geotechnical engineers.
|
Suction-anchor design for floating OWT, branch
|
FastAPI + React + Three.js + Tailwind, branch
|
Op3 framework + V&V : 228 passed, 11 skipped, 0 failed
op3.anchors : 134 passed
op3_studio backend : 67 passed
────────────────────────────────────────────────────
Total : 429 passed, 11 skipped, 0 failed
Note: pre-existing CI failures (missing NREL reference decks, ...) have been resolved with
pytest.skipifguards in April 2026. CI now stays green on minimal checkouts that don't include the multi-GB reference data. Seedocs/AI_CODE_AUDIT_DEFENSE.mdfor outstanding dissertation-defense review items.
flowchart TD
FIELD["`**🌊 FIELD MEASUREMENT**
OMA accelerometers on nacelle
→ measured natural frequency f₁`"]
CPT["`**🌱 SITE CHARACTERIZATION**
CPT profile · cyclic loading history
→ layered soil state (Gₛ, φ, sᵤ)`"]
subgraph MODEL [" 🏗️ Op³ FORWARD MODEL "]
direction TB
FE["`**OptumGX**
3D FE limit analysis
capacity + dissipation`"]
BNWF["`**OpenSeesPy**
1D BNWF structural dynamics
Mode A / B / C / D`"]
AHSE["`**OpenFAST v5**
aero-hydro-servo-elastic
coupled simulation`"]
FE -- "CSV" --> BNWF -- "6×6 K_SSI" --> AHSE
end
DB["`**📦 MC DATABASE**
1,794 Monte Carlo samples
scour × soil × capacity × f₁`"]
BAYES["`**🎯 BAYESIAN UPDATE**
posterior on scour depth
credible intervals · VoI`"]
ACT["`**⚠️ PRESCRIPTIVE ACTION**
monitor · inspect · backfill
defensible decision under uncertainty`"]
CPT --> MODEL
MODEL --> DB
FIELD --> BAYES
DB --> BAYES
BAYES --> ACT
style FIELD fill:#1a1a2e,stroke:#58a6ff,color:#58a6ff
style CPT fill:#1a1a2e,stroke:#58a6ff,color:#58a6ff
style FE fill:#2e1a1a,stroke:#ff7b72,color:#ff7b72
style BNWF fill:#0a2910,stroke:#3fb950,color:#3fb950
style AHSE fill:#0a2910,stroke:#3fb950,color:#3fb950
style DB fill:#1a1a2e,stroke:#58a6ff,color:#58a6ff
style BAYES fill:#1a0a29,stroke:#bc8cff,color:#bc8cff
style ACT fill:#2e1a1a,stroke:#e3b341,color:#e3b341
git clone https://github.com/ksk5429/numerical_model.git
cd numerical_model/op3_studio
cp .env.example .env # add ANTHROPIC_API_KEY for AI chat
docker compose up # backend :8000 + frontend :5173Open http://localhost:5173, drag the scour slider, click "Generate report", type a question at the AI chat. The OpenAPI explorer is at http://localhost:8000/docs.
pip install op3-frameworkFixed-bottom foundation (tripod suction bucket, monopile, jacket):
# End-to-end OpenFAST v5 coupled simulation
python scripts/run_openfast.py site_a --tmax 5
python scripts/run_dlc11_partial.py --tmax 600 --speeds 8 12 18
# Standards conformance + full V&V suite
python scripts/dnv_st_0126_conformance.py --all
python scripts/release_validation_report.py # 18/19 PASS in ~42 sFloating-platform suction anchor (new):
from op3.anchors import (
SuctionAnchor, UndrainedClayProfile, MooringLoad,
anchor_capacity, installation_analysis,
)
anchor = SuctionAnchor(diameter_m=5.0, skirt_length_m=15.0,
padeye_depth_m=10.0, submerged_weight_kN=250.0)
soil = UndrainedClayProfile(su_mudline_kPa=5.0,
su_gradient_kPa_per_m=1.5)
load = MooringLoad(tension_kN=4000.0, angle_at_padeye_deg=25.0)
r = anchor_capacity(anchor, soil, method="dnv_rp_e303", load=load)
print(f"T_ult = {r.T_ult_kN:.0f} kN, FoS = {r.factor_of_safety(4000.0):.2f}")
inst = installation_analysis(anchor, soil, water_depth_m=200.0)
print(f"feasible: {inst.feasible}, "
f"max suction: {inst.max_suction_required_kPa:.1f} kPa")Five runnable examples in examples/anchor_0X_*.py.
🐍 Python API — build a PISA foundation and compose a tower model
from op3 import compose_tower_model
from op3.foundations import foundation_from_pisa
from op3.standards.pisa import SoilState
# 1. Build a PISA-derived foundation
profile = [
SoilState(0.0, 5.0e7, 35, "sand"),
SoilState(15.0, 1.0e8, 35, "sand"),
SoilState(36.0, 1.5e8, 36, "sand"),
]
foundation = foundation_from_pisa(diameter_m=6.0, embed_length_m=36.0, soil_profile=profile)
# 2. Compose a tower model
model = compose_tower_model(
rotor="nrel_5mw_baseline",
tower="nrel_5mw_oc3_tower",
foundation=foundation,
)
# 3. Run analyses
freqs = model.eigen(n_modes=3)
print(f"f1 = {freqs[0]:.4f} Hz") # fixed: 0.3158, PISA: 0.3157
K_6x6 = model.extract_6x6_stiffness() # condense to head stiffness
pushover = model.pushover(target_disp_m=0.5) # static pushover
transient = model.transient(duration_s=10.0) # free vibrationFour foundation modes (Fixed / 6×6 / BNWF / Dissipation-weighted) · 6 standards (DNV · ISO · API · OWA · PISA · HSsmall) · Bayesian + PCE + Monte Carlo UQ · direct OpenFAST SoilDyn export · Apache-2.0.
Feature comparison vs SACS / PLAXIS / OpenSeesPy / OpenFAST
| Capability | Op³ | SACS | PLAXIS | OpenSeesPy | OpenFAST |
|---|---|---|---|---|---|
| Four foundation modes (Fixed / 6x6 / BNWF / Dissipation-weighted) | ✅ | partial | partial | manual | ❌ |
| PISA (Burd 2020 / Byrne 2020) with depth functions | ✅ | ❌ | commercial | ❌ | ❌ |
| Cyclic Hardin-Drnevich layered on PISA | ✅ | ❌ | ❌ | ❌ | ❌ |
| DNV / ISO / API / OWA / PISA / HSsmall standards | 6 | proprietary | 1-2 | ❌ | ❌ |
| Mode D dissipation-weighted BNWF (novel) | ✅ | ❌ | ❌ | ❌ | ❌ |
| Direct Op³ → SoilDyn export | ✅ | ❌ | ❌ | ❌ | native |
| Monte Carlo soil propagation | ✅ | ❌ | manual | manual | ❌ |
| Hermite polynomial chaos expansion | ✅ | ❌ | ❌ | manual | ❌ |
| Grid Bayesian calibration | ✅ | ❌ | ❌ | manual | ❌ |
| V&V test suite | 140 | proprietary | proprietary | user-built | ~200 r-test |
| License | Apache-2.0 | commercial | commercial | BSD-3 | Apache-2.0 |
| Python-native | ✅ | ❌ | ❌ | wrapper | wrapper |
92% V&V · 140 tests · OpenFAST v5 end-to-end · OC6 Phase II match to 0.5% · PISA field validation at 10–30× error reduction · Sphinx docs + tutorials.
Detailed release notes
- 35 / 38 cross-validation benchmarks verified (92%) against 20+ published sources (Fu & Bienen 2017, Vulpe 2015, Doherty 2005, Houlsby 2005, Jalbi 2018, Gazetas 2018, and 14 more)
- 140 unit tests pass across 15 modules (code verification, consistency, sensitivity, extended invariants, PISA, cyclic degradation, HSsmall, Mode D, UQ, reproducibility snapshot, framework integration, ...)
- 4 / 4 calibration regression against published references with all four examples within 4% of the most stringent (NREL 5 MW OC3 at -0.4% vs Jonkman & Musial 2010)
- OpenFAST v5.0.0 end-to-end on SiteA tripod + SoilDyn with Op³ PISA-derived 6×6 stiffness, 8-module coupled simulation
- DLC 1.1 partial sweep at U = {8, 12, 18} m/s — 3 / 3 PASS; full 12-speed × 600 s run scaled overnight
- 35 / 36 DNV-ST-0126 conformance (single failure is the real SiteA 1P resonance finding, not a bug)
- OC6 Phase II benchmark (Bergua 2021 NREL/TP-5000-79989): Op³ K_zz matches to 1.3%, f1_clamped to 0.5%
- PISA field-test cross-validation (McAdam 2020 + Byrne 2020) with depth-function + eccentric-load-compliance corrections reducing prior-release errors by 10–30× on short rigid piles
- End-to-end Bayesian calibration of NREL 5 MW OC3 tower EI: posterior mean 1.014 ± 0.076, 5%-95% credible interval [0.888, 1.145]
- Sphinx documentation (~5000 lines across 9 RST pages + 6 tutorial notebooks), ReadTheDocs-ready and GitHub Pages-deployable
Op³ is cross-validated against 39 independent benchmarks from 20+ published sources. 35 of 38 in-scope benchmarks verified (92%).
Headline matches: NcV = 6.006 (+1.1%) · NcM = 1.468 (−0.8%) · KR/(R³G) = 17.28 (+3.1%) · field Kr = 177 MN·m/rad (−21%, first field validation).
Full report: validation/cross_validations/VV_REPORT.md
📊 Full benchmark table (16 categories)
| Category | Benchmarks | Error range | Sources |
|---|---|---|---|
| Eigenvalue (f1) | #1--5 | 1.2--13% | Jonkman 2010, Gaertner 2020, Kim 2025 |
| Bearing capacity (OptumGX FELA) | #14--15 | 0.8--7.8% | Fu & Bienen 2017, Vulpe 2015 |
| Foundation stiffness | #16--17, #20 | 0.1--26% | Jalbi 2018, Gazetas 2018, Doherty 2005 |
| Field trial | #19 | -21% | Houlsby 2005 (Bothkennar) |
| Scour sensitivity | #10--11 | within published ranges | Zaaijer 2006, Prendergast 2015 |
| Design compliance | #13 | 0% | DNV-ST-0126 (2021) |
| PISA clay stiffness | #6 | 16--32% | Burd et al. 2020 |
| VH envelope | #8 | -7.7% | Houlsby & Byrne / Vulpe 2015 |
| p_ult(z) profile | #21 | consistent | OptumGX plate extraction |
| Centrifuge yield moment | #22 | -0.7% | DJ Kim et al. 2014 |
| Full-scale tripod f1 | #24 | -0.2% | Seo et al. 2020 |
| Walney 1 monopile f1 | #25 | -2.1% | Arany et al. 2015 |
| Suction bucket scour sensitivity | #26 | within range | Cheng et al. 2024 |
| f_meas/f_design ratio | #27 | +0.3% | Kallehave et al. 2015 |
| Cyclic rotation (N=100, N=1M) | #28 | 3.7--4.3% | Jeong et al. 2021 |
| OC4 jacket f1 (fixed-base) | #29 | +1.9% | Popko et al. 2012 |
flowchart TB
T["`**39 Benchmarks · 20+ sources**
35 / 38 in-scope verified (92%)`"]
T --> C1["`✅ <b>Eigenvalue f₁</b> · 1.2–13%<br/>Jonkman 2010 · Gaertner 2020`"]
C1 --> C2["`✅ <b>Bearing capacity</b> · NcV +1.1% · NcM −0.8%<br/>Fu & Bienen 2017 · Vulpe 2015`"]
C2 --> C3["`✅ <b>PISA field trial</b> · 10–30× error reduction<br/>McAdam 2020 · Byrne 2020`"]
C3 --> C4["`✅ <b>OC6 Phase II</b> · K_zz 1.3% · f₁ 0.5%<br/>Bergua 2021 NREL/TP-5000-79989`"]
C4 --> C5["`✅ <b>Centrifuge yield M</b> · −0.7%<br/>DJ Kim 2014`"]
C5 --> C6["`✅ <b>Cyclic N=10⁶</b> · 3.7–4.3%<br/>Jeong 2021`"]
C6 --> C7["`✅ <b>Full-scale tripod f₁</b> · −0.2%<br/>Seo 2020`"]
C7 --> C8["`✅ <b>DNV-ST-0126</b> · 35/36 conformance`"]
C8 --> C9["`🟡 <b>Field Kr</b> · −21%<br/>Houlsby 2005 · first field validation`"]
style T fill:#1a1a2e,stroke:#58a6ff,color:#58a6ff
style C1 fill:#0a2910,stroke:#3fb950,color:#3fb950
style C2 fill:#0a2910,stroke:#3fb950,color:#3fb950
style C3 fill:#0a2910,stroke:#3fb950,color:#3fb950
style C4 fill:#0a2910,stroke:#3fb950,color:#3fb950
style C5 fill:#0a2910,stroke:#3fb950,color:#3fb950
style C6 fill:#0a2910,stroke:#3fb950,color:#3fb950
style C7 fill:#0a2910,stroke:#3fb950,color:#3fb950
style C8 fill:#0a2910,stroke:#3fb950,color:#3fb950
style C9 fill:#2e2410,stroke:#e3b341,color:#e3b341
Reproduce all results:
python validation/cross_validations/run_all_cross_validations.pySee CHANGELOG.md for the full release history and docs/DEVELOPER_NOTES.md for the implementation journal of Track C phases 1 through 8.
| Link | |
|---|---|
| 🌐 Hosted docs | op3-framework.readthedocs.io · Pages mirror |
| 📘 User manual | docs/sphinx/user_manual.rst |
| 🔬 Technical reference | docs/sphinx/technical_reference.rst |
| 📓 Tutorials | docs/tutorials/ — 6 Jupyter notebooks |
| 🧪 Mode D paper-draft | docs/MODE_D_DISSIPATION_WEIGHTED.md |
Auto-deployed on every push via
.readthedocs.yaml and
.github/workflows/docs-deploy.yml.
All documentation pages
| Page | Content |
|---|---|
| Environment setup | Clone, install, OpenFAST bootstrap, r-test bootstrap |
| User manual | Worked examples: every foundation mode, every standard, UQ tools, OpenFAST coupling |
| Technical reference | Units, coordinates, DOFs, PISA math, Hermite PCE, Hardin-Drnevich, Rayleigh cantilever |
| Scientific report | Narrative, distinctive contributions, OC6 + PISA validation findings, limitations |
| Troubleshooting / FAQ | ~30 common issues across install / OpenSees / OpenFAST / PISA / UQ / V&V |
| Contributing guide | V&V-or-it-didn't-happen rule, commit discipline, release process |
| Developer notes | Full implementation journal, all 8 Track C phases with lessons learned |
| Mode D formulation | Novel dissipation-weighted BNWF paper-draft |
| Tutorials | 6 Jupyter notebooks: quickstart, foundation modes, UQ, calibration, SoilDyn, DLC sweeps |
pip install op3-framework
python -c "import op3; print(op3.__version__)" # 1.0.0-rc2Or clone + develop-install + run the full V&V suite in ~42 s:
git clone https://github.com/ksk5429/numerical_model.git
cd numerical_model && pip install -e ".[test,docs]"
PYTHONUTF8=1 python scripts/release_validation_report.py
# → 18/19 PASS, 0 mandatory FAIL🔧 Bootstrap OpenFAST v5 + r-test (for coupled simulations)
# OpenFAST v5.0.0 binary
mkdir -p tools/openfast
curl -L -o tools/openfast/OpenFAST.exe \
https://github.com/OpenFAST/openfast/releases/download/v5.0.0/OpenFAST.exe
# r-test suite
mkdir -p tools/r-test_v5 && cd tools/r-test_v5
git clone --depth=1 --branch v5.0.0 https://github.com/OpenFAST/r-test.git
cd ../..See docs/sphinx/environment.rst for
platform notes.
📁 Full tree
numerical_model/
├── README.md this file
├── LICENSE Apache-2.0
├── CITATION.cff citation metadata
├── pyproject.toml PEP 517 package
│
├── op3/ main Op³ framework
│ ├── composer.py TowerModel: eigen / pushover / transient
│ ├── foundations.py Foundation dataclass + factory
│ ├── opensees_foundations/ OpenSeesPy builder + ElastoDyn loader
│ ├── openfast_coupling/ Op³ → SubDyn / SoilDyn bridge
│ ├── standards/ DNV · ISO · API · OWA · PISA · HSsmall
│ ├── uq/ Propagation · PCE · Bayesian
│ ├── sacs_interface/ SACS jacket deck parser
│ ├── optumgx_interface/ OptumGX scripts (licence holders)
│ └── config/site_a.yaml single source of truth
│
├── data/ OptumGX persisted outputs
│ ├── integrated_database_1794.csv master MC database
│ └── fem_results/ small result CSVs
│
├── site_a_ref4mw/ SiteA 4 MW class decks (v4 / v5)
├── nrel_reference/ NREL + IEA reference turbines
├── validation/ V&V tests + benchmark artefacts
├── examples/ 11 turbine build.py files
├── tests/ 140 pytest assertions
├── scripts/ Runners · audits · release tooling
├── docs/ Sphinx + tutorials + Mode D notes
├── paper/ JOSS-format paper + BibTeX
└── .github/workflows/ CI · docs deploy · release validation
Historical introduction (preserved from v0.1)
The original Op³ framework was developed around the Gunsan offshore wind turbine site (KEPCO demonstration, UNISON U136) between 2023 and 2026 as part of the author's PhD dissertation at Seoul National University.
| Solver | License | Runnable by anyone? |
|---|---|---|
| OpenSeesPy | BSD-3-Clause | ✅ pip install openseespy |
| OpenFAST | Apache-2.0 | ✅ download v5.0.0 binary |
| OptumGX | Commercial — academic licence | ❌ licence holders only |
OptumGX runs once, upstream, to produce the capacity / dissipation
CSVs committed under data/fem_results/ and
data/integrated_database_1794.csv.
The open-source path (OpenSeesPy + OpenFAST) reproduces every
headline result without OptumGX installed.
How the commercial-solver constraint is handled
A third party without OptumGX can still:
- ✅ Read the persisted OptumGX output CSVs directly
- ✅ Run the complete OpenSeesPy foundation analysis pipeline
- ✅ Run the complete OpenFAST coupling pipeline
- ✅ Reproduce every headline numerical result in the dissertation
- ❌ Not re-run the OptumGX simulations from scratch
- ❌ Not change OptumGX input parameters (only the pre-computed parameter envelope is available)
The OptumGX interface scripts in op3/optumgx_interface/
are provided as reference for licence holders who wish to extend the
parameter envelope. They import from the commercial optumgx Python
API; attempts to run them without the licence fail at import time
with a clear error message.
flowchart TD
subgraph UPSTREAM ["🔒 UPSTREAM — runs once, license-gated"]
OGX["`**OptumGX**
3D FE limit analysis · commercial
· bearing capacity envelopes (V, H, M)
· depth-resolved contact pressures
· plastic dissipation fields
· p_ult(z) profiles`"]
end
subgraph OPEN ["🐍 OPEN-SOURCE PATH — runs for anyone"]
direction TB
OSP["`**OpenSeesPy** · BSD-3
1D BNWF structural dynamics
· eigenmodes · pushover
· transient response
· 6×6 condensation at mudline`"]
OF["`**OpenFAST v5** · Apache-2.0
aero-hydro-servo-elastic
· rotor + tower + SubDyn
· wind + wave time-domain
· 8-module coupled simulation`"]
OSP -- "6×6 K_SSI" --> OF
end
DB["`**📦 integrated_database_1794.csv**
1,794 Monte Carlo samples
capacity × stiffness × scour × f₁
reproducible without OptumGX`"]
OGX == "CSV committed to repo" ==> OSP
OGX -.->|"persisted"| DB
OSP -.->|"persisted"| DB
OF -.->|"persisted"| DB
style OGX fill:#2e1a1a,stroke:#ff7b72,color:#ff7b72
style OSP fill:#0a2910,stroke:#3fb950,color:#3fb950
style OF fill:#0a2910,stroke:#3fb950,color:#3fb950
style DB fill:#1a1a2e,stroke:#58a6ff,color:#58a6ff
Two boundary crossings need to be made explicit because they are where most of the engineering work of Op³ actually lives:
-
OptumGX → OpenSeesPy. How does a 3D finite-element capacity analysis translate into 1D beam-on-nonlinear-Winkler-foundation spring parameters? The framework offers four foundation modules described in the next section; each one represents a different level of fidelity and computational cost.
-
OpenSeesPy → OpenFAST. How does the structural dynamic response of the foundation enter the rotor-nacelle-tower-substructure coupled simulation? The framework extracts a 6×6 stiffness matrix (or a frequency-dependent impedance function) from the OpenSeesPy model and injects it into the OpenFAST SubDyn module as a substructure interface condition.
Op³ exposes four ways to represent the foundation in OpenSeesPy, arranged in increasing order of fidelity and computational cost. The choice is made at runtime via a single configuration flag, so the rest of the OpenSeesPy tower model remains identical across all four modes — only the foundation boundary condition changes.
flowchart TD
A["`**Mode A** · 🧱 Fixed base
⚡ fastest · upper-bound reference`"]
B["`**Mode B** · 🔗 6×6 lumped K
⚡ fast · SubDyn-native · PISA paradigm`"]
C["`**Mode C** · 🌾 Distributed BNWF
🟡 medium · depth-resolved p-y / t-z
scour progression per depth`"]
D["`**Mode D** ★ novel · 🎯 Dissipation-weighted
🐢 slow · energy-consistent · research-grade`"]
A ==> B ==> C ==> D
style A fill:#1a1a2e,stroke:#8b949e,color:#8b949e
style B fill:#0a2910,stroke:#58a6ff,color:#58a6ff
style C fill:#0a2910,stroke:#3fb950,color:#3fb950
style D fill:#1a0a29,stroke:#bc8cff,color:#bc8cff
⬇ increasing fidelity + computational cost
| Mode | Name | Fidelity | Runtime | Use case |
|---|---|---|---|---|
| A | Fixed base | Lowest | Fastest | Upper-bound reference, sanity check, rapid design iteration |
| B | 6×6 lumped stiffness | Low | Fast | Frequency-domain SSI for OpenFAST SubDyn; matches the PISA paradigm |
| C | Distributed BNWF springs | Medium | Medium | Depth-resolved stiffness; captures scour progression at each depth |
| D | Dissipation-weighted generalized BNWF | High | Slow | Full energy-consistent coupling with OptumGX plastic dissipation field |
All four modes share the same tower, rotor, and nacelle inertia
properties from op3/config/site_a.yaml.
Only the foundation representation changes, which makes it trivial to
compare the effect of foundation modeling choice on the predicted
natural frequency, mode shape, or transient response.
📘 Per-mode code examples
from op3.opensees_foundations import build_tower_model
model = build_tower_model(foundation_mode='fixed')
model.eigen(1)Base fixed at mudline. No soil contribution, no scour sensitivity. Upper-bound f₁ against which all other modes are compared.
model = build_tower_model(
foundation_mode='stiffness_6x6',
stiffness_matrix='data/fem_results/K_6x6_baseline.csv',
)Six-DOF linear spring at the base node (translation + rotation + coupling). This is the representation OpenFAST SubDyn accepts directly and the one the PISA programme uses for rigid bucket-like foundations [@burd2020pisasand; @byrne2020pisaclay]. Scour progression is modelled by swapping in a different 6×6 matrix per scour level.
model = build_tower_model(
foundation_mode='distributed_bnwf',
spring_profile='data/fem_results/opensees_spring_stiffness.csv',
scour_depth=1.5,
)Nonlinear p-y and t-z springs distributed along the skirt depth. Stiffnesses and capacities are calibrated from the OptumGX contact-pressure database, then scaled by a depth-dependent stress-correction factor as the scoured mudline moves downward. Chapter 6 core representation.
model = build_tower_model(
foundation_mode='dissipation_weighted',
ogx_dissipation='data/fem_results/dissipation_profile.csv',
ogx_capacity='data/fem_results/power_law_parameters.csv',
scour_depth=1.5,
)Extends Mode C with a depth-dependent participation factor derived
from the OptumGX plastic dissipation field at collapse — the
generalised cavity-expansion framework of Appendix A (uniform
plastic-zone assumption replaced by a spatially varying weight
function). Stiffness, ultimate resistance, and half-displacement
all derive from a single energy-consistent expression with sᵤ
cancelling in the y₅₀ parameter. Recommended for research use.
from op3.opensees_foundations import compare_foundation_modes
results = compare_foundation_modes(
modes=['fixed', 'stiffness_6x6', 'distributed_bnwf', 'dissipation_weighted'],
scour_levels=[0.0, 0.5, 1.0, 1.5, 2.0],
)See examples/compare_foundation_modes.py
for the full reproducer of dissertation Table 6.X.
Op³ now covers floating-platform anchors as well as fixed-bottom
foundations. The op3.anchors package implements suction anchor
capacity, installation, padeye optimisation, cyclic degradation, and
MoorPy coupling for floating offshore wind mooring design.
| Feature | Fixed-bottom (existing) | Floating-platform (new) |
|---|---|---|
| Geometry | Tripod suction bucket (L/D ~ 0.5-1) | Single suction anchor (L/D ~ 1-6) |
| Loading | V/H/M from tower | Inclined tension at padeye from mooring |
| Standard | DNV-ST-0126, OWA bearing | DNV-RP-E303, API RP 2SK |
| Capacity methods | OWA, PISA, FE | DNV-RP-E303, Murff-Hamilton, API RP 2SK, Aubeny 2003, FE-calibrated |
| Data flow | Foundation → tower → OpenFAST | OpenFAST/MoorPy → anchor (downstream) |
from op3.anchors import (
SuctionAnchor, UndrainedClayProfile, MooringLoad,
anchor_capacity, installation_analysis,
optimal_padeye_analytical, optimal_padeye_from_dissipation,
cyclic_capacity_reduction, anchor_safety_factor_timeseries,
)
anchor = SuctionAnchor(diameter_m=5.0, skirt_length_m=15.0,
padeye_depth_m=10.0, submerged_weight_kN=500.0)
soil = UndrainedClayProfile(su_mudline_kPa=5.0,
su_gradient_kPa_per_m=1.5)
# 1. capacity (5 methods)
r = anchor_capacity(anchor, soil, method='dnv_rp_e303',
load_angle_deg=30.0)
# 2. installation feasibility
inst = installation_analysis(anchor, soil, water_depth_m=200.0)
# 3. optimal padeye
z_p = optimal_padeye_analytical(anchor, soil,
method='supachawarote_2005')The module follows the same "no synthetic data" rule as the rest of
Op³: finite-element features require a real OptumGX output CSV
produced by op3/anchors/optumgx_anchor_run.py
(run from inside the OptumGX GUI), and mooring coupling uses the
real MoorPy package. See
docs/ANCHOR_OPTUMGX_GUIDE.md for
the LLM-assisted workflow where the user operates OptumGX and an
LLM drives everything downstream.
optimal_padeye_from_dissipation() is a new method that derives the
optimal padeye depth from the Op³ Mode D plastic-dissipation field:
z* = ∫₀ᴸ z ψ(z) dz / ∫₀ᴸ ψ(z) dz
where ψ(z) is the depth-distributed plastic dissipation weight
produced by the OptumGX driver. This extends the existing Mode D
formulation to anchor design and is the single novel theoretical
contribution of the anchor module.
| Script | What it demonstrates |
|---|---|
examples/anchor_01_basic_capacity.py |
Four analytical capacity methods + V-H envelope comparison |
examples/anchor_02_installation_check.py |
Self-weight + suction feasibility + plug-heave |
examples/anchor_03_padeye_optimization.py |
Analytical vs sensitivity sweep vs dissipation-centroid |
examples/anchor_04_cyclic_storm.py |
Andersen 2015 cyclic reduction applied to DNV envelope |
examples/anchor_05_moorpy_coupling.py |
End-to-end: MoorPy catenary → Op³ DNV FoS check |
The coupling is two-way in principle but one-way in practice: the foundation stiffness from OpenSeesPy is extracted as a 6×6 linear matrix at the mudline (or as a frequency-dependent impedance function) and written to a SubDyn input file. OpenFAST then runs the full aero-hydro-servo-elastic simulation with the foundation fully characterized by that matrix. This is computationally efficient because OpenFAST does not need to re-compute the foundation response at each time step.
flowchart LR
subgraph SEES ["🐍 OpenSeesPy (BSD-3)"]
direction TB
BNWF["`**BNWF model**
Mode C or D
distributed p-y / t-z
per scour level`"]
EIG["`**Eigenvalue +
static condensation**
at mudline node`"]
K["`**6×6 K_SSI**
head stiffness matrix
→ SubDyn input`"]
BNWF --> EIG --> K
end
subgraph FAST ["⚙️ OpenFAST v5 (Apache-2.0)"]
direction TB
SUB["`**SubDyn**
substructure interface
accepts 6×6 K directly`"]
SIM["`**AHSE coupled sim**
rotor + tower + waves + wind
time-domain response
8-module integration`"]
SUB --> SIM
end
K ==>|"one-way coupling<br/>per scour level"| SUB
style BNWF fill:#0a2910,stroke:#3fb950,color:#3fb950
style EIG fill:#0a2910,stroke:#3fb950,color:#3fb950
style K fill:#1a1a2e,stroke:#58a6ff,color:#58a6ff
style SUB fill:#0a2910,stroke:#3fb950,color:#3fb950
style SIM fill:#0a2910,stroke:#3fb950,color:#3fb950
The extraction is handled by
op3/openfast_coupling/opensees_stiffness_extractor.py.
The SubDyn input file is generated by
op3/openfast_coupling/build_site_a_subdyn.py.
Both scripts are driven by the single-source-of-truth YAML at
op3/config/site_a.yaml.
Op³'s prescriptive layer closes the loop between the forward model and field observation. Given a measured natural frequency from nacelle OMA, the framework performs a grid-Bayesian update on scour depth, then computes the Value of Information to choose between monitor, inspect, and backfill actions.
flowchart TD
subgraph PRIOR ["📚 PRIOR KNOWLEDGE"]
direction TB
CPT["`**CPT profile**
layered Gₛ, φ, sᵤ`"]
HIST["`**Scour history**
prior p(S) from
hydrodynamic model`"]
end
subgraph FWD ["🔄 FORWARD MODEL (Op³ MC)"]
direction TB
MC["`**1,794 Monte Carlo runs**
soil × scour × capacity`"]
SURR["`**Surrogate f₁(S, θ)**
PCE + encoder projection`"]
MC --> SURR
end
OBS["`**📡 OBSERVATION**
measured f₁ from OMA
with σ_meas`"]
LIK["`**🎯 LIKELIHOOD**
p(f₁_obs | S, θ)
Gaussian residual`"]
POST["`**📈 POSTERIOR**
p(S | f₁_obs)
grid Bayesian update
mean + 5–95% CI`"]
VOI["`**💰 VALUE OF INFORMATION**
expected loss vs cost:
· no action
· additional inspection
· backfill now`"]
DEC{"`**⚠️ DECISION**
VoI > threshold?`"}
A1["`✅ <b>Monitor</b><br/>continue OMA<br/>next cycle`"]
A2["`🔍 <b>Inspect</b><br/>ROV / bathymetry<br/>reduce prior variance`"]
A3["`🔧 <b>Backfill</b><br/>restore embedment<br/>reset scour state`"]
PRIOR --> LIK
FWD --> LIK
OBS --> LIK
LIK --> POST
POST --> VOI
VOI --> DEC
DEC -->|"low"| A1
DEC -->|"medium"| A2
DEC -->|"high"| A3
style CPT fill:#1a1a2e,stroke:#58a6ff,color:#58a6ff
style HIST fill:#1a1a2e,stroke:#58a6ff,color:#58a6ff
style MC fill:#0a2910,stroke:#3fb950,color:#3fb950
style SURR fill:#0a2910,stroke:#3fb950,color:#3fb950
style OBS fill:#1a1a2e,stroke:#58a6ff,color:#58a6ff
style LIK fill:#1a0a29,stroke:#bc8cff,color:#bc8cff
style POST fill:#1a0a29,stroke:#bc8cff,color:#bc8cff
style VOI fill:#2e2410,stroke:#e3b341,color:#e3b341
style DEC fill:#2e1a1a,stroke:#ff7b72,color:#ff7b72
style A1 fill:#0a2910,stroke:#3fb950,color:#3fb950
style A2 fill:#2e2410,stroke:#e3b341,color:#e3b341
style A3 fill:#2e1a1a,stroke:#ff7b72,color:#ff7b72
The grid Bayesian implementation and VoI calculator live in
PHD/ch7/; site-A posteriors for the Gunsan monitoring
campaign are cached at
PHD/ch7/site_a_bayesian_scour_real_mc.json.
The encoder learns a low-dimensional latent representation of the forward-model database, then projects field OMA observations into the same space. Nearest-neighbour retrieval answers the core dissertation question — "which FEM simulation best matches this sensor reading?"
flowchart TD
subgraph TRAIN [" 🧠 TRAINING (offline) "]
direction TB
MC["`**1,794 Monte Carlo runs**
Op³ forward model
scour × soil × capacity`"]
SIG["`**Simulated signatures**
f₁, f₂, mode shapes
per configuration`"]
ENC["`**Supervised encoder**
N-dim → 2-D latent
dim₁ corr = 0.9976`"]
MC --> SIG --> ENC
end
subgraph FIELD [" 📡 FIELD DEPLOYMENT "]
direction TB
OMA["`**OMA on nacelle**
measured f₁, f₂
mode shapes`"]
Z["`**Field point z_obs**
projected into latent space`"]
OMA --> Z
end
subgraph MATCH [" 🎯 NEAREST-NEIGHBOUR RETRIEVAL "]
direction TB
KNN["`**Top-K configurations**
closest simulations
in latent space`"]
PARAMS["`**Inferred state**
scour S · Gₛ · φ · boundary`"]
KNN --> PARAMS
end
ENC --> KNN
Z --> KNN
RESULT["`**🔎 FIELD RESULT · SiteA**
S/D ≈ 0.06–0.13
→ continue monitoring`"]
PARAMS --> RESULT
style MC fill:#0a2910,stroke:#3fb950,color:#3fb950
style SIG fill:#0a2910,stroke:#3fb950,color:#3fb950
style ENC fill:#1a0a29,stroke:#bc8cff,color:#bc8cff
style OMA fill:#1a1a2e,stroke:#58a6ff,color:#58a6ff
style Z fill:#1a1a2e,stroke:#58a6ff,color:#58a6ff
style KNN fill:#1a0a29,stroke:#bc8cff,color:#bc8cff
style PARAMS fill:#2e2410,stroke:#e3b341,color:#e3b341
style RESULT fill:#2e1a1a,stroke:#ff7b72,color:#ff7b72
Training scripts and cached encoder weights live under
PHD/ch8/; the SiteA field retrieval with S/D ≈ 0.06–0.13
is the Chapter 8 headline result.
Op³ bundles 9 reference decks (NREL 5 MW Baseline · OC3 Monopile ·
IEA-scaled 1.72/1.79/2.3/2.8 MW · Vestas V27 · SiteA 4 MW tripod)
with every OpenFAST input deck verified structurally complete at
commit time. See
validation/benchmarks/NREL_BENCHMARK.md
for per-model verification status.
NREL / IEA / Vestas deck inventory
| Model | Power | Rotor | Foundation | Path | Status |
|---|---|---|---|---|---|
| NREL 5MW Baseline (fixed base) | 5.0 MW | 126 m | Fixed base | nrel_reference/openfast_rtest/5MW_Baseline/ |
✅ |
| NREL 5MW OC3 Monopile + WavesIrr | 5.0 MW | 126 m | Monopile | nrel_reference/openfast_rtest/5MW_OC3Mnpl_DLL_WTurb_WavesIrr/ |
✅ |
| NREL 1.72-103 | 1.72 MW | 103 m | Land-based monopile | nrel_reference/iea_scaled/NREL-1.72-103/ |
✅ |
| NREL 1.79-100 | 1.79 MW | 100 m | Land-based monopile | nrel_reference/iea_scaled/NREL-1.79-100/ |
✅ |
| NREL 2.3-116 | 2.3 MW | 116 m | Land-based monopile | nrel_reference/iea_scaled/NREL-2.3-116/ |
✅ |
| NREL 2.8-127 (HH 87 m) | 2.8 MW | 127 m | Land-based monopile | nrel_reference/iea_scaled/NREL-2.8-127/OpenFAST_hh87/ |
✅ |
| NREL 2.8-127 (HH 120 m) | 2.8 MW | 127 m | Land-based monopile | nrel_reference/iea_scaled/NREL-2.8-127/OpenFAST_hh120/ |
✅ |
| Vestas V27 (historical baseline) | 225 kW | 27 m | Land-based | nrel_reference/vestas/V27/ |
✅ |
| SiteA 4 MW class (subject under test) | 4 MW class | 136 m | Tripod suction bucket | site_a_ref4mw/openfast_deck/ |
tested |
SiteA vs NREL side-by-side
Full comparison in
validation/benchmarks/SITE_A_VS_NREL.md.
| Property | NREL 5MW r-test | NREL OC3 Monopile | NREL 2.8-127 | SiteA 4 MW class |
|---|---|---|---|---|
| Rated power | 5.00 MW | 5.00 MW | 2.80 MW | 4.20 MW |
| Rotor diameter | 126.0 m | 126.0 m | 127.0 m | 136.0 m |
| Hub height | 90.0 m | 90.0 m | 87.6 m | 96.3 m |
| Rated rotor speed | 12.1 rpm | 12.1 rpm | 10.6 rpm | 13.2 rpm |
| First FA natural freq (Hz) | 0.324 | 0.276 | 0.290 | 0.244 |
| Foundation | Fixed | Monopile | Fixed | Tripod suction bucket |
| SSI coupling | none | none | none | OpenSees BNWF → SubDyn |
| Scour parameterization | none | none | none | 9 levels, 0-4 m |
block-beta
columns 1
block:APP["🎛️ APPLICATION LAYER"]
columns 3
COM["composer\nTowerModel\neigen · pushover · transient"] FND["foundations\nFactory + dataclass"] SAC["sacs_interface\nJacket deck parser"]
end
space
block:STD["📐 STANDARDS LAYER"]
columns 4
PISA["PISA\nBurd · Byrne 2020"] CYC["cyclic\nHardin-Drnevich\nHSsmall"] DNV["DNV · ISO\nAPI · OWA"] CAP["capacity\nVesic · Fu Bienen"]
end
space
block:PHYS["🏗️ FOUNDATION PHYSICS — four modes"]
columns 4
MA["Mode A\nFixed base"] MB["Mode B\n6×6 lumped K"] MC["Mode C\nDistributed BNWF"] MD["Mode D ★\nDissipation-weighted"]
end
space
block:SOLV["🔧 SOLVER BRIDGES"]
columns 3
OSI["opensees_foundations\nOpenSeesPy builder"] OFC["openfast_coupling\n→ SubDyn 6×6"] OGI["optumgx_interface\n(license holders only)"]
end
space
block:UQ_LAYER["📊 UQ + CALIBRATION"]
columns 3
PRO["Propagation\nMonte Carlo 1,794"] PCE["PCE\nHermite expansion"] BAY["Bayesian\nGrid calibration"]
end
APP --> STD
STD --> PHYS
PHYS --> SOLV
SOLV --> UQ_LAYER
Please cite both the software and the dissertation. A
CITATION.cff is provided for automatic reference
management.
BibTeX
@software{kim2026op3,
author = {Kim, Kyeong Sun},
title = {Op³: OptumGX-OpenSeesPy-OpenFAST integrated numerical
modeling framework for offshore wind turbines},
year = {2026},
version = {1.0.0-rc2},
url = {https://github.com/ksk5429/numerical_model},
doi = {10.5281/zenodo.19476542},
}
@phdthesis{kim2026dissertation,
author = {Kim, Kyeong Sun},
title = {Digital Twin Encoder for Prescriptive Maintenance of
Offshore Wind Turbine Foundations},
school = {Seoul National University},
year = {2026},
type = {Ph.D. Dissertation},
}Op³ is archived on Zenodo for every tagged GitHub release via the
GitHub-Zenodo integration.
The metadata Zenodo uses is in .zenodo.json and
CITATION.cff.
Concept DOI (always points at the latest release):
10.5281/zenodo.19476542
Current releases:
- v0.3.2 (2026-04-08) — Track C industry-grade, see CHANGELOG.md. Each subsequent tag produces an independent version-specific DOI underneath the concept DOI above.
- v0.3.1 (2026-04-08) — Real Bergua / McAdam / Byrne references
- v0.3.0 (2026-04-08) — Track C initial release
Cite the concept DOI for general reference and the version- specific DOI (visible on the Zenodo record landing page) for reproducibility-critical work.
Code in this repository is released under the MIT License.
NREL reference models bundled in nrel_reference/
are redistributed under their original NREL / Apache 2.0 / public
domain licenses. See each subdirectory's README for specifics. The
NREL 5MW Baseline and OC3 monopile decks are from the OpenFAST r-test
suite (Apache 2.0). The IEA-scaled decks are from the NREL Reference
Wind Turbines repository (CC BY).
OpenSeesPy and OpenFAST are separate open-source projects with their own licenses (BSD 3-Clause and Apache 2.0 respectively). This repository does not redistribute their source code.
OptumGX is commercial software; an academic license is required
for the optumgx Python API imported by scripts in
op3/optumgx_interface/. This repository
does not redistribute OptumGX and contains no OptumGX binary artifacts.
Kyeong Sun Kim · Department of Civil and Environmental Engineering,
Seoul National University · kyeongsunkim@snu.ac.kr
Report reproduction issues via GitHub Issues. The author commits to maintaining this repository in working order for at least three years after the dissertation defense.
Funded by the Korea Electric Power Corporation (KEPCO) under "Natural frequency-based scour monitoring for offshore wind turbine foundations." Developed at Seoul National University. NREL reference decks redistributed under their original licences.
Field case study data (Gunsan / KEPCO / MMB / Unison)
The 4 MW-class offshore wind turbine used for the site-specific validation in Chapters 4–8 of the dissertation is the Gunsan Offshore Demonstration Wind Farm unit operated by KEPCO Research Institute (KEPRI) with foundation and support-structure design by Hyundai E&C and Mirae & Company (MMB), and nacelle/rotor hardware from Unison Co., Ltd. (UNISON U136, 4.2 MW, 136 m rotor, 95 m HH, three-bucket suction caisson tripod). Structural drawings, BOM, and geotechnical CPT/OMA data were provided by KEPRI/MMB/Unison for academic use under the KEPCO research agreement.
Proprietary numerical values (tower segment schedule, bucket
OD/skirt length/C-to-C spacing, nacelle mass, site coordinates,
SubDyn 6×6 K matrix) are not redistributed in this public
repository; the framework code loads them at runtime from a private
data tree via op3.data_sources (OP3_PHD_ROOT /
OP3_TOWER_SEGMENTS_CSV). The framework itself (foundation modes
A/B/C/D, PISA/Hardin-Drnevich/HSsmall, Mode D formulation, OpenFAST
coupling, OC6/PISA benchmarks) is fully reproducible with the
shipped NREL 5 MW reference turbine.
If a publication uses Gunsan case study outputs, please additionally acknowledge KEPCO Research Institute, MMB, and Unison Co., Ltd.
Citations for bundled references
- NREL reference wind turbines — NREL/TP-500-38060 (5 MW), NREL/TP-5000-75698 (IEA 15 MW)
- OC6 Phase II benchmark — Bergua et al., NREL/TP-5000-79989 (2021)
- PISA design framework — Burd et al., 2020; Byrne et al., 2020
- HSsmall constitutive model — Benz, 2007