# Transferring Geophysical Hybrid UQ (Block Bootstrap + Geostatistical Initialization + Homotopy) to Calibrate Reduced-Order Floating Wind Models: A Pre-Registered Validation on a 2-DOF Oscillator

## Metadata

- **SPORE ID**: SPR-2026-4F68
- **Domaines**: Geophysical and Geoelectrical Methods x Ocean Engineering
- **Date de generation**: 2026-08-09
- **Panel consensus score**: 6.0/10
- **Novelty score**: 0.85/1.0
- **Panel verdict**: publish_brief

## Abstract

If the geophysical hybrid UQ framework (block-wise bootstrap, geostatistical initial model generation, homotopic continuation) is applied to calibrate reduced-order floating offshore wind structural models, then the resulting posterior predictive distribution of mooring loads will exhibit (a) a coverage probability of the true 99th percentile within [0.94, 0.99] and (b) a median bias reduction of at least 25% compared to a single best-fit calibration, because the hybrid method explicitly separates aleatory (data resampling) from epistemic (initial model ensemble) uncertainty, and homotopy avoids convergence to local minima.

The proposed mechanism involves 4 causal steps: (1) Block bootstrap resampling of measured load/motion time-series data (e.g., tower -> (2) Geostatistical generation (co-kriging with a defined variogram on a normalized p -> (3) For each resampled dataset, a homotopic continuation solver is initialized from  -> (4) The ensemble of converged solutions across all datasets and initializations form.

Literature grounding on 14 verified references yields a novelty score of 0.85 (novel). A 3-phase experimental protocol (budget: €20k-80k, timeline: 8-14 months) is proposed, starting with in silico validation. A panel of 5 expert reviewers reached a consensus score of 6.0/10.

## 1. Hypothese et mecanisme propose

### 1.1 Formulation formelle

If the geophysical hybrid UQ framework (block-wise bootstrap, geostatistical initial model generation, homotopic continuation) is applied to calibrate reduced-order floating offshore wind structural models, then the resulting posterior predictive distribution of mooring loads will exhibit (a) a coverage probability of the true 99th percentile within [0.94, 0.99] and (b) a median bias reduction of at least 25% compared to a single best-fit calibration, because the hybrid method explicitly separates aleatory (data resampling) from epistemic (initial model ensemble) uncertainty, and homotopy avoids convergence to local minima.

### 1.2 Variables

**Variables independantes :**

| Variable | Type | Plage | Unite |
|----------|------|-------|-------|
| Calibration method | categorical | Hybrid UQ (block bootstrap + geostatistical init + homotopy) vs. Single best-fit (least squares) vs. Hamiltonian MCMC | N/A (categorical) |
| Initial model ensemble size | continuous | 50 - 200 | ensemble members |
| Block length for bootstrap | continuous | 10 - 50 | time steps (s) |

**Variables dependantes :**

| Variable | Type | Direction attendue | Unite |
|----------|------|-------------------|-------|
| Coverage probability of the true 99th percentile of mooring load | continuous | increase | probability (0-1) |
| Median absolute bias of predicted 99th percentile mooring load | continuous | decrease | kN |
| Root Mean Square Error (RMSE) of predicted load time series | continuous | decrease | kN |
| Convergence rate of inverse solver | continuous | increase | iterations to convergence |

### 1.3 Chaine causale

1. Step 1: Block bootstrap resampling of measured load/motion time-series data (e.g., tower base bending moment, mooring tension) generates B=1000 resampled datasets, preserving the autocorrelation and periodic structure (using Variable Bandpass Periodic Block Bootstrap for wave-frequency components).
1. Step 2: Geostatistical generation (co-kriging with a defined variogram on a normalized parameter space, or a Latin Hypercube design from expert priors) creates an ensemble of N=100 initial parameter sets (mooring stiffness, damping, hydrodynamic coefficients) that spans the plausible parameter space, representing epistemic uncertainty.
1. Step 3: For each resampled dataset, a homotopic continuation solver is initialized from each of the N initial parameter sets, tracing a path from a simplified (convex) objective function to the full nonlinear objective function, avoiding local minima and converging to a set of candidate posterior solutions.
1. Step 4: The ensemble of converged solutions across all datasets and initializations forms a posterior distribution of parameters and predicted loads, from which the 99th percentile of mooring load is extracted and compared against the true value from a high-fidelity simulation or experimental data.

**Hypotheses cles :**

- Assumption 1: The reduced-order model (e.g., 2-DOF oscillator) is a sufficient representation of the dominant dynamics for the load quantity of interest (mooring tension).
- Assumption 2: The block bootstrap correctly captures the aleatory variability of the measurement noise and environmental loading, given a correctly chosen block length.
- Assumption 3: The geostatistical prior (or expert-based Latin Hypercube) encompasses the true parameter values, ensuring the homotopy solver can reach the global optimum.
- Assumption 4: The homotopy path is smooth and does not introduce artificial local minima that are not present in the original problem.

**Inconnues identifiees :**

- We do not know the optimal block length for the bootstrap that balances variance reduction and preservation of autocorrelation for a given sea state.
- We do not know the minimum ensemble size of initial models required to achieve the target coverage probability.
- We do not know if the homotopy continuation method will scale efficiently to a full 15-DOF floating wind model without a significant increase in computational cost.

### 1.4 Conditions aux limites

- **The reduced-order model must be linearizable around the operating point for the homotopy path to be smooth.** — Strong nonlinearities (e.g., mooring line slack-taut transitions) can create discontinuous objective functions, breaking the homotopy continuation path.
- **The measurement noise must be stationary and ergodic within each bootstrap block.** — Non-stationary noise (e.g., transient events) violates the block bootstrap's assumption of resampling from a stationary distribution.
- **The parameter space must be bounded and continuous for the geostatistical initial model generation to be valid.** — Discrete or categorical parameters (e.g., number of mooring lines) cannot be interpolated via co-kriging.
- **The true parameter values must lie within the convex hull of the initial model ensemble.** — If the initial ensemble does not cover the true parameters, the homotopy solver cannot converge to the global optimum, leading to biased predictions.

### 1.5 Cadre theorique

Bayesian inverse problem theory with explicit separation of aleatory and epistemic uncertainty, operationalized via frequentist resampling (bootstrap) and ensemble initialization (geostatistics), within a homotopy-based global optimization framework.

## 2. Etat de l'art et positionnement

### 2.1 Travaux les plus proches

- **[2024] Uncertainty quantification in electrical resistivity tomography inversion: Hybridizing block-wise bootstrapping with geostatistics** — [10.1093/gji/ggae347](https://doi.org/10.1093/gji/ggae347)
  - Similarite: related
  - Difference cle: This paper applies the hybrid block-wise bootstrap + geostatistical initial models + homotopic continuation to geophysical ERT inversion. The hypothesis proposes transferring this exact framework to floating offshore wind structural model calibration, which is a different domain with different data types (time-series vs. spatial) and different parameters (structural/mooring vs. resistivity).
- **[2022] Uncertainty quantification for regularized inversion of electromagnetic geophysical data. Part I: Motivation and Theory** — [10.1093/gji/ggac241](https://doi.org/10.1093/gji/ggac241)
  - Similarite: related
  - Difference cle: This paper uses a 'randomize-then-optimize' (RTO) approach for UQ in EM geophysical inversion, which is a different algorithmic strategy than block-wise bootstrap + geostatistical initial models. The hypothesis specifically focuses on the block-wise bootstrap + geostatistics combination, not RTO.
- **[2021] Ensemble-Based Electrical Resistivity Tomography with Data and Model Space Compression** — [10.1007/s00024-021-02730-1](https://doi.org/10.1007/s00024-021-02730-1)
  - Similarite: related
  - Difference cle: This paper uses an ensemble-based approach (likely Ensemble Kalman Filter or similar) for ERT UQ, which is a different method than the block-wise bootstrap + geostatistics + homotopy approach. The hypothesis is specifically about transferring the latter combination.

### 2.2 Base de preuves

- **[2024] Uncertainty quantification in electrical resistivity tomography inversion: Hybridizing block-wise bootstrapping with geostatistics** — [10.1093/gji/ggae347](https://doi.org/10.1093/gji/ggae347)
  - Type: direct | Citations: 8
  - This is the core paper that the hypothesis builds upon. It directly demonstrates the hybrid block-wise bootstrap + geostatistical initial models + homotopic continuation approach for UQ in ERT inversion, addressing both epistemic and aleatory uncertainties.
- **[1996] Recent developments in bootstrapping time series** — [10.1080/07474930008800457](https://doi.org/10.1080/07474930008800457)
  - Type: indirect | Citations: 367
  - Provides the theoretical foundation for block bootstrap methods for time series data. This is critical for the transferability of the geophysical method (which uses spatial data) to ocean engineering (which uses temporal data).
- **[2024] Periodically correlated time series and the Variable Bandpass Periodic Block Bootstrap** — [10.1371/journal.pone.0310563](https://doi.org/10.1371/journal.pone.0310563)
  - Type: indirect | Citations: 11
  - Addresses a key challenge in the transfer: ocean engineering data (e.g., wave loads) are often periodically correlated (e.g., due to wave frequency). This paper provides a method to handle such data with block bootstrap.
- **[2025] The Variable Multiple Bandpass Periodic Block Bootstrap for Time Series With Multiple Periodic Correlations** — [10.1155/jpas/9968540](https://doi.org/10.1155/jpas/9968540)
  - Type: indirect | Citations: 9
  - Extends the previous paper to handle multiple periodic correlations, which is relevant for floating wind structures that may have multiple excitation frequencies (e.g., wave, wind, and structural).
- **[2023] The Many Forms of Co-kriging: A Diversity of Multivariate Spatial Estimators** — [10.1007/s11004-023-10104-7](https://doi.org/10.1007/s11004-023-10104-7)
  - Type: indirect | Citations: 22
  - Provides the theoretical basis for the geostatistical generation of initial models. Co-kriging can be used to generate correlated initial model parameter sets (e.g., mooring stiffness and damping) from sparse measurements.
- **[2024] Framework for Aleatory Variability and Epistemic Uncertainty for the Ground-Motion Characterization Based on the Level of Simplification** — [10.1785/0120240141](https://doi.org/10.1785/0120240141)
  - Type: analogous | Citations: 15
  - Provides a framework for separating aleatory and epistemic uncertainty in a different domain (ground-motion characterization). This supports the hypothesis's core mechanism of formally separating these two uncertainty types.
- **[2022] Towards the NASA UQ Challenge 2019: Systematically forward and inverse approaches for uncertainty propagation and quantification** — [10.1016/j.ymssp.2021.108387](https://doi.org/10.1016/j.ymssp.2021.108387)
  - Type: indirect | Citations: 14
  - Demonstrates systematic forward and inverse UQ approaches in a complex engineering system, which is analogous to the proposed approach for floating wind structures.
- **[2024] Adaptive Homotopy-Based Inverse Model for the Geometry Scaling of Microwave Filters** — [10.1109/TMTT.2023.3309391](https://doi.org/10.1109/TMTT.2023.3309391)
  - Type: indirect | Citations: 7
  - Demonstrates the use of homotopy continuation (HC) in an inverse problem to avoid local minima and reuse pre-obtained designs. This supports the hypothesis's use of homotopic continuation for the nonlinear inverse problem.
- **[2020] Structural dynamic nonlinear model and parameter identification based on the stiffness and damping marginal curves** — [10.1002/stc.2540](https://doi.org/10.1002/stc.2540)
  - Type: indirect | Citations: 9
  - Addresses parameter identification in structural dynamics, which is the core inverse problem in ocean engineering. This supports the feasibility of identifying structural parameters from measured responses.
- **[2019] A Probabilistic Uncertainty Estimation Method for Turbulence Parameters Measured by Hot-Wire Anemometry in Short-Duration Wind Tunnels** — [10.1115/1.4044780](https://doi.org/10.1115/1.4044780)
  - Type: indirect | Citations: 8
  - Demonstrates a probabilistic UQ method that separates measurement-related uncertainty from statistical treatment uncertainty in a time-series context, which is analogous to separating aleatory and epistemic uncertainty in load measurements.

### 2.3 Contre-preuves et limitations connues

- **[addressable] Stochastic full waveform inversion with deep generative prior for uncertainty quantification**
  - This paper uses deep generative models as priors for UQ in geophysical inversion, which is an alternative to geostatistical initial models. This does not contradict the hypothesis but suggests that the geostatistical component may be replaceable by more modern methods.
- **[minor] Bootstrap resampling approach to disaggregate analysis of road crashes in Hong Kong.**
  - This paper applies bootstrap resampling to a different domain (road safety) and shows that it improves parameter estimates. It does not contradict the hypothesis but highlights that bootstrap methods are domain-agnostic and can be applied broadly.

### 2.4 Evaluation de nouveaute

- **Score**: 0.85/1.0
- **Verdict**: novel

## 3. Predictions falsifiables

| # | Prediction | Borne quantitative | Methode | H0 | Test statistique |
|---|-----------|-------------------|---------|-----|-----------------|
| 1 | The hybrid UQ method will yield a posterior predictive interval for the 99th percentile of mooring load with a coverage probability of at least 0.94. | Coverage probability ≥ 0.94 (lower bound of 95% CI) for the true 99th percentile, based on 100 synthetic datasets. | Synthetic data from a 2-DOF oscillator model with known parameters and added noise; 100 Monte Carlo repetitions. | H0: The true coverage probability is ≤ 0.90 (i.e., the method is not better than a naive 90% interval). | One-sided z-test for proportions, alpha = 0.05, with power analysis pre-registered to detect a difference of 0.04 from the null. |
| 2 | The hybrid UQ method will reduce the median absolute bias of the predicted 99th percentile mooring load by at least 25% compared to a single best-fit calibration. | Median absolute bias reduction ≥ 25% (e.g., from 10 kN to ≤ 7.5 kN) across 100 synthetic datasets. | Comparison of predicted vs. true 99th percentile load from 100 synthetic datasets. | H0: The median bias of the hybrid method is ≥ 90% of the median bias of the single best-fit method (i.e., no meaningful improvement). | One-sided paired Wilcoxon signed-rank test, alpha = 0.05, with a priori power analysis for a 25% effect size. |
| 3 | The homotopy continuation solver will converge to a solution with a lower final objective function value (RMSE) than a standard multi-start gradient descent in at least 80% of the 1000 bootstrap resamples. | Success rate ≥ 80% (e.g., 800 out of 1000 resamples) where 'success' is defined as achieving a final RMSE at least 5% lower than the best of 50 multi-start runs. | Comparison of final RMSE values from homotopy vs. multi-start on identical bootstrap resamples. | H0: The success rate of homotopy is ≤ 50% (no better than chance). | Exact binomial test, alpha = 0.05. |

## 4. Protocole experimental

**Timeline globale**: 8-14 months
**Budget global**: €20k-80k

### 4.1 Phase 1 — In Silico Validation on Synthetic 2-DOF Data

**Objectif**: To falsify the core mechanism of the hybrid UQ framework in a controlled, high-fidelity simulation environment. This phase tests whether the combination of block bootstrap, geostatistical initialization, and homotopy continuation can achieve the target coverage and bias reduction on a known ground truth, before any physical experiments are considered.

**Methodologie**: 1. **Model Setup**: Implement a 2-DOF oscillator model (surge + pitch) in Python using `scipy.integrate.solve_ivp`. The model will include a nonlinear mooring stiffness term (e.g., a piecewise linear or polynomial function) to represent slack-taut transitions. Define 'true' parameters (mooring stiffness, damping, hydrodynamic added mass) and simulate a 1-hour time series of wave forcing (JONSWAP spectrum, Hs=3m, Tp=10s) and the resulting mooring load response. Add realistic measurement noise (Gaussian, σ=2% of signal amplitude).
2. **Ground Truth Generation**: Generate 100 independent synthetic datasets, each with a different noise realization but the same true parameters. The true 99th percentile of mooring load is computed from a noiseless, long-duration (10-hour) simulation.
3. **Method Implementation**:
   - **Single Best-Fit**: Use `scipy.optimize.least_squares` with a multi-start (50 random starts) approach to calibrate the model to each dataset.
   - **Hybrid UQ**: For each dataset, generate B=1000 bootstrap resamples using the Variable Bandpass Periodic Block Bootstrap (VBPBB) algorithm (implemented in Python, based on the referenced 2024/2025 papers). For each resample, generate N=100 initial parameter sets using Latin Hypercube Sampling (LHS) from a bounded prior space (uniform ±50% around true values). Apply a homotopy continuation solver: start from a simplified quadratic objective (L2 norm of residuals), then gradually transition to the full nonlinear objective (L2 norm with a robust loss function, e.g., Huber) using a path-following algorithm (e.g., `scipy.optimize.least_squares` with a continuation parameter).
4. **Evaluation**: For each of the 100 datasets, compute the posterior predictive distribution of the 99th percentile mooring load. Calculate the coverage probability (does the 95% credible interval contain the true value?) and the median absolute bias. Compare against the single best-fit results.
5. **Computational Tools**: Python 3.10+, `numpy`, `scipy`, `statsmodels`, `pandas`, `matplotlib`. Use `joblib` for parallelization.

- Cout: €500-2000
- Duree: 4-6 weeks
- Equipement: Standard high-performance workstation (16+ cores, 64GB RAM)
- Logiciels: Python 3.10+, NumPy, SciPy, StatsModels, Matplotlib, Joblib

**Criteres de succes :**

- Coverage probability of the true 99th percentile: ≥ 0.94 (lower bound of 95% CI)
- Median absolute bias reduction vs. single best-fit: ≥ 25% reduction
- Homotopy success rate: ≥ 80% of bootstrap resamples

- **GO**: All three success criteria are met. The hybrid method demonstrates superior coverage and bias reduction in the synthetic environment.
- **NO-GO**: Any of the three success criteria fail. The mechanism is falsified in the ideal case, and the approach should be abandoned or fundamentally rethought.
- **PIVOT**: Coverage is met but bias reduction is marginal (e.g., 15-25%). Consider tuning the homotopy path or the geostatistical prior. If homotopy success rate is <80%, investigate the smoothness of the objective function and consider alternative continuation schemes.

### 4.2 Phase 2 — Minimal Experimental Validation on a Physical 2-DOF Oscillator

**Objectif**: To validate the hybrid UQ framework against real, noisy physical data from a simplified mechanical system. This phase tests the assumptions about noise characteristics and model adequacy in a physical setting, bridging the gap between simulation and full-scale application.

**Methodologie**: 1. **Physical Setup**: Construct a small-scale 2-DOF mechanical oscillator (e.g., a cart on a rail with a spring and damper, mounted on a shake table). The system will have a tunable nonlinear spring (e.g., a spring with a variable stiffness or a belt-driven mechanism). The 'mooring load' is measured via a load cell (e.g., Futek LSB200) at the spring attachment point. The 'wave forcing' is the shake table motion (e.g., Quanser Shake Table II).
2. **Data Collection**: Excite the system with a known, repeatable random signal (e.g., band-limited white noise or a JONSWAP-like spectrum) for 10 minutes. Record the forcing signal (shake table displacement) and the response (load cell force) at 100 Hz. Repeat this 20 times to create 20 independent datasets.
3. **Calibration**: Use the same calibration methods as Phase 1 (single best-fit and hybrid UQ) to calibrate the 2-DOF model to each of the 20 datasets. The 'true' 99th percentile is estimated from a very long (1-hour) experimental run.
4. **Evaluation**: Compute the coverage probability and bias reduction across the 20 datasets. Compare the performance of the hybrid method against the single best-fit.
5. **Tools**: Use MATLAB/Simulink or Python with a real-time interface (e.g., `pyserial`) for data acquisition. Use the same Python calibration code from Phase 1, adapted to read the experimental data format.

- Cout: €5k-15k
- Duree: 1-2 months
- Equipement: Small-scale 2-DOF oscillator test rig, Shake table (e.g., Quanser Shake Table II), Load cell (e.g., Futek LSB200, 100N range), Data acquisition system (e.g., NI USB-6009), Power amplifier for shake table
- Logiciels: MATLAB/Simulink or Python with real-time DAQ libraries, Existing Python calibration code from Phase 1

**Criteres de succes :**

- Coverage probability of the true 99th percentile: ≥ 0.90 (relaxed from Phase 1 due to model inadequacy)
- Median absolute bias reduction vs. single best-fit: ≥ 20% reduction
- Model adequacy (RMSE of the best-fit model): RMSE < 10% of the signal's standard deviation

- **GO**: The hybrid method shows a clear improvement over single best-fit (coverage ≥ 0.90 and bias reduction ≥ 20%). The model is deemed adequate (RMSE < 10%).
- **NO-GO**: The hybrid method fails to improve coverage or bias. This indicates a fundamental issue with the method's assumptions (e.g., noise non-stationarity, model inadequacy) that cannot be resolved by tuning.
- **PIVOT**: Coverage is met but bias reduction is <20%. This suggests the method is capturing uncertainty well but the point estimate is biased. Consider improving the model structure or the prior distribution. If model adequacy fails (RMSE > 10%), the 2-DOF model is insufficient; a higher-fidelity model is needed.

### 4.3 Phase 3 — Full Experimental Protocol on a Scaled Floating Offshore Wind Platform

**Objectif**: To validate the hybrid UQ framework on a realistic, scaled floating offshore wind (FOW) platform in a wave basin, demonstrating its practical utility for mooring load prediction. This phase tests the scalability and robustness of the method under realistic environmental loading and complex hydrodynamics.

**Methodologie**: 1. **Model Setup**: Use a scaled FOW platform model (e.g., a 1:50 scale semi-submersible or spar) in a wave basin (e.g., MARIN, IFREMER). The model will be instrumented with load cells on the mooring lines, accelerometers, and a motion capture system. The 'reduced-order model' will be a 3-DOF (surge, heave, pitch) lumped-mass model, calibrated to the physical model.
2. **Data Collection**: Conduct a series of wave basin tests with a range of sea states (e.g., operational, extreme, and survival conditions). For each sea state, run a 3-hour (scaled) test and record all sensor data. Repeat each test 5 times to assess repeatability.
3. **Calibration**: Apply the hybrid UQ framework (with the VBPBB and homotopy solver from Phase 1) to calibrate the reduced-order model to the measured data. Use the full 3-hour time series for calibration. The 'true' 99th percentile of mooring load is estimated from the full 3-hour record.
4. **Validation**: Compare the posterior predictive distribution of the 99th percentile mooring load against the measured value. Calculate coverage probability and bias reduction. Compare against a single best-fit calibration.
5. **Scalability Test**: Assess the computational cost of the hybrid method (wall-clock time) and compare it to the theoretical scaling for a 15-DOF model. This will inform future applications.
6. **Tools**: Use the wave basin's data acquisition system. Use the Python calibration code, optimized for parallel processing on a high-performance computing cluster.

- Cout: €50k-200k+
- Duree: 6-12 months
- Equipement: Scaled FOW platform model (1:50), Wave basin facility with wavemakers, Mooring line load cells (e.g., 100N-1kN range), Motion capture system (e.g., Qualisys, Vicon), Data acquisition system (e.g., NI PXI)
- Logiciels: Python calibration code (from Phase 1), High-performance computing cluster (for parallel bootstrap runs), Wave basin data acquisition software

**Criteres de succes :**

- Coverage probability of the true 99th percentile across all sea states: ≥ 0.94 (as per original hypothesis)
- Median absolute bias reduction vs. single best-fit: ≥ 25% reduction
- Computational scalability: Wall-clock time for calibration < 1 week on a 64-core cluster

- **GO**: Both primary success criteria (coverage ≥ 0.94 and bias reduction ≥ 25%) are met. The method is deemed scalable for practical use.
- **NO-GO**: The method fails to meet the coverage or bias criteria in the realistic wave basin environment. This indicates a fundamental limitation of the approach for FOW applications.
- **PIVOT**: Coverage is met but bias reduction is <25%. This suggests the prior distribution is too narrow or the model is biased. Consider expanding the prior or using a more complex reduced-order model. If scalability fails, investigate algorithmic optimizations (e.g., reduced-order bootstrap, surrogate models).

### 4.4 Quick start : comment demarrer aujourd'hui

- **Peut demarrer maintenant**: Oui
- **Premiere action**: Create a new Python virtual environment and install the core scientific stack (numpy, scipy, pandas, statsmodels, matplotlib, joblib). Then, write a simple 2-DOF oscillator simulation script (surge + pitch) with a nonlinear mooring stiffness term and generate a single synthetic dataset with known parameters and added Gaussian noise.
- **Outils**: Python 3.10+, pip, Virtual environment tool (venv or conda)
- **Donnees ouvertes**: No external data needed; generate synthetic data. For validation of the VBPBB algorithm, use the 'periodically correlated time series' examples from the referenced 2024/2025 papers, which may have supplementary code or data available.

## 5. Analyse d'impact

### 5.1 Impact scientifique

Novelty score: 0.85/1.0 (novel)

### 5.2 Applications industrielles et marche

**Score industriel**: 6.0/10

**Forces :**
- The explicit separation of aleatory (bootstrap) and epistemic (geostatistical ensemble) uncertainty directly addresses a critical pain point for FOW developers: the current industry practice of using single best-fit calibration with safety factors leads to either over-engineered mooring systems (cost overruns of 10-15% on CAPEX) or under-predicted fatigue loads (warranty and insurance risks). The target of 25% bias reduction on the 99th percentile mooring load translates to a tangible reduction in mooring line cross-section or chain grade, which is a direct cost driver.
- The homotopy continuation component offers a computational moat: standard multi-start gradient descent in commercial tools (e.g., OrcaFlex, DeepLines) often fails to find global minima in high-dimensional soil-mooring interaction models. A documented 80% success rate in escaping local minima would allow faster design iterations (reducing engineering hours per load case from days to hours), which is a compelling productivity sell to engineering consultancies like DNV or Aker Solutions.

**Faiblesses :**
- The market is a niche within a niche: the primary buyers are the 5-7 major FOW developers (Equinor, Shell, RWE, Iberdrola) and their engineering contractors. The total addressable market for calibration software/consulting services is likely under €50M/year, not the €1B+ that would attract major software vendors. The 2-DOF oscillator validation, while rigorous, does not prove scalability to the 6-DOF + 1000+ node finite element models used in real FOW design, creating a significant 'valley of death' between Phase 2 and Phase 3.
- The IP position is weak: block bootstrap and geostatistical simulation are established public-domain techniques (used in mining geostatistics since the 1980s). The novelty is purely in the combination and application domain. Without a strong patent on the specific homotopy-ensemble interaction or a proprietary pre-trained surrogate model, competitors (e.g., a well-funded startup like Akselos or a university spin-off) can replicate the approach within 12-18 months of seeing the published protocol.

**Recommandation**: Proceed with Phase 1 (€2k, 4 weeks) as a pure technical de-risking exercise, but do not commit to Phase 3 (wave basin) unless a co-funding agreement is signed with an industrial partner. The commercial strategy should pivot from selling a standalone software tool to embedding this UQ capability as a differentiated 'risk assessment service' within a larger engineering consultancy offering, targeting the specific pain point of mooring system certification. The realistic timeline to a revenue-generating product is 24-36 months, contingent on the 2-DOF results being reproducible on a higher-fidelity model (e.g., a 3-DOF lumped-mass mooring model) before the expensive wave basin test.

### 5.3 Opportunites de financement

| Programme | Agence | Fit | Budget type | Taux succes |
|-----------|--------|-----|-------------|-------------|
| ERC Proof of Concept Grant 2025 | European Research Council | 0.75 | 150 000 EUR (forfait unique) | Environ 40% (taux de succes historique pour les PoC, mais reserve aux laureats ERC existants) |
| ANR Jeunes Chercheuses et Jeunes Chercheurs (JCJC) 2025 | Agence Nationale de la Recherche (France) | 0.65 | 150 000 a 300 000 EUR sur 36-48 mois | Environ 15-20% (taux de succes JCJC 2024) |
| Marie Sklodowska-Curie Actions (MSCA) Postdoctoral Fellowships 2025 - European Fellowships | European Commission (Horizon Europe) | 0.55 | Environ 200 000 EUR pour 24 mois (salaire + couts de recherche) | Environ 15-20% (taux de succes 2023) |

- **ERC Proof of Concept Grant 2025** (European Research Council): Ce programme est ideal pour tester le potentiel de rupture d'une methode issue d'un ERC (ici, probablement un ERC en geophysique ou UQ). Le budget de 150k EUR couvre largement le besoin de 20-80k EUR, et le pre-enregistrement avec GO/NO-GO est exactement le type de demarche a risque maitrise que l'ERC attend pour un PoC. Il faut cependant etre laureat ERC ou avoir un partenaire laureat.
- **ANR Jeunes Chercheuses et Jeunes Chercheurs (JCJC) 2025** (Agence Nationale de la Recherche (France)): Le format JCJC est adapte a un chercheur porteur unique souhaitant structurer une equipe autour d'une thematique innovante. Le budget est plus eleve que le besoin, mais la duree de 36 mois permet d'elargir le protocole a des essais complementaires (ex: campagnes parametriques) et de renforcer la partie modelisation reduite. Le caractere pre-enregistre et la rigueur statistique sont des atouts pour l'ANR.
- **Marie Sklodowska-Curie Actions (MSCA) Postdoctoral Fellowships 2025 - European Fellowships** (European Commission (Horizon Europe)): Ce programme finance un post-doctorant pour developper un projet de recherche personnalise. L'hypothese est parfaitement adaptee a un chercheur post-doctoral en ingenierie civiere/mecanique avec un background en statistiques. Le budget permet de couvrir les trois phases, y compris l'acces a un bassin de houle (via des collaborations). Le caractere interdisciplinaire (geophysique + genie oceanique) est un plus pour les MSCA.

**Recommandation financement**: Cibler en priorite un financement de type 'Proof of Concept' ou 'Pathfinder' pour valider le mecanisme sur le 2-DOF, tout en construisant des maintenant un consortium avec un bassin d'essais (ex: Centrale Nantes / ECN) et un partenaire industriel pour la phase 3. Le dossier doit insister sur la rupture methodologique (separation aleatoire/epistemique) et sur l'impact potentiel en termes de cout de calcul pour le dimensionnement des ancrages.

## 6. Panel Review Summary

| Reviewer | Score | Verdict | Point cle |
|----------|-------|---------|-----------|
| methodologist | 6.8/10 | weak_accept | The pre-registration of explicit, falsifiable predictions with predefined GO/NO-GO/PIVOT criteria is exemplary. This structure forces a clear decision pathway and mitigates confirmation bias, a rare and commendable practice in computational mechanics. |
| domain_expert | 7.2/10 | weak_accept | La formulation est remarquablement claire et opérationnalisée : les critères de succès (couverture de probabilité [0.94, 0.99], réduction de biais ≥ 25%) sont quantifiés et testables, ce qui est rare et précieux dans la littérature sur l'UQ. Cette pré-spécification force une rigueur épistémique qui manque souvent aux études de calibration. |
| contrarian | 3.5/10 | weak_reject | The explicit separation of aleatory (bootstrap) and epistemic (initial ensemble) uncertainty is methodologically principled and represents a genuine improvement over monolithic UQ approaches that conflate the two. |
| industrialist | 6.0/10 | weak_accept | The explicit separation of aleatory (bootstrap) and epistemic (geostatistical ensemble) uncertainty directly addresses a critical pain point for FOW developers: the current industry practice of using single best-fit calibration with safety factors leads to either over-engineered mooring systems (cost overruns of 10-15% on CAPEX) or under-predicted fatigue loads (warranty and insurance risks). The target of 25% bias reduction on the 99th percentile mooring load translates to a tangible reduction in mooring line cross-section or chain grade, which is a direct cost driver. |
| funding_strategist | 6.5/10 | weak_accept | Originalite methodologique : le transfert d'un cadre UQ hybride geophysique (bootstrap par blocs + initialisation geostatistique + homotopie) vers le calage de modeles d'ordre reduit pour l'eolien flottant est une vraie nouveaute, non couverte par les appels classiques sur les jumeaux numeriques. |

**Consensus score**: 6.0/10
**Verdict final**: publish_brief

### 6.1 Consensus

- Tous les reviewers reconnaissent la qualite methodologique du protocole pre-enregistre avec criteres GO/NO-GO explicites et quantitatifs, ainsi que la separation claire entre incertitudes aleatoires et epistemiques.
- Le consensus est unanime sur le fait que le test 2-DOF est insuffisant pour valider la robustesse de l'approche : le Contrarian demontre que l'homotopie n'est pas necessaire sur un probleme a faible dimension, et le domain expert souligne qu'un modele 6-DOF serait plus discriminant.
- Le manque de puissance statistique (100 jeux de donnees synthetiques) est identifie par le methodologist comme critique pour la revendication principale de couverture, et aucun reviewer ne conteste cette fragilite statistique.
- L'absence de partenaire industriel ou de bassin d'essais identifie des la phase 2 est unanimement vue comme un handicap majeur pour la credibilite et le financement de la phase 3.

### 6.2 Points de desaccord

- Le Contrarian affirme que le bootstrap par blocs est fondamentalement inadapte pour estimer les extremes (impossibilite de generer de nouveaux extremes au-dela des donnees observees), tandis que le methodologist et le domain expert considerent que le probleme peut etre resolu par un terme de 'model discrepancy' explicite ou une modelisation parametrique des queues de distribution.
- Le Contrarian juge la comparaison homotopie vs multi-start comme un 'strawman' (baseline trop faible), alors que l'industrialist y voit un avantage competitif reel pour les outils commerciaux existants (OrcaFlex, DeepLines), suggerant que la faiblesse de la baseline est en fait representative de la pratique industrielle actuelle.
- Le funding strategist considere le budget de 20-80k EUR comme un atout pour un financement de type Proof of Concept, tandis que l'industrialist estime que ce montant est trop faible pour couvrir les couts reels de la phase 3 (bassin de houle) et recommande de conditionner cette phase a un co-financement industriel.

### 6.3 Critical path

Le facteur le plus determinant est la demonstration que l'homotopie est reellement necessaire sur un probleme presentant des minima locaux averes, et que le bootstrap peut capturer l'incertitude sur les extremes. Sans un diagnostic explicite de la topologie de la fonction objectif (montrant des minima multiples) et une solution au probleme de generation d'extremes (par exemple via un modele de valeurs extremes parametrique), l'hypothese reste circulaire et non testable de maniere convaincante. La validation sur un 2-DOF ne peut pas etablir la pertinence de la methode pour le cas 15-DOF reel.

**Recommandation finale**: Le panel recommande de rejeter l'hypothese en l'etat, non pas parce que l'approche est depourvue de merite, mais parce que la validation proposee est structurellement inapte a demontrer la these centrale. Le test 2-DOF ne peut pas etablir la necessite de l'homotopie (probleme probablement unimodal), et le bootstrap par blocs ne peut pas generer les extremes necessaires pour valider la couverture du 99eme percentile. Les critiques convergentes du Contrarian (circularite, strawman) et du methodologist (puissance statistique insuffisante) sont dirimantes. Une version revisee devrait (1) utiliser un modele intermediaire 6-DOF avec couplages non-lineaires demontrant des minima locaux, (2) introduire un terme de 'model discrepancy' ou un modele parametrique de valeurs extremes pour les queues, (3) augmenter le nombre de jeux de donnees synthetiques a au moins 500, et (4) associer un partenaire industriel ou un bassin d'essais des la phase 2. Ces modifications fondamentales depassent le cadre d'une simple revision et necessitent une reformulation substantielle du protocole.

## 7. Gap Manifest residuel

### 7.1 Data gaps

- The specific implementation of homotopic continuation for a floating wind structural model has not been demonstrated; the only homotopy paper (a65b8b3c) is in a different domain (microwave filters).
- The handling of periodically correlated time series (e.g., wave-frequency loads) in the block bootstrap step needs further investigation; the periodic block bootstrap papers (d87adc3e, 05ed68c9) are recent and not yet applied to ocean engineering.

### 7.2 Competence gaps

- Phase 1: Python programming
- Phase 1: Time series analysis
- Phase 1: Bayesian statistics
- Phase 1: Optimization algorithms
- Phase 2: Mechanical engineering
- Phase 2: Experimental design
- Phase 2: Data acquisition and signal processing
- Phase 3: Ocean engineering
- Phase 3: Hydrodynamics
- Phase 3: Model testing
- Phase 3: High-performance computing

### 7.3 Epistemic gaps

- We do not know the optimal block length for the bootstrap that balances variance reduction and preservation of autocorrelation for a given sea state.
- We do not know the minimum ensemble size of initial models required to achieve the target coverage probability.
- We do not know if the homotopy continuation method will scale efficiently to a full 15-DOF floating wind model without a significant increase in computational cost.

## References

[1] Z. Khabaz, R. Ghanati, et al. (2024). *Uncertainty quantification in electrical resistivity tomography inversion: Hybridizing block-wise bootstrapping with geostatistics*. DOI: [10.1093/gji/ggae347](https://doi.org/10.1093/gji/ggae347)
[2] Jeremy Berkowitz, L. Kilian (1996). *Recent developments in bootstrapping time series*. DOI: [10.1080/07474930008800457](https://doi.org/10.1080/07474930008800457)
[3] Edward L Valachovic (2024). *Periodically correlated time series and the Variable Bandpass Periodic Block Bootstrap*. DOI: [10.1371/journal.pone.0310563](https://doi.org/10.1371/journal.pone.0310563)
[4] Edward L Valachovic (2025). *The Variable Multiple Bandpass Periodic Block Bootstrap for Time Series With Multiple Periodic Correlations*. DOI: [10.1155/jpas/9968540](https://doi.org/10.1155/jpas/9968540)
[5] P. Dowd, E. Pardo‐Igúzquiza (2023). *The Many Forms of Co-kriging: A Diversity of Multivariate Spatial Estimators*. DOI: [10.1007/s11004-023-10104-7](https://doi.org/10.1007/s11004-023-10104-7)
[6] Irene Liou, Norman A. Abrahamson (2024). *Framework for Aleatory Variability and Epistemic Uncertainty for the Ground-Motion Characterization Based on the Level of Simplification*. DOI: [10.1785/0120240141](https://doi.org/10.1785/0120240141)
[7] S. Bi, Kui He, et al. (2022). *Towards the NASA UQ Challenge 2019: Systematically forward and inverse approaches for uncertainty propagation and quantification*. DOI: [10.1016/j.ymssp.2021.108387](https://doi.org/10.1016/j.ymssp.2021.108387)
[8] Anlan Liu, Yang Wu, et al. (2024). *Adaptive Homotopy-Based Inverse Model for the Geometry Scaling of Microwave Filters*. DOI: [10.1109/TMTT.2023.3309391](https://doi.org/10.1109/TMTT.2023.3309391)
[9] Zuo-Cai Wang, Yanjun Ding, et al. (2020). *Structural dynamic nonlinear model and parameter identification based on the stiffness and damping marginal curves*. DOI: [10.1002/stc.2540](https://doi.org/10.1002/stc.2540)
[10] E. Boufidi, S. Lavagnoli, et al. (2019). *A Probabilistic Uncertainty Estimation Method for Turbulence Parameters Measured by Hot-Wire Anemometry in Short-Duration Wind Tunnels*. DOI: [10.1115/1.4044780](https://doi.org/10.1115/1.4044780)
[11] Yuke Xie, Hervé Chauris, et al. (2024). *Stochastic full waveform inversion with deep generative prior for uncertainty quantification*. DOI: [10.48550/arXiv.2406.04859](https://doi.org/10.48550/arXiv.2406.04859)
[12] X. Pei, N. Sze, et al. (2016). *Bootstrap resampling approach to disaggregate analysis of road crashes in Hong Kong.*. DOI: [10.1016/j.aap.2015.06.007](https://doi.org/10.1016/j.aap.2015.06.007)
[13] D. Blatter, M. Morzfeld, et al. (2022). *Uncertainty quantification for regularized inversion of electromagnetic geophysical data. Part I: Motivation and Theory*. DOI: [10.1093/gji/ggac241](https://doi.org/10.1093/gji/ggac241)
[14] M. Aleardi, A. Vinciguerra, et al. (2021). *Ensemble-Based Electrical Resistivity Tomography with Data and Model Space Compression*. DOI: [10.1007/s00024-021-02730-1](https://doi.org/10.1007/s00024-021-02730-1)

## Annexes

### A. Detailed Reviewer Reports

#### Methodologist

- **Score**: 6.8/10 | **Verdict**: weak_accept | **Confidence**: 0.88
- **Strengths**: The pre-registration of explicit, falsifiable predictions with predefined GO/NO-GO/PIVOT criteria is exemplary. This structure forces a clear decision pathway and mitigates confirmation bias, a rare and commendable practice in computational mechanics.; The three-phase progression (synthetic → physical oscillator → scaled FOW platform) is a logically sound 'falsification ladder'. It allows for the isolation of methodological flaws (Phase 1) from model inadequacy (Phase 2) and scalability issues (Phase 3), providing a clear diagnostic path.; The explicit use of a 'true' 99th percentile from a noiseless, long-duration simulation in Phase 1 provides a robust ground truth, which is the gold standard for validating coverage properties of any UQ method.
- **Weaknesses**: The statistical power for the primary coverage claim is critically underpowered. Claiming a lower bound of 0.94 for a 95% CI based on only 100 synthetic datasets is statistically fragile. The standard error for a proportion near 0.95 with n=100 is ~0.02, meaning the 95% CI for the coverage estimate itself is roughly [0.91, 0.99]. This is insufficient to distinguish between the target (0.94) and the null (0.90), making the primary hypothesis effectively untestable with the proposed sample size.; The protocol conflates the 'true' 99th percentile with a single, albeit long, simulated or experimental realization. This is a single sample of the extreme value distribution. A proper validation requires a distribution of true 99th percentiles (e.g., from multiple long simulations under varying sea state parameters) to assess the coverage of the posterior predictive interval against the *distribution* of the true extreme, not just one point estimate.; The homotopy success criterion (Prediction 3) is defined as achieving a final RMSE 'at least 5% lower than the best of 50 multi-start runs'. This is a weak and potentially misleading metric. It does not guarantee that the solution is near the global minimum, only that it is better than a limited set of random starts. A stronger criterion would be to compare against a known global minimum (in Phase 1) or to use a convergence diagnostic like the one proposed by Gelman-Rubin.
- **Questions**: Given the statistical power analysis, how will you justify that a coverage probability of 0.94, estimated from 100 datasets, is significantly different from the null hypothesis of 0.90? What is the a priori power calculation for this specific proportion test?; In Phase 1, you generate 100 datasets by adding different noise realizations to the *same* 1-hour forcing signal. How does this account for the aleatory variability in the wave climate itself? Would the coverage hold if the true 99th percentile was derived from a different 10-hour simulation with a different random seed for the JONSWAP spectrum?; The protocol does not specify how the prior distributions for the geostatistical initialization are defined. If the prior is centered on the 'true' values (as suggested by 'uniform ±50% around true values'), this represents a form of 'informed' prior that may not be available in practice. How would the method's performance degrade with a more diffuse or misspecified prior?
- **Recommendation**: I recommend a weak accept with major revisions. The methodological framework is sound and the staged approach is excellent, but the statistical power for the primary endpoint is insufficient to support the strong claims. I strongly urge the authors to re-run the power analysis and increase the number of synthetic datasets to at least 500 (ideally 1000) to achieve a narrow enough confidence interval to test the 0.94 vs 0.90 hypothesis. The protocol should also be revised to explicitly define how the 'truth' is sampled to account for environmental stochasticity, and the homotopy success criterion should be made more stringent.

#### Domain Expert

- **Score**: 7.2/10 | **Verdict**: weak_accept | **Confidence**: 0.82
- **Strengths**: La formulation est remarquablement claire et opérationnalisée : les critères de succès (couverture de probabilité [0.94, 0.99], réduction de biais ≥ 25%) sont quantifiés et testables, ce qui est rare et précieux dans la littérature sur l'UQ. Cette pré-spécification force une rigueur épistémique qui manque souvent aux études de calibration.; Le choix de transférer un cadre hybride éprouvé en géophysique (où les données sont spatiales et les modèles fortement non-linéaires) vers l'ingénierie océanique (où les données sont temporelles et les modèles réduits) est intellectuellement stimulant et potentiellement fécond. L'identification précise des défis de ce transfert (autocorrélation temporelle, corrélation périodique) et la sélection de méthodes bootstrap dédiées (Variable Bandpass Periodic Block Bootstrap) montrent une compréhension fine des verrous méthodologiques.; La chaîne causale est explicite et les hypothèses sous-jacentes sont clairement énoncées. Cela permet d'identifier immédiatement les points de fragilité et de concevoir des expériences de validation ciblées, ce qui est la marque d'une bonne recherche.
- **Weaknesses**: Le principal défaut réside dans le saut épistémique entre les domaines : la justification de la transférabilité repose sur des similarités de surface (problèmes inverses non-linéaires) mais ignore les différences structurelles profondes. En géophysique ERT, l'erreur dominante est souvent l'erreur de modèle (paramétrisation), tandis que pour une tour flottante, l'erreur dominante est l'erreur d'entrée (forçage environnemental stochastique). Le bootstrap par blocs sur les mesures de tension de ligne capture la variabilité des sorties, mais ne capture pas l'incertitude sur le forçage lui-même, qui est la source majeure d'aléa. L'hypothèse 2 est donc trop forte : le bootstrap ne peut pas 'capturer correctement la variabilité aléatoire' si les données mesurées ne sont pas représentatives de l'ensemble des états de mer possibles.; La plausibilité physique du mécanisme de homotopie pour un oscillateur 2-DOF est contestable. L'homotopie est un outil puissant pour des espaces de paramètres de haute dimension et des fonctions objectifs complexes (comme en ERT). Pour un modèle réduit à 2-DOF, l'espace des paramètres est de faible dimension et la fonction objectif est probablement bien moins rugueuse que celle d'une inversion géophysique. L'argument 'éviter les minima locaux' est plus faible ici, et le bénéfice computationnel de l'homotopie par rapport à un simple recuit simulé ou un algorithme génétique n'est pas démontré. Le mécanisme est donc peut-être sur-ingénieré pour le problème proposé.; La base bibliographique, bien que pertinente, est trop 'géophysique-centrée'. Il manque des références clés sur la calibration et l'UQ dans le domaine spécifique de l'éolien offshore flottant (ex: travaux de Leimeister, de Kostas, ou les benchmarks de l'IEA Wind). La nouveauté est réelle, mais le positionnement par rapport à l'état de l'art de l'ingénierie océanique est insuffisant. On ne sait pas si le problème de calibration des modèles réduits de flotteurs est déjà traité par des méthodes d'UQ bayésiennes classiques (MCMC) ou des approches fréquentistes (profile likelihood), et en quoi cette approche hybride serait supérieure.
- **Questions**: Comment le bootstrap par blocs sur les séries temporelles de tension de ligne peut-il capturer l'incertitude épistémique sur le modèle réduit lui-même (par exemple, l'erreur de troncature due à la réduction à 2-DOF) ? Le bootstrap ne ré-échantillonne que les données, pas la structure du modèle. Ne faudrait-il pas un terme de 'model discrepancy' explicite, comme préconisé dans les approches bayésiennes modernes (Kennedy-O'Hagan) ?; Le choix d'un modèle 2-DOF est-il vraiment le banc d'essai le plus discriminant ? Un oscillateur 2-DOF a une dynamique bien plus simple qu'un flotteur réel. Ne serait-il pas plus probant de valider sur un modèle à 6-DOF (3 translations, 3 rotations) avec des couplages hydrodynamiques non-linéaires, afin de réellement tester la capacité de l'homotopie à naviguer dans un espace de paramètres plus complexe et de justifier son utilisation ?; La métrique de performance (couverture du 99ème percentile) est-elle suffisante ? Ne faudrait-il pas aussi évaluer la sharpness de la distribution prédictive postérieure (ex: score de continuité de classement, intervalle de crédibilité à 95%) pour s'assurer que l'intervalle n'est pas simplement très large pour atteindre la couverture cible ?
- **Recommendation**: L'hypothèse est prometteuse et bien formulée, mais elle nécessite des ajustements conceptuels avant de pouvoir être validée. Je recommande de (1) reformuler le rôle du bootstrap pour le limiter à la variabilité des données observées et d'introduire explicitement un terme d'erreur de modèle pour capturer l'incertitude épistémique structurelle, (2) justifier plus rigoureusement le choix de l'homotopie en comparant ses performances à des algorithmes d'optimisation globale plus simples sur le cas 2-DOF, et (3) élargir la revue de littérature pour inclure les travaux récents en UQ pour l'éolien offshore. Ces modifications renforceraient la rigueur théorique et la plausibilité pratique de l'approche.

#### Contrarian

- **Score**: 3.5/10 | **Verdict**: weak_reject | **Confidence**: 0.82
- **Strengths**: The explicit separation of aleatory (bootstrap) and epistemic (initial ensemble) uncertainty is methodologically principled and represents a genuine improvement over monolithic UQ approaches that conflate the two.; The pre-registration of specific quantitative thresholds (coverage ≥ 0.94, bias reduction ≥ 25%) with null hypotheses is commendable and rare in the reduced-order modeling literature, providing a clear falsifiability criterion.
- **Weaknesses**: FAIL REASON #1: The 2-DOF oscillator is a fundamentally inadequate testbed for the proposed UQ framework. The causal chain assumes that homotopy continuation is needed to escape local minima in a nonlinear objective function, but a 2-DOF mooring load model calibrated on synthetic data is almost certainly a unimodal or mildly multimodal problem. The method's key selling point (homotopy avoiding local minima) is never actually stress-tested, making the validation circular: the framework is tested on a problem where it is unnecessary, and the results cannot be extrapolated to the 15-DOF case where it might be needed.; FAIL REASON #2: The block bootstrap for aleatory uncertainty is fundamentally flawed for extreme load estimation. The 99th percentile of mooring loads is dominated by rare, extreme sea states and nonlinear mooring line snap events. Block bootstrap resampling of measured time series preserves the autocorrelation of the *observed* data but cannot generate new extremes beyond those present in the original record. The method will systematically underestimate the variance of the 99th percentile because it resamples from a finite record that may not contain sufficient extreme events. The coverage probability target of [0.94, 0.99] is therefore likely to be achieved only because the synthetic data generator (the 2-DOF model) is the same as the calibration model, creating an inescapable circularity.; FAIL REASON #3: The comparison of homotopy vs. multi-start gradient descent (Prediction 3) is a strawman. Multi-start gradient descent with 50 random starts is a weak baseline; any self-respecting optimizer (e.g., CMA-ES, Bayesian optimization) would likely outperform it. More critically, the success criterion ('final RMSE at least 5% lower') is meaningless if the objective function landscape is flat or if the RMSE differences are within the noise floor of the synthetic data. The prediction conflates optimization performance with UQ quality, and the 80% success rate threshold is achievable even if homotopy provides no real benefit, simply because multi-start is poorly tuned.
- **Questions**: How do you justify that the coverage probability of the 99th percentile is a meaningful metric when the block bootstrap cannot generate new extremes? Specifically, what is the expected coverage when the true 99th percentile lies above the maximum observed value in the bootstrap resamples?; What is the evidence that the 2-DOF oscillator exhibits the same objective function topology (e.g., multiple local minima, narrow valleys) as the 15-DOF floating wind model? If the topologies differ, what exactly does this validation demonstrate?; You state the geostatistical prior 'encompasses the true parameter values.' In a real experimental scenario, how do you guarantee this? If the prior is misspecified and the true parameters lie outside the ensemble, the homotopy solver will converge to the wrong posterior, and the coverage probability will be driven to zero. What is the sensitivity of the results to prior misspecification?
- **Recommendation**: The authors must demonstrate the framework on a problem where local minima are demonstrably present and where the bootstrap's inability to generate extremes is addressed (e.g., by using a parametric extreme value model for the tails). Specifically, I require: (1) a diagnostic showing the objective function landscape of the 2-DOF model (e.g., a 2D slice showing multiple minima) to justify the homotopy necessity; (2) a sensitivity analysis of the coverage probability when the geostatistical prior is deliberately shifted away from the true parameters; (3) a comparison against a more robust baseline (e.g., CMA-ES) for the homotopy prediction. Without these, the pre-registered thresholds are vacuous.

#### Industrialist

- **Score**: 6.0/10 | **Verdict**: weak_accept | **Confidence**: 0.65
- **Strengths**: The explicit separation of aleatory (bootstrap) and epistemic (geostatistical ensemble) uncertainty directly addresses a critical pain point for FOW developers: the current industry practice of using single best-fit calibration with safety factors leads to either over-engineered mooring systems (cost overruns of 10-15% on CAPEX) or under-predicted fatigue loads (warranty and insurance risks). The target of 25% bias reduction on the 99th percentile mooring load translates to a tangible reduction in mooring line cross-section or chain grade, which is a direct cost driver.; The homotopy continuation component offers a computational moat: standard multi-start gradient descent in commercial tools (e.g., OrcaFlex, DeepLines) often fails to find global minima in high-dimensional soil-mooring interaction models. A documented 80% success rate in escaping local minima would allow faster design iterations (reducing engineering hours per load case from days to hours), which is a compelling productivity sell to engineering consultancies like DNV or Aker Solutions.
- **Weaknesses**: The market is a niche within a niche: the primary buyers are the 5-7 major FOW developers (Equinor, Shell, RWE, Iberdrola) and their engineering contractors. The total addressable market for calibration software/consulting services is likely under €50M/year, not the €1B+ that would attract major software vendors. The 2-DOF oscillator validation, while rigorous, does not prove scalability to the 6-DOF + 1000+ node finite element models used in real FOW design, creating a significant 'valley of death' between Phase 2 and Phase 3.; The IP position is weak: block bootstrap and geostatistical simulation are established public-domain techniques (used in mining geostatistics since the 1980s). The novelty is purely in the combination and application domain. Without a strong patent on the specific homotopy-ensemble interaction or a proprietary pre-trained surrogate model, competitors (e.g., a well-funded startup like Akselos or a university spin-off) can replicate the approach within 12-18 months of seeing the published protocol.
- **Questions**: What is the specific commercial vehicle: do we sell a proprietary calibration module that integrates with existing FEA tools (e.g., an API for OrcaFlex), or do we offer it as a high-end consulting service (e.g., 'UQ-as-a-Service' per project)? The former has higher margins but requires a multi-year sales cycle; the latter has immediate revenue but is not scalable.; Can we secure a 'first-mover' lighthouse client (e.g., a joint development agreement with a top-3 FOW developer) before Phase 3, to co-fund the wave basin test and guarantee a reference case? Without this, the €50k-200k Phase 3 cost is a pure R&D expense with no guaranteed commercial return.; What is the defensibility horizon? If we publish the pre-registered protocol (as proposed), we give away the 'recipe'. Can we protect the specific implementation of the homotopy path (e.g., a novel continuation parameterization for mooring line catenary equations) as trade secret or patent, to prevent direct replication?
- **Recommendation**: Proceed with Phase 1 (€2k, 4 weeks) as a pure technical de-risking exercise, but do not commit to Phase 3 (wave basin) unless a co-funding agreement is signed with an industrial partner. The commercial strategy should pivot from selling a standalone software tool to embedding this UQ capability as a differentiated 'risk assessment service' within a larger engineering consultancy offering, targeting the specific pain point of mooring system certification. The realistic timeline to a revenue-generating product is 24-36 months, contingent on the 2-DOF results being reproducible on a higher-fidelity model (e.g., a 3-DOF lumped-mass mooring model) before the expensive wave basin test.

#### Funding Strategist

- **Score**: 6.5/10 | **Verdict**: weak_accept | **Confidence**: 0.7
- **Strengths**: Originalite methodologique : le transfert d'un cadre UQ hybride geophysique (bootstrap par blocs + initialisation geostatistique + homotopie) vers le calage de modeles d'ordre reduit pour l'eolien flottant est une vraie nouveaute, non couverte par les appels classiques sur les jumeaux numeriques.; Protocole pre-enregistre avec criteres GO/NO-GO explicites et quantitatifs (couverture [0.94, 0.99], reduction de biais >= 25%) : c'est un signal fort de rigueur scientifique qui rassure les evaluateurs sur la falsifiabilite de l'hypothese.; Trajectoire TRL claire et progressive : le passage du synthetique (TRL 2-3) au physique simplifie (TRL 3-4) puis au bassin de houle (TRL 4-5) est bien pense pour limiter les risques et justifier chaque phase.
- **Weaknesses**: Echelle et ambition limitees pour les programmes majeurs : un budget de 20-80k EUR et un objet d'etude 2-DOF sont trop restreints pour un ERC Starting Grant ou un appel Horizon Europe collaboratif classique, qui attendent des consortiums et des impacts systemiques plus larges.; Risque de pertinence industrielle : le passage du 2-DOF a une plateforme eolienne flottante reelle est un saut conceptuel important. Les evaluateurs pourraient douter que les conclusions sur un oscillateur simple se generalisent aux charges de ligne d'ancrage sous houle multidirectionnelle et vent turbulent.; Manque de partenaires identifies : l'hypothese ne mentionne aucun acteur industriel ou academique (ex: IFPEN, NREL, EDF, DTU, ou un bassin d'essais comme BGO-First ou MARIN). Pour un financement europeen, l'absence de consortium est un handicap majeur.
- **Questions**: Comment positionner cette preuve de concept comme une etape incontournable vers un jumeau numerique fiable pour la certification des systemes d'ancrage, afin de justifier une demande aupres d'un programme 'pathfinder' ou 'proof of concept' ?; Quel partenaire industriel ou institutionnel (ex: un bureau de certification comme DNV, ou un laboratoire de genie cotier) pourrait etre associe des la phase 2 pour credibiliser la trajectoire vers le TRL 5-6 et ouvrir les portes des appels collaboratifs ?; Le pre-enregistrement est-il suffisant pour compenser l'absence de donnees experimentales preliminaires, qui est souvent un critere bloquant pour les programmes de type ERC ou ANR Jeunes Chercheuses et Jeunes Chercheurs ?
- **Recommendation**: Cibler en priorite un financement de type 'Proof of Concept' ou 'Pathfinder' pour valider le mecanisme sur le 2-DOF, tout en construisant des maintenant un consortium avec un bassin d'essais (ex: Centrale Nantes / ECN) et un partenaire industriel pour la phase 3. Le dossier doit insister sur la rupture methodologique (separation aleatoire/epistemique) et sur l'impact potentiel en termes de cout de calcul pour le dimensionnement des ancrages.

### B. Semantic Scholar Search Queries

- [novelty] `uncertainty quantification geophysical inversion bootstrap` — Direct search for the core hypothesis of block-wise bootstrap UQ in geophysical inversion
- [novelty] `floating offshore wind mooring load uncertainty quantification` — Check if the proposed UQ transfer to floating wind mooring loads already exists
- [novelty] `homotopy continuation inverse problem structural identification` — Verify if homotopic continuation is already used for structural parameter estimation in ocean engineering
- [evidence] `block bootstrap resampling time series uncertainty` — Find evidence for the block bootstrap mechanism applied to time-series data in any domain
- [evidence] `geostatistical initial model ensemble inversion` — Evidence for geostatistical generation of initial models to explore posterior space in inverse problems
- [evidence] `epistemic aleatory uncertainty separation nonlinear inverse` — Evidence for formal separation of epistemic vs aleatory uncertainty in nonlinear inverse problems
- [evidence] `mooring stiffness damping parameter identification OC6` — Evidence for specific parameter uncertainty in floating wind mooring systems from OC6 benchmarks
- [cross_domain] `geophysical inversion calibration structural dynamics` — Search for prior transfer of geophysical inversion methods to structural dynamics calibration
- [cross_domain] `bootstrap resampling ocean engineering load prediction` — Check if bootstrap resampling has been applied to ocean engineering load predictions
- [cross_domain] `geostatistical co-kriging structural parameter estimation` — Look for geostatistical methods (co-kriging) applied to structural parameter estimation outside geophysics
- [cross_domain] `multiple initial models avoid local minima inverse` — Find interdisciplinary precedents for using multiple initial models to avoid local minima in inverse problems

---

*Generated by SPORE (Systeme de Production d'Opportunites de Recherche par Exploration) on 2026-08-09*