This paper, presented at the OMAE 2023 conference, explores how machine learning can move physical model testing from "intuition-based" to "data-driven." By using Gaussian Processes and Bayesian optimization, the researchers developed a method to design more efficient experiments for studying the complex nonlinear forces that cause "ringing" in offshore wind turbine foundations.
TL;DR
The researchers developed a Bayesian sampling strategy to optimize the design of wave tank experiments for offshore wind turbines. By integrating previous experimental results and real-world metocean data into a Gaussian Process model, they created a "smart" experimental design that captures complex nonlinear wave loads more accurately and efficiently than traditional manual methods.
What Problem Does This Paper Address?
In offshore engineering, physical model tests in wave tanks are the gold standard for understanding how structures like wind turbine monopiles will behave in the ocean. However, these tests are expensive and time-consuming. Engineers usually rely on their own experience or a "grid search" (testing a uniform range of conditions) to decide which wave heights, periods, and cylinder sizes to test.
The specific challenge here is nonlinear wave loading, particularly "ringing"—high-frequency vibrations that occur when steep waves hit a structure. Ringing can lead to structural resonance and fatigue, but capturing the physics behind it requires very specific, often extreme, test conditions. The authors wanted to find a way to pick the most "informative" test cases—those that provide the most physical insight with the fewest possible runs.
Methods
The authors employed a Gaussian Process (GP) surrogate model combined with Bayesian Interference.
- Surrogate Modeling: They built a GP model to interpolate between existing data points. Unlike a simple curve fit, a GP provides both a prediction and a measure of uncertainty (variance) for any given point in the parameter space ($kA$ for wave steepness, $kd$ for water depth, and $kR$ for cylinder size).
- Modified Acquisition Function: This is the core innovation. Usually, Bayesian optimization looks for points with the highest uncertainty. However, in the ocean, some "uncertain" points are physically impossible (e.g., waves so steep they break before hitting the structure). The authors modified the acquisition function to weight uncertainty against field data likelihood. This ensures the model suggests test cases that are both informative and representative of real-world sea states.
- Harmonic Extraction: They used a four-phase harmonic extraction method to separate the tricky nonlinear force components from the standard linear ones, allowing the ML model to focus specifically on the higher-order physics.
Datasets
- Prior Experimental Data: Datasets from three previous experimental campaigns were used to "prime" the GP model.
- Field Data: 20 years (2004–2024) of ERA5 metocean reanalysis data from 36 UK offshore wind farm locations. This provided the "real-world" context for wave heights and periods.
- New Experimental Data: The study concluded with a new set of experiments conducted at the Kelvin Hydrodynamics Laboratory (University of Strathclyde) using a 76m wave flume to validate the ML-designed test matrix.
Key Insights and Ideas
- Interpretable ML Decisions: The ML model made decisions that aligned with physical intuition but were more precise. For example, it assigned low importance to very low-steepness waves (where physics is linear and well-understood) and very high-steepness waves (which would break before reaching the cylinder).
- Optimizing Hardware: The model didn't just pick wave conditions; it helped pick the cylinder size. The acquisition function suggested an optimal radius of $R \approx 0.3m$. However, the researchers had to adjust this to $R=0.2m$ to avoid side-wall reflections in their specific tank—highlighting the need for "human-in-the-loop" constraints.
- Asymptotic Behavior: The model identified that the effect of water depth ($kd$) on nonlinear loading has an asymptotic behavior, meaning the importance of testing deeper water shifts significantly as you move from shallow to intermediate depths.
Key Results
- Higher Accuracy: The Bayesian-designed experiment outperformed the traditional grid search in capturing higher-order force coefficients.
- Efficiency: As shown in the paper's error evolution plots (Figure 9), the relative prediction error dropped much faster when using Bayesian sampling compared to a standard grid search.
- Realistic Testing: By incorporating ERA5 data, the experiment avoided "wasting" tank time on wave conditions that are mathematically possible but never actually occur in the North Sea.
Limitations and Future Work
- Physical Constraints: The current GP model doesn't "know" how hard it is to change a cylinder. In a real lab, changing a wave setting takes seconds, but changing a physical cylinder takes a whole day. The authors suggest future acquisition functions should include a "penalty" for hardware changes to optimize for actual lab time.
- Unidirectional Focus: The study used unidirectional waves. In the real ocean, waves are "spread" (coming from multiple directions), which adds another layer of complexity for future ML models to tackle.
- Geographic Bias: Because the field data was specific to the UK, the resulting experimental design is optimized for North Sea conditions. Different regions might require different "informed" designs.
Summary updated