🤖 AI

FLOATBench: A Dataset and Benchmark for Floating Offshore Wind Turbine Tower Fatigue

This paper introduces FLOATBench, a public dataset and benchmark comprising over 580,000 fatigue-damage labels from high-fidelity simulations of 22 MW floating offshore wind turbines, designed to standardize the evaluation of surrogate models through a novel regime-aware protocol that addresses the limitations of traditional random-split validation in deep-water wind energy applications.

Original authors: João Alves Ribeiro, Bruno Alves Ribeiro, Francisco Pimenta, Sérgio M. O. Tavares, Faez Ahmed

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: João Alves Ribeiro, Bruno Alves Ribeiro, Francisco Pimenta, Sérgio M. O. Tavares, Faez Ahmed

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are an engineer trying to build a giant, floating wind turbine that can survive in the deep ocean. These turbines are massive (22 megawatts, which is huge), and they are constantly being battered by wind and waves. Over 25 years, this constant shaking causes "fatigue," like metal getting tired and eventually cracking. If you get the math wrong, the tower could break, costing millions and causing safety issues.

To design these towers, engineers use super-complex computer simulations (called OpenFAST). But running one simulation takes about an hour of computer time. To design a tower properly, you need to run thousands of simulations for different wind speeds, wave heights, and wave periods. It's like trying to taste every single possible flavor of ice cream to find the perfect one, but making a single scoop takes an hour.

The Problem: No Common Recipe
The paper explains that researchers were all trying to build "shortcuts" (called surrogate models) to predict fatigue without running the slow simulations. However, everyone was using their own private set of data, their own testing methods, and their own rules. It was like everyone playing a different game of chess with different rules; you couldn't tell who was actually the best player.

The Solution: FLOATBench
The authors created FLOATBench, a public "standardized test" for these shortcuts. Think of it as a giant, shared exam for computer models.

Here is what makes it special, using simple analogies:

1. The "Exam" is Huge and Realistic

Instead of a small quiz, they created a massive dataset with 582,000 test questions.

The Questions: Each question asks, "If the wind blows at this speed and the waves are this high, how much damage will this specific part of the tower suffer?"
The Source: They didn't guess the answers. They ran 19,404 real, high-fidelity simulations (the "slow" hour-long ones) to generate the correct answers (the "labels").
The Variety: They tested three different tower designs (a baseline one and two optimized ones) and looked at 30 different sections of the tower from the bottom to the top.

2. The "Tricky" Part: The Edge Cases

Most standard tests just shuffle the questions randomly. If you study for a test by memorizing the middle of the book, you might pass a random quiz. But in the real world, storms happen at the edges of what we expect.

FLOATBench divides the test into three zones:

In-Train (The Classroom): These are conditions the model has seen before.
Interpolation (The Hallway): These are conditions between what the model has seen.
Extrapolation (The Wild West): These are extreme conditions the model has never seen before (e.g., a massive storm).

The Big Discovery:
The paper found that if you just use a random test (like a standard school exam), the "best" model looks like a genius. But when you test them on the "Wild West" (the extreme storms), the rankings flip!

The Metaphor: Imagine a student who memorizes the textbook perfectly (Rank #1 in class). But when you take them out into a blizzard (Extrapolation), they freeze and fail. Meanwhile, a student who understands the principles of weather (a Neural Network) might be average in class but survives the blizzard.
The Result: The "Global Winner" (the best model overall) often fails miserably at the most dangerous, extreme conditions. FLOATBench exposes this so engineers don't pick a model that looks good on paper but fails in a real storm.

3. The "Transfer" Test

The benchmark also tests if a model trained on one tower design can learn to predict damage for a different tower design.

The Metaphor: It's like teaching a mechanic to fix a specific type of car, and then asking them to fix a completely different car model without any new training.
The Result: The paper found that if you train the model on a "standard" tower, it can predict damage for the new, optimized towers. But if you train it only on the optimized towers, it fails completely when asked to predict damage for the standard one. It's like a mechanic who only knows how to fix sports cars and has no idea how to fix a truck.

Summary of What They Did

Built a Library: They created a massive, public library of 582,000 fatigue answers derived from real physics simulations.
Created a Fair Test: They designed a testing protocol that forces models to prove they can handle extreme, unseen conditions (the "Extrapolation" zones), not just the easy ones.
Ranked the Models: They tested 8 different types of AI models (like XGBoost, Neural Networks, etc.) and found that the "best" model depends entirely on where you are testing it. A model that wins the general test might lose the "survive the storm" test.

Why It Matters
This isn't just about wind turbines; it's about creating a fair way to test AI in engineering. It stops researchers from claiming their model is "the best" just because it's good at guessing the middle of the data. It forces them to prove their model is safe at the edges, where real-world disasters happen.

The paper concludes that FLOATBench is the first time the floating wind industry has had a shared, fair benchmark to compare these AI shortcuts, ensuring that the models we trust to design our future energy infrastructure are actually reliable.

Technical Summary of FLOATBench

Problem Statement
The deployment of floating offshore wind turbines (FOWTs) in deep waters is essential for accessing the majority of global offshore wind resources. As turbine designs scale toward 22 MW, tower fatigue becomes a critical design bottleneck due to the complex coupling of aero-hydro-servo-elastic loads. Accurate fatigue-damage prediction is vital for certification and optimization, yet the field lacks a shared surrogate benchmark. Existing studies utilize disparate simulation sets, data splits, and metrics, making it impossible to compare competing methods. Furthermore, standard random train/test splits fail to evaluate model reliability in extrapolation regimes—conditions outside the training envelope—which are operationally critical for engineering surrogates deployed over physical wind/wave operating envelopes.

Methodology
The authors present FLOATBench, a public tabular benchmark and evaluation protocol designed to address these gaps.

Dataset Construction: The dataset comprises 582,120 per-section fatigue-damage labels derived from 19,404 high-fidelity OpenFAST simulations. These simulations cover three distinct 22 MW FOWT tower geometries (a baseline REF and two fatigue-aware re-designs, OPT1 and OPT2) mounted on a semi-submersible platform. Each tower is simulated across a grid of 1,078 aligned wind/wave operating points (22 mean wind speeds × 7 significant wave heights × 7 peak wave periods) with six turbulence seeds per point. The output is aggregated at 30 discrete cross-sections per tower, converting bending moment time-series into cumulative fatigue damage ( $D$ ) via rainflow counting and Miner's rule.
Regime-Aware Partitioning: Unlike standard benchmarks that use random splits, FLOATBench employs a regime-aware partition based on an $\alpha$ -shape (concave hull) of the training data in the wind and wave subspaces. Test points are stratified into three regimes: In-Train (IT), Interpolation (IP), and Extrapolation (EX). This ensures the test set explicitly covers the boundaries and corners of the operating envelope, creating a structured covariate shift.
Evaluation Protocol: The benchmark defines three protocol levels:
1. E1 (Random Split): A baseline random partition to establish global performance.
2. E2 (Within-Tower Regime-Aware): Evaluation on the regime-stratified test set, reporting performance per regime (e.g., the worst-case EX_EX regime) and per tower section (base vs. top).
3. E3 (Cross-Tower Transfer): A leave-one-tower-out scheme where models are trained on two geometries and evaluated on the third to test transferability across design families.
Surrogate Models: The benchmark evaluates a suite of tabular surrogates (XGBoost, LightGBM, CatBoost, Neural Networks, TabM, etc.) trained via AutoGluon-Tabular, using environmental features and section geometry as inputs to predict damage (in DEL format).

Key Results
The evaluation reveals significant discrepancies between global accuracy and boundary reliability:

Rank Inversions: Models that rank highest under random splits (E1) often fail in extrapolation regimes. For instance, the global rank-1 model ( $WeightedEnsemble\_L2$ ) drops to rank 23/11/11 in the worst-case EX_EX regime across the three towers, while neural network variants (e.g., $NeuralNetFastAI\_r102\_BAG\_L1$ ) emerge as top performers in extrapolation despite lower global ranks.
Regime Sensitivity: The wind axis drives the majority of extrapolation error (6–20× increase from IT to EX), whereas the wave axis remains relatively flat. Neural network families demonstrate superior robustness in wind extrapolation compared to tree-based models like TabM.
Section and Geometry Dependence: The "best" model varies by tower section; a model winning at the tower base may lose at the top. Similarly, cross-tower transfer (E3) shows high sensitivity to geometry: training on designs including the baseline REF allows for reasonable transfer to optimized designs, but training without REF results in a 4–6× increase in error when predicting the baseline geometry.

Significance and Claims
The authors claim that FLOATBench is the first public scientific-ML benchmark for FOWT tower fatigue at the 22 MW scale, offering a multi-geometry, per-section dataset with a regime-aware evaluation protocol. The primary contribution is the demonstration that random-split leaderboards systematically miss critical failures in extrapolation and boundary regimes. By exposing rank shifts between global and subgroup-specific performance, the benchmark provides a more rigorous standard for selecting surrogates that ensure reliability in real-world deployment conditions where operating points frequently drift beyond the training cloud. The authors position the dataset and evaluation harness as a foundation for adjudicating competing tabular surrogates and suggest the regime-aware protocol generalizes to other engineering surrogate tasks defined over physical operating envelopes.

1. The "Exam" is Huge and Realistic

2. The "Tricky" Part: The Edge Cases

3. The "Transfer" Test

Summary of What They Did

Technical Summary of FLOATBench

More like this