B1 - Integrated design process for offshore structures (endet 2024)

In project B01, a data-driven framework was developed to support the conceptual design of offshore jacket substructures. Given the complexity and cost-sensitivity of offshore wind projects, the early conceptual phase poses significant challenges. Traditional methods rely heavily on engineering intuition and experience, often resulting in designs prone to bias and requiring costly modifications later. To address these issues, our research integrated state-of-the-art machine learning (ML) and data augmentation techniques with systematically collected global data, providing a new pathway for automated, data-informed conceptual design.

A key achievement was the construction of a high-quality global dataset of jacket designs, enriched with synthetic structures generated through advanced computational methods. Real-world data gaps were addressed by leveraging online resources, engineering reports, and estimation techniques (e.g., using MICE for missing values), while the addition of synthetic structures was accomplished via multi-objective genetic algorithms (NSGAII) and deep generative models (e.g., Mixed Deep Gaussian Mixture Models, MDGMM). These approaches not only broadened the design space beyond what real data alone could offer but also enhanced the robustness and predictive capability of downstream ML models. By incorporating physically informed objectives, such as load-bearing capacity under wind and wave loads, as well as cost estimations based on total mass, into the generative processes, the synthetic designs retained engineering plausibility and relevance. Additionally, a user-friendly framework was developed that allows engineers to influence the synthetic design generation process by customizing the design space, optimization objectives, hyperparameters, and the final population of designs.

The ML-based predictive framework we developed uses a variety of algorithmic approaches, including random forests (RF), extreme gradient boosting (XGBoost), and multi-layer perceptron (MLP), to estimate crucial jacket structural features from given boundary conditions. Additionally, interpretability techniques, such as feature importance analyses, were employed to provide transparency into the decision-making process, guiding engineers and stakeholders in understanding why certain designs were recommended. These predictive models were integrated into a data-driven conceptual design workflow, enabling quick, reliable, and justifiable early-stage decisions that reflect both existing patterns (captured in real data) and innovative possibilities (explored by synthetic data and generative optimization methods).

The result is a methodology that not only accelerates the conceptual design of offshore jacket substructures and improves its efficiency but also supports strategic decision-making with quantifiable evidence. By incorporating simulation-driven objectives, physical constraints, and generative modeling, the approach delivers well-founded design recommendations that align with economic, structural, and environmental considerations. This lays a more reliable and cost-effective foundation for the early phases of offshore wind farm development, advancing the state-of-the-art in offshore structural engineering.

At the end of the project, several collaborations with other subprojects (SP) in the CRC1463 enriched the research outcomes. With SP A03, code and knowledge exchange enabled the simulation of load-bearing capacities under wave conditions, yielding valuable insights into structural performance. Through cooperation with SP A06, parameters from real and synthetic jacket substructures were provided to facilitate vessel-substructure coupled motion simulations, building the basis for future surrogate modeling of installability. The partnership with SP B02 enhanced our dataset by incorporating key load cases derived from specialized simulations of reference wind turbines. Additionally, our generative design optimization process would be integrated into the digital twin methodologies by SP Z01, ensuring that design recommendations integrate into broader digital lifecycle management frameworks.

During the process of the project, the members of B01 took part in various activities related to the CRC, such as working as part of internal working groups and organizing events and workshops. Specifically, Prof. Dr. Eirini Ntoutsi, in cooperation with Manolis Panagiotou, organized a Machine Learning lecture, as well as, a hands-on workshop on Python and Machine Learning, for the members of the CRC. Additionally, Manolis, as a member of the working group “PhD Seminar”, aided in the organization of the “CRC networking workshop” in Bremen in 2022. Furthermore, the working group successfully secured and organized the 19^th EAWE PhD Seminar in Hanover in 2023. As part of the working group, Manolis was additionally responsible for maintaining the website of the event. At the same time, Han Qian, as a member of the working group “Research Data Management”, developed “Guideline for Handling Research Data: CRC 1463 Integrated Design and Operation Methodology for Offshore Megastructures”, which were approved by the CRC and published.

The work under the subproject B01 lead to publications in various workshops, conferences, and journals. The full track of publications is listed below.

.