PhD position in wind farm fatigue load surrogating through graph-based machine learning

The Faculty of Engineering, Department Toegepaste mechanica, is looking for a PhD-student with a doctoral grant. 

More concretely your work package, for the preparation of a doctorate, contains: 

We are looking for a highly motivated and skilled PhD researcher to work on structural surrogates of offshore wind foundations through graph-based machine learning.

Our goal is to perform full-structure load predictions for wind turbines, specifically the foundations, with the ultimate objective of including structural health information in windfarm asset management to optimise structural lifetime consumption while guaranteeing optimal power production.

You will work on the cutting edge of both wind energy and machine learning, two of the fastest growing scientific disciplines, to develop graph-based surrogates of offshore wind turbines’ foundations. As newer offshore wind farms are coming online with tighter fatigue designs and more aggressive control strategies, your research will seek to assist decision-making (e.g. during operation, but also during design) to prevent over-consumption of fatigue life while balancing optimal production. The monitoring of fatigue loads along the entire substructure is a highly complex task, as sparse instrumentation does not guarantee direct sensing at all fatigue-critical locations of the substructure’s primary steel. Analytical solutions – so-called virtual sensing – have seen considerable developments over the years, but some fundamental assumptions to turn the problem into a tractable one haven’t changed: time-invariance is still assumed. Naturally, as turbines are dynamic systems evolving over time (due to operation), this doesn’t hold. Therefore, there is much to be gained by surrogating the full structure flexibly: a generalist surrogate capable of estimating and extrapolating loads along the substructure, for any type of turbine, given specific geometric, inflow and sea-state information. Furthermore, such a machine learning surrogate can speed-up both design and operational employment.

This doctoral research will thus leverage the power of graph neural networks – a novel ML architecture, capable of learning fundamental physical behaviour by modelling systems as graphs (i.e. relationally interdependent systems) and encoding nonlinearities in these. The group has plentiful in-house simulation capabilities of numerical models and access to extensive real-world monitoring data. The focus will be therefore on the development and critical assessment of the surrogate models, especially their validation, how effectively they generalize and extrapolate knowledge, and how might they be improved through transfer learning. Fundamental proof-of-concept studies using reduced order models (ROM) and even including physics-informed graphs are to be equated in an initial stage.

Your research will be supervised by Prof. Christof Devriendt, head of the research group on structural integrity monitoring for offshore structures and Dr. Francisco de Nolasco, postdoctoral researcher working on ML algorithms for lifetime estimation (GNNs). We are part of the Acoustics & Vibration Research Group itself part of the Department of Mechanical Engineering, Faculty of Engineering of the Vrije Universiteit Brussel (VUB). Our research group is a founding member of the Offshore Wind Infrastructure Application Lab (OWI-Lab) and, as part of it and our projects, you’ll be embedded in an international research setting, with access and the possibility to collaborate with experts all around the world. This research fits within the European project WILLOW, focused on developing data-driven smart (fatigue- aware) curtailment tools and the national funded projects Smartlife and Supersized, leveraging model and data-driven digital twins for smart asset management and lifetime optimization of offshore windfarms.

For this function, our Brussels Humanities, Sciences & Engineering Campus (Elsene) will serve as your home base.