Session: 08-04-02 Machine Learning and Optimization for Marine Design and Operation

Paper Number: 126721

126721 - Highly Efficient Probabilistic Analyses of Floating Offshore Marine Drilling Operations Using Active Learning Methods 

Garlid et al. (2022) previously showed that it is feasible to determine marine riser disconnect criteria for a mobile offshore drilling unit (MODU) using the structural reliability analysis (SRA) method. The method was presented through a case study of a notional MODU moored, with static thruster assist, in shallow water and subjected to harsh environments. The case study was performed with a series of simplifications that might introduce unwanted conservatism for the end user, such as co-linear environmental forces and static thruster settings. Further, only ten environmental load cases were used to populate the response surfaces, leading to a relatively coarse response surface grid. Still, a significant amount of time domain analyses was needed to populate the response surfaces that describe the relevant short-term response statistics. Although more computationally expensive, the SRA method offers several advantages over the conventional frequency-domain method currently employed in the industry. These advantages include the possibility to consider non-linear effects more accurately when considering extreme responses and the possibility to consider the coupling effects between the different components, such as the MODU and marine riser. Therefore, the SRA method is more accurate, allowing operators to run more efficient marine operations and possibly extending operational windows. In this paper, the authors will develop the SRA method further by utilising active learning when performing the response surface generation to significantly reduce computational efforts. Active learning is a type of machine learning that performs the data sampling in an active manner. The algorithm will sample the data points in iterations. After each iteration, the algorithm will choose the new sampling points for the next iteration with the aim of minimising the error function. By doing this, the algorithm only trains using the most relevant data points and avoids using data points that do not have significant contributions to the response surface; in this paper, the Kriging model will be used to model the response surface. When employing active learning, the computational efforts required by the SRA method are expected to decrease by as much as over 95%. This would make the SRA method closer in performance in terms of computational efforts against the conventional frequency domain but with the added advantages of a high-fidelity method that can consider full-coupling and non-linear effects.

Presenting Author: Stian Garlid University of Stavanger

Presenting Author Biography: Mr. Stian Garlid has a Master's degree in Marine and Offshore Technology from The Norwegian University of Science and Technology (NTNU), and has worked 13 years in the industry for Equinor. Stian currently works in the engineering department within the drilling and well division where he is working on a Ph.D. project related to operational criteria for mobile offshore drilling units.

Authors:

Stian Garlid University of Stavanger
Chao Ren University of Stavanger
Yihan Xing University of Stavanger