Surrogate modelling of large-scale wave energy converter arrays using geometric feature engineering

Summary

The article demonstrates that enriching raw device coordinates with geometric features encoding spatial relationships, such as distance and orientation between wave energy converters (WECs), consistently improves the accuracy of a surrogate model predicting two-level absorbed power in arrays of 49 and 100 devices under Perth and Sydney wave climates.

Why it matters

A single hydraulic simulation for a 100-device WEC layout takes approximately one hour, making large-scale design exploration prohibitively expensive. A more accurate and data-efficient surrogate model directly accelerates early-stage layout optimization, addressing one of the critical bottlenecks in the large-scale commercialization of wave energy.

Highlights

• Geometric feature enrichment outperforms coordinate-only inputs in both prediction accuracy and explained variance across both array sizes.
• Models retain significant predictive performance even when trained on substantially reduced datasets, indicating improved data efficiency.
• The approach is validated against frequency-domain potential flow simulations in two distinct offshore wave climates in Perth and Sydney.

Critical take

Adding hand-crafted relational features to a table-based machine learning pipeline is a well-known technique in spatial machine learning. The baseline model using only coordinates may be considered a weak target. A more meaningful comparison would involve a graph neural network or attention-based model that learns spatial structure end-to-end. Furthermore, the article does not specify the surrogate model's architecture or report uncertainty quantification. Consequently, it remains unclear how robust the accuracy gains are under unseen wave directions or array topologies.