CPT Prediction from Seismic Data using Machine Learning – Comparing Joint versus Separate Prediction Approaches

This study evaluates machine learning-based prediction of Cone Penetration Test (CPT) parameters (qc, fs, u2) from seismic data at an offshore wind farm site in the German North Sea. Two approaches are compared: separate prediction of each CPT parameter using individual neural networks, and joint prediction using a single multi-output neural network. The joint prediction approach produces more geologically realistic combinations of parameters, potentially supporting more reliable soil classification and geotechnical interpretation. In contrast, the separate approach can result in inconsistent parameter relationships in certain zones. Additionally, the joint model significantly enhances efficiency by eliminating the need for multiple unit-specific models and post-processing steps. This makes it particularly well-suited for large-scale applications involving 2D or 3D seismic data. Overall, joint CPT prediction offers a more robust and scalable solution for subsurface characterization in offshore wind farm development.