Data Science Applications in Geo-Intelligence-Paper

Paper

Pdf Summary

The paper examines how machine-learning methods can support offshore geotechnical engineering, using two practical examples: predicting pile-driving blowcounts and classifying soil types from cone penetration test (CPT) data.<br /><br />A prediction competition used installation records from 114 North Sea piles, with CPT measurements, pile dimensions and hammer data as inputs. The strongest submissions achieved useful accuracy, and several different methods performed well. Engineering-informed feature design was especially important: transforming CPT measurements into cumulative shaft resistance improved predictions, while a simple linear model using engineered features performed competitively. More complex neural networks generally performed poorly on this relatively small dataset.<br /><br />For soil classification, models were trained on Borssele CPT data to assign samples to nine soil classes. Linear classifiers struggled with classes requiring curved decision boundaries. Decision trees captured more complex patterns, but deeper trees overfit the training data and generalized less well. Feature selection and cross-validation helped improve and assess performance.<br /><br />The paper emphasizes that a model’s successful training does not guarantee reliable predictions. Engineers should evaluate accuracy on held-out data, use cross-validation appropriately—particularly for depth-dependent data—and check for underfitting and overfitting. Interpretability also matters: transparent models can be compared with established engineering methods, while black-box models require additional tools to assess their behavior.<br /><br />Wider adoption depends on high-quality data collection, improved data storage and exchange, access to shared datasets, and stronger data-science skills among geotechnical engineers. Machine learning can complement engineering knowledge and established methods, help recalibrate them as new data become available, and make better use of project experience. It is not a substitute for understanding soil behavior or engineering judgment.

Keywords

machine learning

offshore geotechnical engineering

pile-driving blowcounts

cone penetration test

CPT data

soil classification

feature engineering

cross-validation

overfitting

interpretability