Performance of traditional and machine learning-based transformation models for undrained shear strength

In geotechnical engineering, transformation models are often used as first estimates of parameters and to verify the order of magnitude of field and laboratory tests, which reliability might be constrained by many uncertainties. The undrained shear strength has been for long of particular interest for such models. The traditional transformation models for undrained shear strength are often rather simple. Still, the geotechnical community does not seem to have agreed upon which models to use. In particular, the question of including index properties to the models seems to be open. In the paper, the performance of traditional transformation models is compared to that of machine learning (ML)-based models. In addition, the influence of data coherence is studied by using two datasets of different quality. The ML-based transformation models proved to perform better than traditional ones for both datasets. Clearly, most dominant variables in the transformation model are the preconsolidation pressure and the effective vertical stress. Although including additional variable often may well improve the performance of the training set, the prediction of the testing sets generally tends to worsen, indicating overtraining. The risks for overtraining increase with incoherent data.