Deep learning models for efficient geotechnical predictions: reducing training effort and data requirements with transfer learning

Accurate predictions of geotechnical failure loads, such as bearing capacity and slope stability, are important for safe infrastructure design and risk management. Recently, the use of ready-to-use deep-learning models as surrogate models to improve computational efficiency of risk analysis considering spatial variability has been proposed. These deep-learning models were trained with big datasets that covers all possible material properties and boundary conditions for a particular kind of problem, so it is ready to make predictions without further training. However, training such a large model requires a big dataset and substantial computational efforts. This study introduces a novel framework by employing existing deep neural networks and transfer learning techniques to effectively address these issues. Specifically, the pre-trained MobileNetV2 for image classification is used as a base. It is found that parametric ReLU (PReLU) and locally connected (LC) layers are important for our tasks. The PReLU activation can mitigate neuron deactivation by allowing the model to learn negative input slopes, while LC layers enabled the extraction of localized features – critical for accurately representing spatial variability in soil properties. Compared to a handcrafted locally connected network (MAPE ≈ 2%), the proposed deep neural network achieves similar predictive accuracy (MAPE ≈ 3%) but reduced training times (only 10% time required on the same computer). Notably, training with only 50 to 60% of the data maintained stable performance, and even with as little as 8.5% of the original dataset, satisfactory accuracy was achieved. Furthermore, this transfer learning approach generalized seamlessly to various problems without significant modifications–both bearing capacity and slope stability problems are tested. The results highlight the accuracy, efficiency, and generalization of the proposed framework, indicating the potential to simplify geotechnical engineering analyses and accelerate decision-making in practice.