Practice of Artificial Intelligence in Geotechnical Engineering

Deep learning (DL) methods have emerged as one of the most popular and effective artificial intelligence algorithms for solving a range of problems in various industries. In geotechnical engineering, DL methods are becoming increasingly important for addressing complex issues. However, conventional DL methods have limitations in representing multi-level and sequential data, which are typical in geotechnical engineering. These limitations can lead to inaccurate and inefficient models. To overcome these limitations, we propose a novel REMSE loss function and an adaptive activation function tanhLU. These techniques are designed to address the differential performance on multi-level data and vanishing gradient problems. By understanding the underlying mechanisms of these problems, we were able to develop solutions that can significantly improve the performance of DL models in geotechnical engineering. During the talk, I will present several examples of how these improved DL methods can be used to model the monotonic stress-strain relationships of soils and sequential behaviours of shield tunnelling. These examples demonstrate the effectiveness of our proposed techniques and their potential for a range of applications in geotechnical engineering. Overall, the development of improved DL methods is crucial for advancing our understanding of complex phenomena in geotechnical engineering and solving important practical problems.