Green AI for Energy-Efficient Ground Investigation: A Greedy Algorithm-Optimized AI Model for Subsurface Data Prediction

Current ground investigation practice in geotechnical engineering is highly energy- and time-intensive, which is environmentally unfriendly. Several weeks are usually required to perform the investigation, and along this process, acquiring just one parameter of a soil sample may consume more than 100 Wh. Addressing this challenge, this study aimed at reducing energy consumption of ground investigation following Green AI concept, and an AI model for subsurface data prediction, which was optimized with greedy algorithm, was therefore developed. This model inputs soil parameters obtained by low-energy methods and outputs soil parameters (i.e., shear strength parameters) obtained by high-energy methods in traditional ground investigation practice. The model was established using ensemble learning techniques and a large dataset with fifteen types of parameters, which was obtained from 924 samples via a series of laboratory tests in a traditional ground investigation. Meanwhile, considering the variability on ensemble model architecture, greedy algorithm was adopted to optimize the architecture from a pool of nine base learners and six ensemble strategies. Thereby, the best-performing AI model was eventually identified, and it achieved a R2 value of 0.881 over the testing dataset, which accounts for 30% of the overall parameter dataset. With this model, the energy consumption needed to acquire the subsurface data was significantly reduced by 71%. The findings of this study pave the way for a lower-carbon but more intelligent ground investigation.