Abstract: Accurate prediction of the permeability coefficient is crucial for evaluating the compaction quality of earthworks. However, during the compaction process, on-site testing is often time-consuming and expensive, leading to fewer samples, which affects prediction accuracy. Moreover, most current predictive models have limited capabilities and tend to be black-box models with poor explainability. To overcome these issues, in this study, we proposed a new method to predict the permeability coefficient of earth-rock material based on an improved generative adversarial network (GAN) and explainable osprey optimization algorithm–Huber loss–light gradient boosting machine (OOA–HL–LightGBM). Firstly, by introducing the Wasserstein distance as the loss function into the conditional generative adversarial network (CGAN), the Wasserstein conditional generative adversarial network (WCGAN) was proposed to generate high-quality data, addressing the issue of insufficient information caused by small samples. Furthermore, by incorporating material and compaction parameters as inputs, a high-accuracy permeability coefficient prediction model was developed using LightGBM with the Huber loss function and the OOA. Finally, the Shapley additive explanation (SHAP) method was introduced into OOA–HL–LightGBM to analyze the specific roles of different features within the dataset to enhance the credibility of the prediction results. The proposed method was applied to a large-scale high-core rockfill dam in southwestern China to thoroughly verify its effectiveness and superiority.

Key words: Permeability coefficient prediction; Light gradient boosting machine (LightGBM); Wasserstein conditional generative adversarial network (WCGAN); Shapley additive explanation (SHAP)

Please provide your name, email address and a comment

Your Name: Affili: E-Mail Address:

Comment (Please comment in English):


Verification Code(click to change a pic):