DataSheet1_Machine-learning models to predict P- and S-wave velocity profiles for Japan as an example.PDF

Wave velocity profiles are significant for various fields, including rock engineering, petroleum engineering, and earthquake engineering. However, direct measurements of wave velocities are often constrained by time, cost, and site conditions. If wave velocity measurements are unavailable, they need to be estimated based on other known proxies. This paper proposes machine learning (ML) approaches to predict the compression and shear wave velocities (VP and VS, respectively) in Japan. We utilize borehole databases from two seismograph networks of Japan: Kyoshin Network (K-NET) and Kiban Kyoshin Network (KiK-net). We consider various factors such as depth, N-value, density, slope angle, elevation, geology, soil/rock type, and site coordinates. We use three ML techniques: Gradient Boosting (GB), Random Forest (RF), and Artificial Neural Network (ANN) to develop predictive models for both VP and VS and evaluate the performances of the models based on root mean squared errors and the five-fold cross-validation method. The GB-based model provides the best estimation of VP and VS for both seismograph networks. Among the considered factors, the depth, standard penetration test (SPT) N-value, and density have the strongest influence on the wave velocity estimation for K-NET. For KiK-net, the depth and site longitude have the strongest influence. The study confirms the applicability of commonly used machine-learning techniques in predicting wave velocities, and implies that exploring additional factors will enhance the performance.

波速剖面在岩石工程、石油工程与地震工程等诸多领域均具备重要应用价值。然而，波速的直接实测往往受限于时间、成本与现场场地条件。若无法获取波速实测数据，则需基于其他已知替代指标开展估算工作。本文提出机器学习（Machine Learning, ML）方法，用于预测日本地区的压缩波与剪切波波速（分别记为VP与VS）。本研究采用日本两大地震台网的钻孔数据库：强震观测网（Kyoshin Network, K-NET）与基盘强震观测网（Kiban Kyoshin Network, KiK-net）。研究选取了深度、N值、密度、坡度、海拔、地质条件、岩土类型以及场地坐标等多类影响因素。本研究选用三类机器学习技术：梯度提升树（Gradient Boosting, GB）、随机森林（Random Forest, RF）与人工神经网络（Artificial Neural Network, ANN），分别构建压缩波与剪切波波速的预测模型，并基于均方根误差与五折交叉验证法评估模型性能。针对两类地震台网的波速估算任务，基于梯度提升树的模型均取得了最优的估算效果。在所选影响因素中，对于K-NET台网的波速估算，深度、标准贯入试验（Standard Penetration Test, SPT）N值与密度的影响最为显著；而对于KiK-net台网，则为深度与场地经度的影响最为突出。本研究证实了常用机器学习技术在波速预测任务中的适用性，并表明引入更多额外影响因素可进一步提升模型的预测性能。