Intelligent prediction of rock mechanical parameters using LWD data with a hybrid LSTM-MoE model

作者：Sheng Li, Houzhen Wei, Rui Xu, SUN Fengyi, Mei Bai, Xiaolong Ma · 发表于：Geomechanics and Geoengineering · 年份：2026 · DOI：10.1080/17486025.2026.2664024 · 被引用次数：1 · 研究领域：Rock Mechanics and Modeling、Machine Learning in Materials Science、Tunneling and Rock Mechanics

Rapid and precise estimation of rock mechanical parameters, such as elastic modulus and Poisson’s ratio, is a critical yet unresolved challenge in geotechnical engineering. Traditional laboratory and in situ tests are often slow, expensive, provide only discrete, point-based data, and can suffer from sample disturbance or size effects. In this study, a rock drilling model compliant with the Drucker–Prager (D-P) criterion is first built via finite element simulation. After validation, a hybrid Long Short‑Term Memory (LSTM) and Mixture‑of‑Experts (MoE) neural network is developed to predict the elastic modulus and Poisson’s ratio directly from weight on bit (WOB) and torque time series. Verified through laboratory tests on seven typical sandstone lithologies, this model demonstrates significantly higher prediction accuracy than conventional methods (R2 = 0.9351 for elastic modulus, R2 = 0.9317 for Poisson’s ratio). This approach provides reliable numerical simulation data and is capable of acquiring rock mechanics parameters reliably, continuously, and in real time, meeting the urgent demand for real-time, while-drilling data in engineering practices such as tunnel rock mass classification, slope stability evaluation, and foundation bearing capacity computation.