Experimental Characterization and Hybrid LSTM‐RF Modeling of Time‐Dependent Interlayer Behavior in Mass Concrete Under Extreme Multi‐Physical Environments

ABSTRACT Addressing the unclear mechanisms and insufficient prediction accuracy regarding the effects of extreme weather on the interlayer performance of mass concrete in Northwest China, this study proposes a novel framework for interlayer performance prediction and risk early warning, integrating multi‐condition physical experiments with machine learning methods. The research simulates five typical extreme environmental conditions (high temperature, strong wind, rapid cooling, high temperature with strong wind, and cold wave with strong wind), and systematically quantifies the influence of various interlayer treatment measures (natural curing, thermal insulation covering, artificial grooves, and incorporation of PVA fibers) on the time‐varying characteristics of concrete moisture content, penetration resistance development, splitting tensile strength, and crack resistance. Experimental results indicate that extreme compound conditions lead to a maximum reduction of 62.8% in interlayer splitting strength, with the combined effect of cold wave and strong wind being the most significant. The synergistic use of thermal insulation blankets and PVA fibers can achieve a strength recovery rate of up to 85.3%. Based on multi‐source experimental data features, an LSTM‐RF hybrid prediction model was constructed, where the long short‐term memory (LSTM) network specifically processes the time‐series features of moisture content and penetration resistance (prediction R 2 > 0.92). The established “physical experiment–digital modeling” dual‐driven approach provides a quantifiable decision‐making basis for concrete construction in extreme environments.