Automated optimization of underground cavern excavation sequences via deep reinforcement learning with graph-based physical response prediction

This is the authors' abstract. We don't add key points for this paper.

Safe and efficient excavation of underground cavern groups is a critical issue in hydropower and other large underground infrastructure projects. Inappropriate excavation sequences may induce adverse mechanical responses in the surrounding rock mass, elevating the risk of engineering hazards. Yet excavation sequence optimization remains challenging owing to the vast solution space, the high cost of numerical simulation, and the difficulty of handling practical engineering constraints. This study proposes Auto-Exca, a deep reinforcement learning framework assisted by a spatiotemporal graph surrogate model for automatic excavation sequence optimization. The framework encodes the evolving spatial interactions and rock mass parameters of the cavern group as a graph structure and formulates the optimization task as a Markov decision process. A dataset generated through three-dimensional FLAC3D simulations is used to train the surrogate model to predict 19 physical response variables for different excavation sequences. With this model providing rapid stability evaluation, a dual-stage reward mechanism integrates stepwise construction feedback with final stability assessment to jointly optimize efficiency and safety. Application to a real engineering project shows that the identified scheme requires fewer excavation decision steps and improves the majority of the evaluated stability indicators relative to the original design. Notably, by learning from reward signals, Auto-Exca not only recovers established engineering practices, such as prioritizing the excavation of major caverns, but also identifies more refined temporal coordination strategies. The proposed framework may support intelligent excavation planning in other complex underground projects after retraining with data from the target project.