<p>The stability investigation of drifts for deep geological radioactive waste disposal has been extensively studied in the last two decades, and the previous works highlighted its high dependence not only on the short- and long-term behavior but also on the uncertainty of surrounding rock’s properties. Numerical modeling and Artificial Intelligence (AI) have been largely chosen for the deterministic analysis and used as the prediction model of the uncertainty quantification and stochastic modeling process. While numerical simulation can provide results that respect a high-fidelity of physical behavior, the expensive computational time presents an obstacle for the uncertainty quantification and reliability analysis. The AI-based surrogates have been shown their usefulness when they can replace numerical modeling as the black box for probabilistic inversion and stochastic modeling. However, it still lacks a relevant AI methodology that can efficiently predict the complex response of this underground structure involving time-dependence and anisotropy of surrounding rock mass. This study presents a novel approach that integrates in the updated Bayesian inference process a new ANN-based surrogate, called ANN-NBEATS model. While this ANN-NBEATS surrogate allows to reduce the time consuming of the prediction of drift behavior over time, the updated Bayesian inversion can significantly improve the uncertainty quantification accuracy by updating in a sequential manner the posterior distribution of the short- and long-term mechanical properties of the rock formation. The results demonstrate that the ANN-NBEATS surrogate outperforms conventional modelling of deep tunnel behavior forecasting and its integration in the updated Bayesian inference can offer a powerful tool for uncertainty quantification of complex behavior of geological rock formation.</p>