Normal ageing and Alzheimer’s disease (AD) are increasingly linked to altered large-scale brain dynamics, yet most imaging biomarkers remain based on static connectivity or regional atrophy. Here we propose a deep learning framework to quantify brain non-equilibrium—time-irreversible neural dynamics that reflect continuous energy and information exchange—using neuroimaging time series (primarily resting-state fMRI) and individual connectomes. The method learns subject-specific latent dynamics and optimises objectives sensitive to time directionality (e.g., forward–backward sequence discrimination) and entropy-producing / irreversibility-related signatures, while incorporating graph-structured priors to respect network topology. Using large-scale cohorts spanning cognitively normal ageing, mild cognitive impairment, and AD, we will (i) estimate individual non-equilibrium indices, (ii) evaluate associations with cognition and clinical stage, and (iii) localise network-level contributions via interpretable attribution and perturbation analyses. This project requires substantial HPC resources for repeated cross-validation, hyperparameter optimisation, and large-scale permutation/sensitivity testing across preprocessing and connectome definitions. The resulting scalable pipeline aims to deliver robust non-equilibrium biomarkers for ageing and AD, and an open, reproducible workflow for future methodological development