Defining and quantifying a given brain state remains an unsolved and fundamental problem in modern neuroscience. Such a characterization could provide us with the fundamental laws of the complex human brain function in health and disease. In recent years, several theoretical frameworks have been developed to describe brain states, including wakefulness, sleep, stroke, anesthesia, aging, or neuropsychiatric disorders. These investigations have shown that the most crucial aspect of such frameworks is disentangling the dynamical complexity and causal mechanisms underlying brain states; nevertheless, they remain largely unknown. A deeper understanding of the mechanisms underlying brain function in health and disease would help design novel stratification solutions (i.e., divide a patient population into distinct subgroups based on their disease characteristics) and develop more personalized therapies for alleviating or even reverting brain disorders. In this project, we propose developing a novel thermodynamics framework based on physics that combines theoretical neuroscience, artificial intelligence, and clinical neuroscience. The framework will be applied to human neuroimaging data from patients suffering from brain disorders (attention deficit hyperactivity disorder, stroke, bipolar disorder, autism, and schizophrenia) and healthy participants (middle-aged and older adults) emphasizing women's health (from premenopausal to postmenopausal stages). In particular, the novelty of the framework is based on assessing the non-equilibrium nature of whole-brain dynamics in low-dimensional space (i.e., the latent space of brain dynamics). This new idea is important given that the latent space reduces the noise and dimensionality of brain signals. Empirical biomarkers will be  extracted by obtaining the directionality of information flow that breaks the balance of the underlying hierarchy in brain states in latent space. Furthermore, the framework will include a new generation of non-stationary whole-brain models in reduced latent space. This approach will provide model biomarkers by obtaining a robust estimation of the production entropy and, thus, the causal mechanistic principles of nonreversibility underlying brain states. Moreover, we will combine such whole-brain models with artificial perturbations to force transitions from brain disorders to healthy states by exhaustively exploring region-by-region stimulation in low dimensional space. This will allow us to uncover the potential mechanisms to reverse the dynamics in brain disorders towards more healthy regimes. Finally, we will use the empirical- and model-biomarkers extracted with deep learning algorithms to define better stratification psychiatry. This can help physicians with patient stratification and thus design more personalized treatments and optimize primary care management. Overall, using methods of thermodynamics from physics to characterize brain states can potentially provide novel biomarkers for brain disorders. We hypothesize that better signatures of dynamical complexity and causal mechanisms of non-reversibility in different brain states could arise from estimating the differential effects of the external (extrinsic) environment on the internal (intrinsic) brain dynamics in a low-dimensional space. This approach can provide tools for moving towards more naturalistic neuroscience, benefiting our understanding of brain disorders.