Combining first-principles methods with statistical mechanics is very important for accurately modeling molecules and materials under finite temperature and pressure, which closely match real-world experiments and natural environments. Here, through efficient first-principles molecular dynamics simulations and enhanced sampling techniques, we discovered that molecules related to life, such as glycine, ribose, urea, and uracil analogs, can form in C-H-O-N supercritical fluids. Additionally, our findings explain why ribose in RNA molecules adopts a five-membered ring structure rather than the more stable six-membered ring observed under ambient conditions. Furthermore, we developed first-principles Markov state models to elucidate the reaction mechanisms and dynamics of CO₂ dissolved in supercritical water, both in bulk and under nanoconfinement. Unlike previous simulations using enhanced sampling methods, our approach employs unsupervised machine learning techniques to automatically identify complex reaction coordinates and pathways involving multiple intermediates, without relying on prior assumptions. Our study provides valuable insights into the reaction kinetic network of aqueous carbon, yielding significant implications for the deep carbon cycle and the sequestration of CO₂. Considering that our simulations are all conducted under thermal equilibrium conditions, we recently proposed a deep learning variational model that can directly generate a temperature-differentiable canonical ensemble. This model can be combined with any explicit density generative model to obtain the system's partition function without notable bias. Our new method achieves accuracy comparable to Markov chain Monte Carlo but is more efficient.