Molecular dynamics (MD) simulations are a fundamental tool for studying the microscopic dynamics and thermodynamic properties of molecular systems. However, obtaining reliable long-time behavior from direct simulations can be computationally demanding, particularly when relevant processes involve rare events or widely separated timescales. This project investigates machine learning methods for learning molecular dynamics from simulation data, with the objective of enabling accurate and efficient prediction of long-time dynamics and equilibrium properties. A central focus is the development and analysis of diffusion-based generative models for learning Markov transition processes from molecular dynamics trajectories. Such models provide a flexible framework for approximating conditional transition distributions at finite lag times, potentially overcoming limitations of conventional linear or low-dimensional approximations of transfer operators. The learned transition models will be studied with respect to their ability to reproduce important dynamical and statistical properties of the underlying molecular system, including Markovianity, equilibrium distributions, relaxation dynamics, and thermodynamic consistency. As an initial research direction, the project will investigate detailed-balance-based reweighting methods for estimating thermodynamic equilibrium distributions from finite and potentially non-equilibrium MD trajectories. The approach exploits the symmetry imposed by detailed balance on forward and time-reversed transition probabilities to learn configuration-dependent statistical weights. This provides a complementary approach to conventional transfer-operator methods, while avoiding the need to explicitly solve an eigenvalue problem and allowing more expressive machine learning models to be employed. The project will combine theoretical analysis with numerical experiments on molecular simulation datasets. Particular attention will be given to the accuracy, stability, and scalability of the proposed methods, as well as to their ability to preserve physically relevant properties when extrapolated to long timescales. The overall goal is to develop principled machine learning methodologies for molecular simulation that combine the flexibility of modern generative models with the physical structure and constraints imposed by molecular dynamics. The resulting methods are intended to provide improved tools for studying long-time molecular behavior and estimating thermodynamic properties from computationally accessible simulation data. Main supervisor: Simon Olsson, Chalmers University of Technology.