Understanding transition paths in molecular systems is key to uncovering rare but essential events that govern transitions between metastable states, such as conformational changes and chemical reactions. Traditional molecular dynamics (MD) simulations often struggle to capture these phenomena due to the immense timescales and dimensionality involved. In this work, we present a framework that leverages Generative Flow Networks (GFlowNets) to learn distributions over transition trajectories efficiently and robustly. Our method integrates GFlowNets trained via the Sub-Trajectory Balance (SubTB) objective with physics-informed molecular dynamics simulations using OpenMM. By treating each trajectory as a stochastic path through a learned energy landscape, we can generate physically plausible transitions between metastable states while respecting thermodynamic consistency. This is accomplished by sampling state sequences from Langevin dynamics and using potential energy evaluations to compute Boltzmann-weighted rewards. Key improvements include adaptive control of drift magnitudes, stability-aware integration with small time steps, and robust handling of backward probabilities. We apply this framework to the canonical problem of alanine dipeptide isomerization, demonstrating the model's ability to capture meaningful transition behavior under physical constraints. This approach provides a promising pathway toward learning interpretable, sample-efficient dynamics in high-dimensional molecular systems, with broad implications in chemistry, biophysics, and materials science.