Autonomous Exploration of Molecular Energy Landscapes using Reinforcement Learning and Diffusion Modeling

Scientific progress in molecular design increasingly depends on our ability to navigate vast chemical spaces with intelligence and autonomy. This thesis investigates how machine learning—particularly reinforcement learning (RL)—can be leveraged to discover stable molecular structures from first principles. We move beyond traditional simulation and dataset-driven paradigms by proposing self-guided agents that learn to generate molecules atom-by-atom, driven only by feedback from quantum-informed energy evaluations. Combining ideas from generative modeling, 3D geometric learning, and agent-based optimization, we explore the design of molecular generation policies capable of autonomous exploration and adaptation. A central contribution of this work is the development and evaluation of a molecular reinforcement learning framework, the results of which are presented in a dedicated research paper. Through this study, we aim to demonstrate both the potential and the current limitations of first-principles learning in molecular discovery, while outlining directions for future development of autonomous scientific systems.