Physics-enhanced reinforcement learning for real-time optimal control of dynamical systems — ThinkLLM