AbstractThis article develops a safe reinforcement learning (SRL) algorithm for optimal control of nonlinear systems with input constraints. First, we design a novel performance index function by taking advantage of control Lyapunov‐barrier functions (CLBF) with inherent safety and stability properties to ensure closed‐loop stability and safety during operation under the optimal control policy. Additionally, since it is challenging to represent the CLBF‐based value function as an explicit function of process states, neural networks (NNs) are used to approximate the value function using the process operational data that indicate safe and unsafe operations. Theoretical results on the stability, safety, and optimality of the SRL algorithm are developed, accounting for the approximation error of the NN‐based value function. Finally, the efficacy of the proposed safe optimal control scheme is shown using an application to a chemical process example.