This paper presents a modified neural network structure and a learning algorithm for direct inverse control of nonlinear and dynamic chemical processes. The major drawback of the typical direct inverse control, the difficuly of estimating the Jacobian for the plant, has been solved by introducing one-step ahead prediction model and batch updating rules. Using process prediction model offers the direct inverse control the ability to handle the processes with large time delays and quality variables that are difficult to measure on-line. The learning algorithm of the proposed architecture can be viewed as a receding horizon control. The proposed methods are illustrated by application case studies for optimal feedback control of batch reactor startup operations. The case studies have shown the proposed methods are better in terms of control performance and more robust to measurement noises, input disturbances, and model-plant mismatch than the previous ones.