Decoupling control provides an advantage of reducing control-loop interactions in multivariable control. The control system can be easily designed and operated. However, most decouplers are sensitive to modeling errors, especially for the system with relative gains being far away from unity. In practical application, decoupling control still presents some difficulties. This work is to design a neural network decoupler for decoupling control. The neural network decoupler is insensitive to modeling errors and does not have the unstable poles that are due to mathematical operations based on the transfer function model. Implementation of the neural network decoupler to a continuous stirred tank reactor (CSTR) demonstrates that decoupling control with the decoupler gives satisfactory results in control performance. The neural network decoupler improves the applicability of decoupling control.