This letter proposes a knowledge-guided optimization framework based on dual artificial neural network (ANN) surrogates to enable fast reconfigurable intelligent surface (RIS) units design. In the proposed framework, we first train two ANN surrogates to achieve a fast and accurate prediction of the electromagnetic (EM) amplitude and phase responses of the RIS unit corresponding to the two extreme working states of the tunable component. Then, leveraging the prior knowledge that the EM responses monotonically vary with the tunable component, accurate predictions of the maximum reflective amplitude loss and phase shift of the RIS unit can be achieved. This allows us to utilize the trained ANNs to perform surrogate-assisted dual-objective optimization, which minimizes the reflective amplitude loss while maximizing the phase shift. The proposed framework is applied to design optimizations of two multibit RIS units, demonstrating significant improvements in design efficiency and solution optimality compared to direct EM optimization. Numerical simulation and experimental results further confirm the proposed model’s accuracy.