Since direct reverse modeling methods suffer low precision, poor stability and other short comings, we propose a Bayesian regularization neural network reverse modeling method for L1/2 norm regularization. L1/2 norm regularization makes the network structure sparse, so it can reduce network size and accelerate the training speed of the network. When the proportionality coefficient is greater than 0.5, using the Bayesian regularization method can make the output of the network more smooth, and improve the stability and generalization ability of the network. We apply this method to the design of Doherty power amplifier with known Doherty main power amplifier efficiency and output matching terminals S11 and S21. Simulations corresponding to the output power and f can simplify the design process. Experimental results show that  the mean square error of the output power, the f related to S11, and the f related to S21 obtained from the proposed model are reduced by 8.83%, 9.30% and 9% respectively in comparison with the direct reverse modeling method, and the running time is reduced by 99.34%, 99.40% and 99.23%. It can also solve the multisolution problem in design and be used in RF microwave devices design.