Abstract: The availability of applying BP neural network algorithm to boost the optimization process of multi-composition neutron shielding materials,in which conventional Monte Carlo simulation would cause severe time consumption,has been discussed.A typical 3-layer BP neural network model has been established with 300 random mass components of composite materials and their corresponding dose values calculated by Monte Carlo as training samples.The absolute deviations between the predicted dose value and the sample value are within ±2.For validation samples outside the training samples,the absolute deviations expend to -6.4 to 5.2.According to the deviation distribution statistics,for over 70% of the total samples,the relative deviation absolute values are within 2%.The calculation accuracy and generalization ability of the neural network model are qualitatively determined to meet the requirements of the optimization algorithm.The cross-validation method can improve the calculation accuracy of the training samples,whereas increasing the calculation deviation of the verification samples,indicating that the balance between the fitting degree and generalization ability of the samples should be considered during the establishment of the neural network.

Key words: BP neural network, Monte Carlo, neutron shielding, material composition