Taking the split Hopkinson pressure bar(SHPB) test data as training samples, an implicit dynamic constitutive model of steel fiber reinforced concrete(SFRC) was proposed with the improved BP neural network program. On the basis of ABAQUS/Explicit, a userdefined material subroutine VUMAT was compiled with FORTRAN language, embedding with the neural network constitutive model into finite element calculation. With SFRC structure＇s response under impact load as case study, it is shown that the constructed constitutive model and its finite element method are effective.