A new finite element analysis principle based on neural networks is proposed. It can realize the finite element analysis in real time. The neural computation of a simple structure was simulated on digital computer, the effect of input parameters on convergent speed and precision was analyzed. Before solving a certain FE problem, the linking weight matrix of the neural net, i.e., the global stiffness matrix of the structure analyzed, must be defined. In this paper, a method which uses the BP net to compute the element stiffness matrix in real time is proposed.