The mathematic model of accelerometer with neural network which has the characteristic of self-learning, self-adapting, nonlinear transformation is established and is used to calibrate the static coefficients of the accelerometer. It is proved that the calibrated values are exactly enough to compensate the accelerometer's output.