Abstract: An inversion algorithm with neural network is put forward, which is applied to determining the parameters in unidirectional fiber composite by ultrasonic method. Based on Chistoffel equation, the method treats the composite as material of hexagonal symmetry. The unknown five elastic constants C₁₁, C₁₂, C₁₃, C₃₃ and C₄₄ are reconstructed by inversion from the measured velocities of sound propagating in the composite plate. The inversion algorithm lets the neural network repeatedly learn the training set renewing with better samples. Thus it overcomes the errors in the neural network forecast process which is caused by the limited number of learning samples, it improves calculation accuracy consequently. The presented method is used in the measurement and inversion of the unidirectional composite S2/DET85. The calculated velocity values according to the constructed elastic constants are in good agreement with the experimental results. This method realizes evolutionary inversion automatically in ultrasonic measuring the feature of unidirectional fiber composite.