Abstract In this paper, a new method of finite element model updating using neural networks is presented. Many previous model updating techniques have exhibited inconsistent performance when subjected to noisy experimental data. From this background it is clear that a successful model updating method must be resistant to experimental noise. A well-known property of neural networks is robustness in the presence of noise, and it is hoped to exploit this property for model updating purposes. The proposed updating method is tested on a simple simulated model, both in the absence and presence of noise, with promising results. A further advantage of this updating method is the ability to work with a limited number of experimentally measured degrees of freedom and modes.