Michaels and Smith (1997) suggested to use neural networks to invert surface waves, inferring the sub-surface properties. Artificial neural networks are computer programs that simulate the biological neural networks. Calderón-Macías et al. (2000) also used them to inverse electrical data. From input stimuli (= observables values), it provides an output set of values (= model parameters). As a human brain, it needs education to react correctly in each situations. Hence, the neural network used for surface wave inversion is trained with series of the synthetic signals for which the model is perfectly known. To summarize, the network is a generic mean of mapping observable to model parameters.
A correct behaviour is obtained only if the network has been trained with synthetic models close to the true model. Hence, this method cannot be used to scan all potential models that correspond to experimental data. Moreover, the error propagation cannot be included in an easy way and the non-uniqueness is never handled.