GNGTS 2023 - Atti del 41° Convegno Nazionale

Session 3.3 ______ ___ GNGTS 2023 (GA+NN), both performed in a DCT compressed domain, shows that, although data misfit and model misfit increase in GA+NN, the reconstruction of the resistivity distribution is satisfactory and reproduces the main features visible in the result of the GA inversion. Fig. 2 also represents the prediction error of the inversions, defined as the absolute value of the difference between true and predicted model. In particular, Fig. 2b shows the ability reached by the CNN in predicting the data misfit value of models on which the net was trained or not. Fig. 3 compares the results in terms of data, seen as pseudosections and vectors. The most important advantage of the proposed method is the reduction of the computational effort required: a GA inversion needs about 130 minutes, whilst GA+NN (15000 models dataset generation, training and inversion) takes about 70 minutes. Most of this time is required to create the dataset of models, whilst the training phase takes about 5 minutes and the inversion only 1 minute. The number of forward modelling required drops from almost 32000 for GA to 16000 for GA+NN. Conclusions In this work, we have presented an algorithm, named GA+NN, that uses a CNN to substitute the forward modelling and the data misfit computation steps in an inversion performed with a global optimization method. We have applied this method to the inversion of ERT data in a synthetic case. Our work shows how a CNN can be used in this context to speed up the inversion without losing the ability to get reliable results, but reducing its computational effort, halving both forward modelling computations and the overall computational time with respect to a standard GA inversion in which a FE code perform the forward modelling evaluation for each considered model. What makes this approach promising is the possibility to obtain preliminary results in a short time (less than 10 minutes for training and inversion). To make this method more effective, it might help to have some datasets of models with different characteristics, in order to adapt the training dataset to the specific geologic situation. Indeed, training the CNN with models from prior distributions whose statistical properties are completely different from the ones of the true model affects the quality of the result. A further step to continue this research path is, of course, the application of GA+NN to real data. To improve the quality of the results, we could apply the NN not in the whole inversion, but only in the first phase or to only a portion of the models and, possibly, make use of techniques like transfer learning. Moreover, since most of the time required by GA+NN relies on the training dataset generation, further investigations with the aim to accelerate this phase could be useful.