A Genetic-Evolutionary ADAM (G-ADAM) optimizer for semisupervised convolutional neural network-based acoustic impedance inversion

Recently, deep neural networks have been widely used for seismic inversion due to the availability of computational resources. However, having the capability to perform complex non-linear mapping it requires large amount of labeled data (well-log data) which in general is rarely available at every location. To address the challenge, we proposed a semi-supervised learning approach with a hybrid optimizer, i.e., Genetic-evolutionary adaptive moment estimation (G-ADAM) for acoustic impedance estimation from seismic data. In this approach, reflectivity and seismic loss functions are used to constrain the framework. Seismic loss function holds the dominant role in updating weights and biases of the network. Therefore, it circumvents the need of large amount of well-log/labeled data. Performance of the proposed G-ADAM based framework and its comparison with ADAM was done based on Marmousi2 synthetic model and F3 real seismic dataset. A better accuracy (∼98%) was achieved by G-ADAM optimizer than the 96% accuracy of ADAM. However, the convergence of G-ADAM was relatively slow as compared to ADAM. The results indicated that with G-ADAM best optimal solution can be achieved. Therefore, the proposed framework could be used as impedance inversion tool for seismic reservoir characterization.