PDARTS: busca de arquitetura neural diferenciável para inversão sísmica

Abstract

Seismic inversion is an inverse problem that minimizes both amplitude and phase differences between simulated signals and real observed signals through an optimization problem. Due to its high computational cost and the industry demand for higher resolution, researchers often explore ways to accelerate the process and improve its accuracy. In the last decade, deep learning emerged as a promising alternative for seismic inversion; however, still demanding laborious trial and error processes, integration of domain knowledge, and hyperparameter tuning. To improve model architectures for this task, this paper introduces PDARTS (Projected Differentiable Architecture Search), a method inspired by DARTS which aims to leverage the generalization capability of neural blocks such as Fourier and U-Fourier blocks for the seismic inversion task. Because Fourier-based blocks rely on Fourier transforms, which require square inputs, the asymmetric shape (height, width) of original seismic data necessitates enforcing a square input format to ensure proper operation. To achieve this, PDARTS employs fixed encoder layers connected to a projection layer to reshape spatial dimensions and a convolutional layer to mitigate noise. The decoder, along with the final layers of the encoder, is implemented as a DARTS supernetwork, where neural architecture search (NAS) is conducted to explore the potential of Fourier and U-Fourier neural blocks, aiming to discover new neural networks with better performance. Experiments demonstrated that the best architecture discovered by our method, PDARTSNet, outperforms current state-of-the-art neural networks for seismic inversion. Furthermore, it demonstrates that despite the increased number of network parameters and consequent higher computational costs, PDARTS has proven capable of discovering neural networks with superior performance for seismic inversion.