New paradigms in seismic data preconditioning and imaging via Deep Neural Networks

地球物理方法的主要目标是估计地球内部的图像。这些图像是探索资源和了解地下过程的核心信息。例如，在应用地震学中，进行了一项将地震波注入地下的实验。地质界面反射地震波，并将其传播回地表。地球表面的传感器记录了这些反射波。传感器收集的数据通过计算方法进行处理，将记录的信号转换为地球内部的图像。将地震数据映射到地下图像的计算方法主要分为两大类：预条件方法和反演方法。预处理和反演需要用额外的相关先验信息来解决数值问题。这种先验信息传统上是通过约束来提供的，这些约束确保具有给定特征的解(例如，平滑度)。近年来，机器学习领域提供了一些技术，允许学习上述先验信息，并通过这种方法找到涉及数据处理和倒置的许多数值问题的有效和准确答案。与提议的计划特别相关的是深度神经网络。这些都是强大的机器学习工具，在图像分类和分割任务中具有令人难以置信的性能。深度神经网络允许执行复杂的任务，如人脸识别、自动驾驶和诊断医学成像所需的任务。拟议的研究计划将研究将深度神经网络结合到解决应用地震学中出现的数据预处理和反问题的不同方法。我们将探索基于深度神经网络的正则化方法来寻找地震反问题的快速稳定解。同样，我们还计划研究基于深度神经网络的信号处理方法，并将这些方法用于地震信号增强和地震数据重建等预处理任务。我将在地震数据处理、地震成像和机器学习的基础上培训3名博士生、3名硕士学生、3名PDF和5名本科生研究助理。这一研究计划将增加关于机器学习在地学中的应用的科学知识，特别是在地震数据预处理和成像方法的发展方面。我预计，通过拟议计划培训的学生将在学术界或行业寻求职业生涯，并获得足够灵活的培训，以将他们的技能应用于任何需要工具从大数据集中获取、增强和提取有价值信息的行业。

The main goal of geophysical methods is to estimate images of the interior of the earth. These images are central information for exploring resources and understanding processes that occur in the subsurface. For instance, in applied seismology, an experiment is carried out where seismic waves are injected into the subsurface. Geological interfaces reflect seismic waves and propagate them back to the surface. Sensors on the earth's surface record these reflected waves. The data gathered by sensors are processed via computational methods that turn recorded signals into images of the earth's interior. Computational approaches to map seismic data into subsurface images are categorized into two main groups: preconditioning and inversion methods. Preconditioning and inversion require solving numerical problems with additional relevant prior information. This prior information is traditionally provided via constraints that ensure solutions with a given character (e.g. smoothness). In recent years, the field of machine learning has provided technologies that permit learning the aforementioned prior information and, by this means, finding efficient and accurate answers to many numerical problems that involve processing and inversion of data. Of particular relevance to the proposed plan are Deep Neural Networks. These are robust machine learning tools with incredible performance in image classification and segmentation tasks. Deep Neural Networks allow performing complex tasks such as those required for face recognition, autonomous driving, and diagnostic medical imaging. The proposed research plan will investigate different means to incorporate Deep Neural Networks into the solution of data preconditioning and inverse problems that arise in applied seismology. We will explore regularization methods based on Deep Neural Networks to find fast and stable solutions to seismic inverse problems. Similarly, we also plan to investigate signal processing methods based on Deep Neural Networks and use these methods for preconditioning tasks such as seismic signal enhancement and seismic data reconstruction. I will be training 3 Ph.D. students, 3 MSc students, 3 PDFs, and 5 undergraduate research assistants in the foundations of seismic data processing, seismic imaging, and machine learning. This reseach program will increase scientific knowledge on the application of machine learning in geosciences, particularly in the development of seismic data preconditioning and imaging methods. I expect students trained via the proposed plan to pursue careers in either academia or industry and to gain sufficiently flexible training to apply their skills to any industry requiring tools to acquire, enhance and extract valuable information from large datasets.