A physics-based neural network approach for geophysical inversions

地球物理成像有助于我们对地球及其地质、构造和火山历史的了解。然而，大多数地球物理成像使用单一数据类型，而不是联合使用多种数据类型，导致图像可能与其他数据不兼容。这种不相容可能导致对地球和潜在地球物理过程的误导或错误推断。该项目通过利用具有物理约束的新型人工智能工具，研究了一种无缝使用重力和地震数据的方法。该方法有望简单有效地应用于许多不同的地球物理成像问题，因此有可能提高我们对地球构造历史的理解。在其他情况下，增强的图像可能有助于确定如何最好地应对地震和火山爆发等自然灾害。这项工作还有助于连接人工智能和地球物理科学界，这两个社区将从更多的相互作用中受益。此外，该项目还支持研究生和本科生，以及普罗维登斯公立学校的推广工作，其中包括少数民族和低收入高中学生。联合反演可以最大限度地减少使用不同类型地球物理数据所固有的权衡，这在实施上是具有挑战性的，部分原因是每种数据类型都需要专门的专业知识，建立每种类型的反演存在独特的困难，以及作为此类研究的一部分必须分析的数据量和模拟。机器学习在解决上述三个挑战方面取得了很大的进步，但在计算上仍然难以实现，对于全波形地震和重力数据的联合反演也是如此。这项工作使用了一种新的机器学习框架，称为物理通知神经网络（pinn）来实现这些数据的联合反演。PINN框架利用对底层物理模型的理解来降低构建神经网络的成本，以准确地近似波传播或泊松方程控制相关数据。基于pinto的联合反演方法的一个测试目标是利用洛杉矶盆地的现场数据，在那里可以使用大量的现场地质和构造现场数据来测试联合反演是否实现了提高真实地质特征分辨率的目标。该项目由地球科学理事会共同资助，旨在支持人工智能/机器学习在地球科学领域的发展。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Geophysical imaging has contributed to much of our understanding of the Earth and its geologic, tectonic, and volcanic history. However, most geophysical imaging utilizes single data types rather than using multiple data types jointly, resulting in images that can be incompatible with other data. Such incompatibilities can lead to misleading or erroneous inferences about the Earth and the underlying geophysical processes. This project investigates a method for allowing seamless usage of both gravity and seismic data jointly by leveraging new artificial intelligence tools that have physics based constraints. The method promises to be easily and efficiently applied to many different geophysical imaging problems and therefore potentially improve our understanding of Earth’s tectonic history. In other contexts, the enhanced images may help determine how to best respond to natural hazards like earthquakes and volcanic eruptions. This work also helps connect the artificial intelligence and geophysics scientific communities, two communities that would benefit from more interactions with each other. In addition, the project supports graduate and undergraduate students, as well as outreach efforts in the Providence public schools that includes minority and low-income high school students.Joint inversions that robustly minimize the tradeoffs inherent in using different types of geophysical data are challenging to implement, in part due to the specialized expertise needed for each data type, the unique difficulties in setting up each type of inversion, and the quantity of data and simulations that must be analyzed as part of such studies. Machine learning has made great strides in addressing all three of the above challenges, but it still often remains computationally impractical to implement robustly and this is true of joint inversions of full-waveform seismic and gravity data. This work uses a new machine learning framework called physics informed neural networks (PINNs) to implement joint inversions of such data. The PINN framework leverages an understanding of the underlying physical model to reduce the cost of building a neural network to accurately approximate the wave propagation or Poisson equation governing the relevant data. One test target of the PINN-based joint inversion methodology is using field data from the Los Angeles Basin, where numerous prior in situ geological and structural field data can be used to test whether the joint inversion accomplishes the goal of improving the resolution of true geologic features. This project is co-funded by the Directorate for Geosciences to support AI/ML advancement in the geosciences.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.