Optimal Geophone Arrays and 3D Seismic Tomography with Physics-Informed Neural Networks for Subterranean Shipyard Mapping

Obtaining an accurate map and assessment of subterranean site conditions is an important component of geotechnical engineering. However, despite recent advances in seismic imaging and tomography this is still very much an open problem and one that can lead the use of large factors of safety and can result in substantial additional project costs and delays as geophysical features may be unexpectedly revealed during later phases of construction. Offshore conditions in ports and shipyards further compound the problem, and very few options exist for practical yet accurate geotechnical site assessment. However, recent advances now offer several ways in which modern machine learning techniques can be applied to provide an advanced seismic imaging and tomography capability, these include:  Recent advances in physics-informed neural network (PINN) 3-D subsurface map generation from seismic geophone data provides an approach that is robust to noise from secondary reflections, and has the potential to overcome many of the traditional of current linearization-based techniques.  The second way that machine learning may be applied to this problem, is in image enhancement. In this way machine learning can be paired with traditional image reconstruction techniques but used in a capacity similar to that of super resolution to provide additional fidelity or robustness to noise is in ways that standard linearization algorithms cannot.  Finally, modern machine learning techniques can be utilized to identify and provide semantic understanding of 3-D volumetric data. For example, identification of void locations and or soil types can be automatically mapped and associated with individual of voxels within a 3-D image representation of the reconstructed soil medium During Phase I, we will assess the feasibility of each of two of the aforementioned techniques, namely tomographic reconstruction and image interpretation (1 and 3, respectively). We’ll perform optimization and automatically assign semantic labeling to identify voids, anomalies and geophysical features to tomographic maps during the Phase I Option. During Phase II, we would propose to also include 3D tomographic image enhancement, however, we would propose to focus on the core challenges (1, 3) during Phase I. In this proposal, we will briefly present the fundamental equations, architectures, and technical approach that we will propose to undertake to implement them as means of quickly determining the feasibility of the approach. We will also identify areas where additional data or effort may be required, and will look to provide a solid foundation for a successful Phase II effort, where data will be gathered in the field under conditions similar to those required by the Navy and used to calibrate the models developed and trained prior. Validation examples will include localization and dimensions of timber-constructed relieving structures, piles, structural details, patterns, and missing elements.