Identification of Induced Polarization affected Transient Electromagnetic Data by Deep Learning

Recent advances in transient electromagnetic (TEM) methodologies have increasingly highlighted the presence of induced polarization (IP) effects across various geological settings. Modeling these effects involves treating subsurface resistivity as frequency-dependent. However, when IP effects are minimal, this complex resistivity simplifies to a real-valued form, allowing the subsurface to be modelled through simpler approaches. Thus, the computational complexity of modeling TEM data significantly varies depending on the presence of IP effects. Additionally, high correlation among IP model parameters increases the number of equivalent models, further complicating the inverse modeling process. The problem exacerbates as TEM data with minimal IP effects can also be modeled using IP parameters due to the correlation issues and inversion limitations, potentially distorting the real-valued resistivity parameter. These complexities necessitate the need for a two-stage process, aligned with Occam's Razor, which advocates adopting the simplest conditions unless more complex ones are necessary. However, manually distinguishing these data sets is complex and time-consuming, and often requires extensive expert analysis.<br/><br/>To address these challenges, we introduce a two-stage deep learning method that automates the identification of TEM data across three categories: data suitable for resistivity-only modeling, data requiring IP modeling, and data that cannot be modeled using either method. The first stage distinguishes between data that can be modeled using simple 1-D resistivity-only conditions and data that falls outside this assumption. Subsequently, the second stage then focuses on the latter, further distinguishing data that necessitates the incorporation of IP parameters and data that remains unmodelable. We test the effectiveness of our methodology on a small tTEM dataset acquired in Kigadye, Tanzania, which covers 12-line kms. The results demonstrate that the proposed approach effectively distinguishes between data that can be modeled using resistivity-only conditions and those requiring IP modeling. With this distinction in place, the next step calls for an inversion methodology that can collectively model these data based on their respective categories.