Controlled source electromagnetic (CSEM) data are used to image the resistivity properties of the subsurface. The technology has developed significantly the last 20 years, with several applications in the petroleum industry, marine mineral mining, freshwater mapping and geotechnical studies. The main obstacle for the use of CSEM is the ambiguity in the results, the measurements must be modelled to represent the 3D geobodies. Several models may satisfy the measurements, and the end users of the results are often left with a lot of uncertainties.
We propose to combine the CSEM technologies with the recent developments in Artificial Intelligence (AI) technologies, as we see great potential for improvements in the inversion and modelling of CSEM data using AI methods in a novel workflow. The advantages of using AI in modelling of CSEM data lies in that a deep learning network trained on traditional processing will be able to handle complex problems in a better way, finding relationships outside the equations. Shorter computation time will also provide cost effective and quicker results, allowing for more alternative models to be tested. Our aim is to estimate and present uncertainties instead of "the model" and in this way improve decision making.
CSEM data is proven to be very valuable in hydrocarbon exploration due to the sensitivity of electromagnetic energy to net reservoir volume (porosity) and hydrocarbon saturation that is derived from the resistivity distribution. CSEM, however, does not measure resistivity directly. To obtain resistivity, an inverse problem must be solved. Under inverse problem we understand the situation where electromagnetic response to given sources is measured, while resistivity distribution is unknown and must be estimated. However, this problem is ill-posed from the mathematical standpoint. CSEM inversion is a non-linear problem, that is, small changes in the measurements can lead to large changes in the estimated model parameters; moreover, non-linearity of the problem is even more pronounced by the presence of noise. In addition, the solution of the problem is non-unique since observations can fit more than one model. Complexity and instability of the traditional inversion process prevents CSEM from being used to its full capacity. In recent years researchers started to explore possibilities to apply deep learning methods to the geophysical inversion. The results demonstrate capabilities of the deep learning to cope with such a challenging problem, however, examples presented are more of an illustrative nature. In this project we will further investigate applicability of ML methodologies to processing and inversion of CSEM data. Our rationale behind the choice of this methodology is connected to the opportunity to reduce significantly processing and inversion time. Our goal in this project is to develop a methodology for efficient probabilistic inversion of CSEM data. Both PINNs and cGANs use already probabilistic principles, and building on top of these methods will allow for the development of the probabilistic inversion methodology that is much faster than the traditional one.