Deep Learning for Joint Geophysical Inversion of Seismic and MT datasets AbhinavPratap Singh 1 *, Divakar Vashisth 1 , Shalivahan Srivastava 1 1 Department of Applied Geophysics, Indian Institute of Technology (Indian School of Mines) Dhanbad Summary Using more than one geophysical technique provides a more reliable way to delineate the subsurface structure than a single geophysical method. In this paper, we present a novel way for the joint inversion of seismic and MT datasets using artificial neural networks. Different rock models are taken to compute the joint relation between the porosity, shale content, velocity and resistivity of the strata. The velocity and resistivity models thus generated were used to compute seismic traces and apparent resistivity curves, the dataset used for training and testing the neural network. Such a method of performing joint inversion is advantageous when we want to check the reliability of our Machine Learning model in regions of low velocity or when a particular physical property does not vary across the layers while another does since, unlike other optimisation schemes used for the geophysical inversion, we do not need to rerun the model for every initial model. The efficacy of artificial neural networks (ANN) to perform the join inversion is tested, and in the process, a particular type of ANN architecture is developed. Introduction Seismic and magnetotelluric (MT) methods are often used to map the subsurface by estimating the geophysical properties, namely velocity and resistivity variations. The seismic method plays a vital role in the delineation of near- surface geology for engineering purposes, hydrocarbon exploration as well as the Earth's crustal structure investigation (Li et al., 2019). MT method has found application in the exploration of oil, groundwater, mineral and geothermal systems (Simpson and Bahr, 2005). However, due to limitations inherent in each of these methods, separate inversions for resistivity and velocity models may not be optimal to fully understand the geology. Seismic interpretation techniques have found difficulty in delineating low-velocity layers (LVL) in the past (Heijst et al., 1994), while the MT method has faced issues with equivalence problem or resolving electrically thin conducting layer (Manglik and Verma, 1998). As there is an inherent ambiguity in the estimation of subsurface structure using a single geophysical method, using more than one method can reduce this ambiguity to a significant extent (Vozoff and Jupp, 2007;Manglik et al., 2011). Hence, we attempt joint inversion of seismic and MT datasets in the present study. Geophysical inverse problems are inherently ill-posed, multidimensional, multimodal and non-linear in nature (Sen and Stoffa, 2013). The solution to such a problem can be estimated using local optimisation (Conjugate gradient, Levenberg-Marquardt algorithm) or global optimisation (Particle Swarm Optimisation, Genetic Algorithm, Simulated Annealing) techniques. These optimisation schemes require computation of forward models in each iteration and subsequently reduce the misfit function between the computed and the observed data. In contrast, a neural network model trained on an acceptable range of geologically plausible models can be used to invert a number of geophysical datasets much more efficiently. This deep learning tool can approximate any linear or non-linear function with arbitrary precision (Van der Baan and Jutten, 2000) and hence have found success in geophysical inverse problems (Russell 2019) and other numerical domains in geophysics (Zhang et al., 2018). In this paper, we intend to use a deep learning approach for joint inversion of seismic and MT datasets. We begin by giving insights about the problem statement, followed by defining the architecture of the neural network model to carry out the joint inversion and then provide a general workflow at the end to perform joint inversion of different geophysical datasets. To carry out inversion using ANN, a number of steps have to be followed. First, a training dataset is generated, followed by NN training with a train set and testing on a validation set to measure their performance before employing them to the actual data. Forward modelling The resistivity and p-wave velocities may vary a lot for different formations and their relationship with lithology is generally complex. This has led to development of several empirical relations in different studies. While for a range of sedimentary rocks, we may agree on some of them, there is a high degree of uncertainty in case of unconsolidated sediments. As the top layer of our rock physics models is the one of unconsolidated sediments, so, we did not consider any particular relation to estimate the p-wave velocity for that layer. Other strata were assumed to be shale bearing water saturated sandstone. For these strata, certain rock physics models were taken to develop a joint relation between porosity, shale volume fraction, p-wave velocity and resistivity and consequently, velocity and resistivity models were generated. The seismic and MT responses were generated for the same lithology/rock physics model to create the training dataset for the neural network. 10.1190/segam2021-3583955.1 Page 1741 © 2021 Society of Exploration Geophysicists First International Meeting for Applied Geoscience & Energy Downloaded 09/03/21 to 3.86.69.137. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/page/policies/terms DOI:10.1190/segam2021-3583955.1