Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

2408.02420v1

Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

Inversion of DC Resistivity Data using Physics-Informed Neural Networks

Rohan Sharma*1, Divakar Vashisth2, Kuldeep Sarkar1 and Upendra Kumar Singh1
1
Department of Applied Geophysics, Indian Institute of Technology Dhanbad, 2Department of Energy
Science and Engineering, Stanford University

email: * rohanlatha29@gmail.com

Summary

The inversion of DC resistivity data is a widely employed method for near-surface characterization.

Recently, deep learning-based inversion techniques have garnered significant attention due to their

capability to elucidate intricate non-linear relationships between geophysical data and model

parameters. Nevertheless, these methods face challenges such as limited training data availability and

the generation of geologically inconsistent solutions. These concerns can be mitigated through the

integration of a physics-informed approach. Moreover, the quantification of prediction uncertainty is

crucial yet often overlooked in deep learning-based inversion methodologies. In this study, we utilized

Convolutional Neural Networks (CNNs) based Physics-Informed Neural Networks (PINNs) to invert

both synthetic and field Schlumberger sounding data while also estimating prediction uncertainty via

Monte Carlo dropout. For both synthetic and field case studies, the median profile estimated by

PINNs is comparable to the results from existing literature, while also providing uncertainty

estimates. Therefore, PINNs demonstrate significant potential for broader applications in near-surface

characterization.

Introduction

The direct-current (DC) resistivity method is a widely utilized technique in near-surface geophysics,

attributed to its high sensitivity to variations in the resistivity of geological features and its cost-

effectiveness. It has been popularly employed in groundwater investigation (Benson et al., 1997) and

mineral exploration (Srigutomo et al., 2016). In DC resistivity surveys, an electric current is injected
into the ground, and the potential difference is measured to calculate the apparent resistivity using

Ohm's law. To obtain the true resistivity model of the subsurface and facilitate interpretation, this

apparent resistivity data is inverted. Some researchers have tried solving this inverse problem using

global optimization methods like simulated annealing (Sen et al., 1993), genetic algorithm (Liu et al.,

2012), and variable weight particle swarm optimizer - grey wolf optimizer (Sarkar et al., 2023).

The advent of machine learning algorithms has proven effective in geophysical inversion, enabling

complex nonlinear mappings between observed geophysical data and subsurface earth model

parameters. In a supervised setting, they have been employed to invert DC resistivity data (El-Qady

and Ushijima, 2001; Singh et al., 2019; Aleardi et al., 2021). However, inversion based on supervised

deep learning presents challenges, such as the limited availability of training data and estimations that

could be geologically and physically inconsistent. Additionally, there is a need for uncertainty

quantification in the predicted model parameters. These considerations are driving the adoption of

physics-informed methodologies in the field of geophysical inversion (Vashisth and Mukerji, 2022;

Liu et al., 2023a; Liu et al., 2023b).

This paper introduces a convolutional neural network (CNN) based, physics-informed deep learning

framework tailored for unsupervised 1D DC resistivity data inversion, drawing inspiration from

Vashisth and Mukerji (2022). Our primary contribution is the development of a lightweight CNN

architecture, applied to both case studies and employed to estimate uncertainties in the predicted

resistivities through Monte Carlo dropout. In the following sections, we outline our employed

methodology and demonstrate the efficacy of our approach through both synthetic and field case

studies.

Methodology

The DC resistivity Physics Informed Neural Network (PINN) architecture comprises of an encoder

and a decoder (Figure 1). The encoder consists of three convolutional blocks, each containing a 1D
convolutional layer, a rectified linear units (ReLU) layer for non-linearity, and a dropout layer with a

dropout rate of 0.2 for regularization and uncertainty estimation. The outputs from the final

convolutional block are flattened and fed into a fully connected layer with the number of output nodes

equal to the number of layered earth model parameters. To define an n-layered earth model, we

require 2n-1 model parameters, consisting of the resistivity (ρ) and thickness (h) of each layer, with

the bottommost layer assumed to be semi-infinite. The fully connected layer’s output is passed

through a sigmoid activation function to constrain the values between 0 and 1. Subsequently, these

values are rescaled to fall within the specified bounds (search space) for resistivities and thicknesses

before being fed into the decoder.

The output of the encoder serves as the input for the decoder, which incorporates the physics of

Schlumberger sounding (Koefoed, 1979). The decoder generates apparent resistivity curves

corresponding to the input resistivity profiles. The training phase makes use of the Adam optimizer

with an initial learning rate of 0.001, aiming to minimize the root mean squared (RMS) error between

the logarithms of the observed and predicted apparent resistivity curves. It is important to note that

incorporating the logarithm of the apparent resistivity values into the loss function was crucial for

achieving better fits, particularly in cases with highly varying resistivities. The learning rate was

reduced by a factor of 0.8 if there was no improvement in the loss for 15 consecutive epochs. Upon

completion of the training process, the encoder's output vector acquires meaningful representations in

the form of the learned layered earth model parameters. This allows the encoder to directly predict

resistivity profiles from the apparent resistivity curves. The values of all hyperparameters within the

architecture were determined through an extensive exploration of various combinations.


Figure 1 Unsupervised DC Resistivity PINN architecture. The encoder is a CNN trained to generate

resistivity profile from a given input apparent resistivity curve. The decoder incorporates the physics

of Schlumberger sounding to guide the encoder’s learning process.

During the prediction phase, dropout is still employed to mimic the behaviour observed during

training. This approach, known as Monte Carlo Dropout, serves to approximate the posterior

distribution of the model parameters, thereby aiding in uncertainty estimation (Gal and Ghahramani,

2016). By repeating this process multiple times (in this case, 10000 times), empirical samples of the

predictions are generated. These samples facilitate the computation of approximate posterior

percentiles. Notably, the 50th percentile (P-50) represents the median value, while the range between

the 5th (P-5) and 95th percentile (P-95) provides an estimate of the network's prediction uncertainty

(Das and Mukerji, 2019).

Results and Discussion

In this study, we utilize the same vertical electric sounding (VES) synthetic and field dataset as

presented in Sarkar et al. (2023). Our analysis involves comparison of the results obtained through
PINNs with those derived from the application of the variable weight particle swarm optimizer - grey

wolf optimizer (vPSOGWO) algorithm, as demonstrated in Sarkar et al. (2023).

The synthetic apparent resistivity dataset, incorporating 10% Gaussian noise (Figure 2a), is derived

from a three-layered earth model (Figure 2b) with resistivities conforming to an H-type curve. Figure

2b demonstrates that the median (P-50) of PINNs predicted resistivity profiles closely matches the

true profile (Figure 2b), with the computed apparent resistivity curve for this profile (Figure 2a)

exhibiting an RMS error of 16.55 Ωm, significantly outperforming the 45.39 Ωm RMS error

associated with vPSOGWO algorithm. Furthermore, the uncertainty in the predicted resistivity

profiles (spread between P-5 and P-95) remains relatively constant with depth, highlighting

consistency in the estimations.

For the field case study (Figure 3), VES data from Digha, Medinipur (West Bengal) is inverted. The

geological profile for the area is characterized by alternating layers of sedimentary rock containing

sand and clay, overlaid by a thin alluvial layer. The apparent resistivity curve (Figure 3a)

corresponding to the PINNs estimated P-50 resistivity profile (Figure 3b) exhibits a strong alignment

with the observed data, yielding an RMS error of 0.49 Ωm, which is comparable to the 0.50 Ωm RMS

error given by vPSOGWO algorithm. Notably, the uncertainty in the PINNs predicted resistivity

profiles decreased considerably below the estimated conductive alluvial layer, indicating a higher

level of confidence in the model's predictions in this region.


Figure 2 (a) The observed (black) and computed DC apparent resistivity curves for the synthetic case

study. (b) The true (black) and estimated resistivity profiles from PINNs and vPSOGWO. The PINNs

estimated median resistivity profile (P-50) and its corresponding apparent resistivities are illustrated

in pink, while the predictions highlighting the uncertainty range (between P-5 and P-95) are shown in

green. The vPSOGWO estimated resistivity profile and apparent resistivities are shown in yellow.

Figure 3 (a) The observed (black) and computed DC apparent resistivity curves for the Digha field

case study. (b) The estimated resistivity profiles from PINNs and vPSOGWO. The PINNs estimated

median resistivity profile (P-50) and its corresponding apparent resistivities are illustrated in pink,

while the predictions highlighting the uncertainty range (between P-5 and P-95) are shown in green.

The vPSOGWO estimated resistivity profile and apparent resistivities are shown in yellow.
The findings from both synthetic and field case studies underscore the robustness and accuracy of the

proposed PINNs architecture for inverting DC resistivity data, with all observed apparent resistivity

data points falling within the model-predicted uncertainty bounds. Future work will explore the

application of PINNs to the inversion of 2D and 3D Electrical Resistivity Tomography (ERT) data.

Conclusions

This study highlights the efficacy of the proposed PINNs architecture in accurately inverting DC

resistivity data, showcasing performance comparable to that reported in existing literature. The

architecture proved to be robust in handling both synthetic and field datasets, predicting consistent

resistivity profiles and providing reliable uncertainty estimates. These results underscore the potential

of PINNs as a reliable and valuable tool for geophysical data inversion tasks and near-surface

characterization.

References

Aleardi, M., Vinciguerra, A., & Hojat, A. [2021]. A convolutional neural network approach to

electrical resistivity tomography. Journal of Applied Geophysics, 193, Oct.

Benson A. K., Payne K. L., & Stubben M. A. [1997]. Mapping Groundwater Contamination using DC

Resistivity and VLF Geophysical Methods-a Case Study. Geophysics, 62, 80-86.

Das, V., & Mukerji, T. [2020]. Petrophysical properties prediction from prestack seismic data using

convolutional neural networks. Geophysics, 85(5), N41-N55.

El-Qady, G. and Ushijima, K. [2001]. Inversion of DC resistivity data using neural networks.

Geophysical Prospecting, 49, 417-430.


Gal, Y., & Ghahramani, Z. [2016]. Dropout as a Bayesian Approximation: Representing Model

Uncertainty in Deep Learning. The 33rd International Conference on Machine Learning, PMLR, 48,

1050-1059.

Koefoed, O. [1979]. Geosounding principles 1. Resistivity sounding measurements. Elsevier Science

Publishing Company, Amsterdam.

Liu, B., Li, S.C., Nie, L.C., Wang, J., L, X., & Zhang, Q.S. [2012]. 3D resistivity inversion using an

improved Genetic Algorithm based on control method of mutation direction. Journal of Applied

Geophysics, 87, 1-8.

Liu, B., Pang, Y., Jiang, P., Liu, Z., Liu, B., Zhang, Y., Cai, Y., & Liu, J. [2023a]. Physics-Driven

Deep Learning Inversion for Direct Current Resistivity Survey Data. IEEE Transactions on

Geoscience and Remote Sensing, 61, 1-11.

Liu, M., Vashisth, D., Grana, D., and Mukerji, T. [2023b]. Joint inversion of geophysical data for

geologic carbon sequestration monitoring: A differentiable physics-informed neural network model.

Journal of Geophysical Research: Solid Earth, 128(3), e2022JB025372.

Sarkar, K., Mukesh, and Singh, U.K. [2023]. Nature-inspired stochastic hybrid technique for joint and

individual inversion of DC and MT data. Sci Rep, 13, 2668.

Sen, M. K., Bhattacharya, B. B., & Stoffa, P. L. [1993]. Nonlinear inversion of resistivity sounding

data. Geophysics, 58, 496–507.

Singh, A. P., Vashisth, D., and Srivastava, S. [2019]. Random Forest Regressor for Layered Earth

data inversion, AGU Fall Meeting 2019, S53D– 0483.


Srigutomo, W., Trimadona, & Pratomo, P. M. [2016]. 2D resistivity and induced polarization

measurement for manganese ore exploration. Journal of Physics: Conference Series, 739, 012138.

Vashisth, D., & Mukerji, T. [2022]. Direct estimation of porosity from seismic data using rock-and

wave-physics-informed neural networks. The Leading Edge, 41(12), 840-846.

You might also like