[Paper Review] Predicting of shear wave velocity using Artificial Neural Networks

Franklin Adjei, Priscilla Afare+1 more|arXiv (Cornell University)|Nov 7, 2020

Seismic Imaging and Inversion Techniques7 references4 citations

TL;DR

This study proposes a multi-variable artificial neural network (ANN) model to predict shear wave velocity (Vs) in the Tano North Field using correlated well log data, overcoming limitations of direct measurement. The ANN outperformed other methods, achieving the highest R² and lowest MAPE, demonstrating superior accuracy in estimating Vs for geomechanical and lithological analysis.

ABSTRACT

Shear wave velocity is an important parameter for determining lithology, porosity and the dynamic properties in geo-mechanical studies. However, due to time and cost limitations, shear wave velocity is not available at all intervals and in all wells. In this paper, well logs with strong correlation to shear wave velocity were determined and used to predict the shear wave velocity for the Tano North Field. Four different methods were used to estimate the shear wave velocity under three different conditions. Then, based on obtained coefficient of determination and average absolute percent relative error between real and predicted values of shear wave velocity, the final results were compared. The results of this work demonstrated that the neural network based on multiple variables can estimate the shear wave velocity better than other methods used.

Motivation & Objective

• To address the scarcity of shear wave velocity data due to high cost and time constraints in geomechanical studies.
• To identify well logs strongly correlated with shear wave velocity for predictive modeling.
• To develop and compare multiple prediction methods, including artificial neural networks, for estimating Vs.
• To evaluate and select the most accurate method based on R² and average absolute percent relative error (MAPE).

Proposed method

• Identified well logs with strong correlation to shear wave velocity, such as compressional wave velocity, density, and porosity.
• Trained a multi-layer perceptron artificial neural network using multiple input variables to predict Vs.
• Applied three different modeling conditions to assess performance under varying data availability and input combinations.
• Evaluated model performance using coefficient of determination (R²) and average absolute percent relative error (MAPE) between predicted and actual Vs values.
• Compared the ANN model against three alternative prediction methods under the same conditions to ensure fair evaluation.
• Used 30 pages of results, 25 figures, and data from the Tano North Field to validate the model's predictive capability.

Experimental results

Research questions

• RQ1 Can artificial neural networks effectively predict shear wave velocity using correlated well log data in the Tano North Field?
• RQ2 How does the performance of a multi-variable ANN compare to other prediction methods in terms of R² and MAPE?
• RQ3 Which combination of well log inputs yields the most accurate shear wave velocity predictions?
• RQ4 Does the inclusion of multiple geophysical parameters improve prediction accuracy over single-variable models?
• RQ5 What is the optimal configuration of input variables and network architecture for accurate Vs estimation?

Key findings

• The multi-variable artificial neural network achieved the highest coefficient of determination (R²), indicating strong predictive accuracy.
• The ANN model recorded the lowest average absolute percent relative error (MAPE), demonstrating superior precision compared to other methods.
• Shear wave velocity predictions were significantly improved when multiple correlated well logs were used as inputs.
• The study confirmed that ANN-based modeling outperformed traditional regression and single-variable approaches in estimating Vs.
• The final model showed robust performance across different data conditions, validating its reliability for geomechanical applications.
• The results support the use of ANNs as a reliable alternative for predicting Vs in data-scarce environments.