Wilton, E & Danielson, J 2025, 'Estimating borehole wall strength from televiewer data using machine learning', in JJ Potter & J Wesseloo (eds), SSIM 2025: Fourth International Slope Stability in Mining Conference, Australian Centre for Geomechanics, Perth, https://doi.org/10.36487/ACG_repo/2535_46
(https://papers.acg.uwa.edu.au/p/2535_46_Wilton/)
Abstract: Acoustic televiewers (ATVs) have become ubiquitous tools within the field of rock slope engineering. ATVs are geophysical probes which create an image of a borehole wall by emitting ultrasonic pulses and recording the reflected waves’ amplitudes and travel times. As denser rocks generally reflect waves of higher amplitudes, amplitude data can serve as useful proxies for rock strength. However, signal quality and interpretation can be influenced by confounding factors such as opaque borehole fluids, the centralisation of the probe within the borehole and the presence of mud or cuttings on the borehole wall.
In this study, a Python-based convolutional neural network, a commonly used architecture in computer vision, was trained to estimate rock strength grades from ATV amplitude logs. These logs were converted to images and matched with manually logged field strength grades to generate the training data for the model. Rock strength grades were based on the International Society of Rock Mechanics (ISRM) 1978 scale, which ranges from R0 (extremely weak) to R6 (extremely strong). The training dataset included data from many different boreholes, probes and project settings. The model’s prediction accuracy was assessed against a validation dataset and complete test surveys. The model reached a minimum mean square error value of 0.36 on the validation dataset and the model’s predictions on average were within approximately onehalfstrength grade of the manually logged grades for the test surveys.
By providing continuous, automated strength grade estimates – even in zones of no core recovery – this approach reduces subjectivity and improves coverage compared with traditional field tests. When benchmarked against manually logged strength grades, the model detected finer-scale changes in rock quality, particularly in faulted zones. This tool can assist in reviewing and corroborating traditional strength logs and forms part of BGC Engineering’s artificial intelligence-powered televiewer data processing platform, Structura.

Keywords: acoustic televiewer, machine learning, rock strength estimation, artificial intelligence, computer vision