Created by Katie Silversides, Senior Geo-data Scientist
TL;DR: Measure while drilling data can be used to extract coal seam depths without adding additional sensors. A 1D convolutional neural network (CNN) was used to locate coal boundaries in the specific energy–achieving 86% accuracy and creating reliable coal surface maps. This work was conducted in collaboration with IMDEX.
Measure while drilling (MWD) data can provide additional information about a coal seam without additional scanning equipment or time. In particular, it can provide critical information about coal depth and thickness, allowing for more accurate mining and helping prevent coal damage from blasting. But what if we could extract this information from data we’re already collecting?
This study uses MWD data collected during blast hole drilling in a coal mine. The aim is to identify holes where the drill stops at the top of the coal seam (touch coals) and where the drill passes through a coal seam (through coals)–both essential for blast design and resource tracking.
Turning drilling data into geological intelligence
Typical MWD measurements can include rate of penetration, torque and force on bit. These can be combined to calculate specific energy (SE) which is a measure of the energy required to drill through the rock. When the drill moves from sandstone into the softer coal, the specific energy drops–a signature we can use a machine to learn automatically. This approach to property prediction from drilling data represents a modern evolution of traditional sub-surface data analysis.
For both touch and through coals, a training dataset was built containing two classes: coal boundaries and not coal. A 1D CNN was trained on this dataset, and then applied to holes from different benches to determine if coals were present.
Training data. A) positive through coal examples. B) positive touch coal examples. C) negative touch coal examples
Refining the results
Additional rules were applied to identify the exact boundary depth and eliminate some of the false picks. These included low probability, inconsistent with nearby holes, high SE min and a smaller than expected SE amplitude change. This refining process proved essential for practical application.
Test bench to demonstrate cleaning: A) SE, red = high energy and blue = low energy. B) Initial results, light blue = CNN identified through coal, dark blue = CNN touch coal, purple = manual top coal and red = manual lower coal. C) Cleaned results (same colours as B).
Results: touch coals
For the touch coals the CNN model had an accuracy of 86% when compared to the manually picked labels, with most of the errors being missed touch coals (13%). These missed coals most often occurred where the MWD did not continue far enough into the coal – the algorithm needs that characteristic energy drop signal to confirm coal presence.
A) Touch coal that was successfully identified. Red lines = manually identified coals and blue lines = CNN picks.
B) An example of where the CNN did not identify a touch coal as the MWD did not continue into the coal but stopped just short.
Results: through coals
The through coals are more difficult to evaluate. In addition to coal bands, there are other soft bands that can produce similar MWD signals and may be identified as coal by the CNN. These bands may be other rock units, thin coal bands that are not considered economic or coal seam splits where only the main split is logged. These can be eliminated using additional cleaning steps, including removal of thin bands.
A hole where the CNN model correctly identified two coal seams, but also identified an additional soft seam at the top of the hole. Red and green lines = manually identified coals, and blue and yellow lines = CNN picks.
Cleaning to remove other soft bands: A) SE, red = high energy and blue = low energy. B) Initial results, light blue = CNN identified through coal, dark blue = CNN touch coal, purple = manual top coal and red = manual lower coal. C) Cleaned results (same colours as B).
Validating against manual interpretation
The manual and CNN coal picks were then used to create gridded surfaces representing the top of the coal seams. The two surfaces were highly similar, indicating that the CNN method of identifying coal picks is suitable for use in this case. This is a strong validation, the automated method produces results comparable to those from experienced interpreters. It’s an example of how modern sub-surface data analysis has evolved to incorporate machine learning alongside traditional techniques.
Gridded surfaces for the top of the through coal (top) and touch coal (lower) based on coal picks. Light blue = CNN top coal, dark blue = CNN lower coal, purple = manual top coal and red = manual lower coal.
The bottom line
This automated approach extracts valuable coal seam location data from routine MWD measurements, no additional sensors or site delays. With 86% accuracy for touch coals and reliable surface mapping for through coals, the method provides real time geological intelligence that supports better blast design and resource tracking. While geological context and manual review remain important for complex cases, this tool demonstrates how machine learning can unlock information that’s already hiding in drilling data.
