SAMPLING, ANALYSIS & LAB ASSAY
An innovation inflection point
From drill core to assay, the application of innovation in the sampling, analysis and laboratory assay parts of the exploration process will ultimately require an integrated mindset
Jason Nitz has held a range of senior leadership positions in operations, technical services and technology over the past 21 years, and he now sees an exciting opportunity to leverage artificial intelligence( AI) to help optimise operations, recently singling out drill core analysis as an area ripe to take advantage of AI-based algorithms.
“ The geology industry sits on a paradox,” he said in a recent LinkedIn post.“ Drill cores are among the most valuable and frequently collected datasets in mineral exploration – some mines generate hundreds of kilometres annually. Yet, for decades, extracting meaningful information from core photographs has remained a labour-intensive, manual process. A single geologist can spend weeks analysing what a trained machine-learning model can process in hours.”
This is changing rapidly, he says, explaining that ML4DrillCore systems represent a fundamental shift in how the mining industry interprets geological data.
He explained:“ By combining computer vision and deep learning, these [ ML4DrillCore ] systems can analyse drill core photographs with speed and consistency that fundamentally redefine what geologists should spend their time doing.”
ML4DrillCore, born out of a collaboration between Luleå University of Technology and a technology vendor, uses convolutional neural networks – pattern-recognition models trained on thousands of annotated core photographs – to automatically identify critical geological features like lithology, structures and faults, ore-bearing indicators
72 and rock quality metrics.
“ The models learn visual patterns – colour, texture, grain structure, veining – with remarkable precision,” Nitz says.“ In peer-reviewed studies, accuracy for lithology classification ranges from 90-98 %, with some systems achieving a less than 3 % error in rock quality estimation.”
It is speed where the business case becomes compelling, according to Nitz.
A recent implementation at a major gold mine in Western Australia processed 85,000 m of drill core imagery using ML4DrillCore in hours. This is a task that would have required months of manual logging, he says.
At another site, the time from raw imagery to complete geological analysis dropped from 2-6 weeks to 24 hours for chip tray analysis alone.
“ To put this in perspective: if a single geologist can carefully log 100 – 200 m of core per day( accounting for the interpretive work required), processing 85,000 m represents roughly 500-plus geologist-days of work,” he said.“ ML4DrillCore compressed this into hours.”
This speed is meaningless unless it allows geologists to conduct more strategic work as a result, which is where the discussion’ s narrative can take the form of“ AI replacing geologists”, instead of“ AI handling the routine work so geologists can do their job”.
Tasks such as photographing core boxes, visually inspecting each segment, classifying rock types based on appearance, recording fracture patterns and documenting alteration zones represent repetitive tasks. Once a system is trained, a machine-learning model
Larvotto Resources recently became the first Australian mining company to install an Elemission ECORE rapid core scanner at its Hillgrove antimony-gold project
executes them with mechanical consistency, free from fatigue or interpretive bias, Nitz says.
This creates space for higher-order geological work such as anomaly investigation, orebody delineation, resource estimation refinement, or strategic exploration decision making.
“ These questions require judgement, domain expertise and integration of multiple data streams – exactly what experienced geologists are trained to do,” Nitz says.“ ML4DrillCore removes the bottleneck preventing them from reaching this work.”
Objectivity is another benefit with this type of AI-based logging.
Nitz explains:“ Manual core logging is inherently subjective. Different geologists may interpret the same fracture pattern differently. The same geologist may log more conservatively on day five than day one. Lithological boundaries that seem obvious in one light can be ambiguous in another.”
A trained ML4DrillCore system, Nitz says, applies the same rules to every image, every time. Over thousands of core metres, this eliminates the rework and reinterpretation cycles required in manual datasets. It also creates a digital audit trail where teams can see exactly what the model detected and why, enabling geologists to override or refine predictions confidently.
Both of these benefits require investment and discipline to be leveraged over the long term.
ML4DrillCore systems require highquality core photography, with standardised lighting, consistent scale markers, depth referencing, etc. The machine-learning platforms also need“ labelled training data” that consists of( ideally) thousands of annotated core images from a specific geological environment. Geological validation is also another‘ must’: models learn patterns, but a geologist must
International Mining | FEBRUARY 2026