Drill and blast is where the mining value chain starts, and getting it right has an outsized impact on everything downstream. Poor fragmentation means oversize material clogging crushers, excessive fines that create problems in processing, and wasted explosives. Good fragmentation means higher crusher throughput, lower energy costs in the grinding circuit, and better ore recovery. The difference between good and bad blast fragmentation can shift total processing costs by 10-15%.

The problem has always been feedback. A drill and blast engineer designs a pattern based on rock mass models, drills it, loads it, fires it — and then waits hours or even days to find out how well it worked. By the time fragmentation data comes back from the processing plant, conditions have changed and the next blast has already been designed. It’s a feedback loop that’s far too slow to drive meaningful optimisation.

Machine learning-based image analysis is finally closing that gap.

How It Works

Camera systems — mounted on excavator buckets, conveyor gantries, or haul truck tipping points — capture images of the muckpile and broken ore. ML algorithms then analyse those images to determine the particle size distribution (PSD) of the fragmented rock. The results are available within minutes of digging, rather than hours or days after the ore reaches the plant.

This isn’t new in concept. Image-based fragmentation analysis systems like Split-Desktop and WipFrag have been around since the 1990s. What’s changed is the accuracy and speed that modern deep learning models bring to the problem. The old systems struggled with fines (material under about 25mm), overlapping particles, and lighting variations. They required careful image setup and still produced results that many engineers viewed with scepticism.

Current ML models handle these challenges far better. Trained on hundreds of thousands of labelled images across different rock types, lighting conditions, and moisture levels, they can estimate PSD with accuracy that consistently matches — and in some conditions exceeds — manual sieving analysis. A peer-reviewed study in the International Journal of Mining Science and Technology found that well-trained ML fragmentation models achieved R-squared values above 0.92 compared to physical sieve analysis across multiple rock types.

Connecting the Blast to the Result

The real value isn’t the fragmentation measurement itself — it’s what you do with it. When drill and blast engineers can see the fragmentation result from this morning’s blast by lunchtime, they can adjust the afternoon’s loading design or tomorrow’s drill pattern. That rapid iteration is what drives improvement.

At a large open cut gold mine in the Eastern Goldfields, the drill and blast team installed cameras on their primary excavator and on the run-of-mine bin. They linked the fragmentation data back to their blast design software, creating a record that connects every blast pattern to its fragmentation outcome. After six months, they’d built a dataset that showed clear correlations between specific pattern modifications and fragmentation improvements for each geological domain on their site.

The engineers told me they’d reduced their average P80 (the size at which 80% of material passes) by 12% across the site, and they’d done it while reducing powder factor in some domains. In other words, they were getting better fragmentation with less explosives in areas where the data showed they’d been over-blasting.

Where AI Strategy Support Makes the Difference

The technology piece — cameras and ML models — is actually the simpler part of this puzzle. The harder part is integrating fragmentation data into the drill and blast workflow in a way that engineers actually use day-to-day.

I’ve seen operations where fragmentation cameras were installed, generated beautiful reports, and sat unused because the data wasn’t connected to anything the blast engineers worked with. The reports went to a shared drive. Nobody looked at them regularly. The feedback loop stayed open.

The operations that get this right invest in the data integration layer. Fragmentation results feed directly into blast design software. Automated reports compare predicted versus actual fragmentation for each blast. Dashboards show trends over time by geological domain, by blast engineer, by shift. AI strategy support from specialists like team400.ai can help mining operations build these integration frameworks so that the ML models don’t just generate data — they drive decisions.

Practical Challenges

Dust is the enemy. Camera lenses on mining equipment get filthy fast, and a dusty lens produces unreliable results. Air blast systems, regular cleaning schedules, and protective housings help, but sensor maintenance is an ongoing requirement. If the camera goes down for a shift, you lose that blast’s data entirely.

Wet conditions create another headache. Water on the muckpile surface changes the apparent colour and texture of particles, and surface water can cause reflections that confuse the algorithms. Most current systems handle moderate moisture reasonably well, but heavy rain or waterlogged muckpiles can produce unreliable readings. Some operations simply flag wet-condition data as lower confidence rather than excluding it entirely.

Calibration against physical sieving remains important. Even the best ML models can develop systematic biases over time as rock types change or cameras age. Running a manual sieve analysis once a month and comparing it to the ML output keeps the system honest.

The Downstream Effect

This is where the broader economic argument lives. Better blast fragmentation doesn’t just make the drill and blast team look good — it flows through the entire processing chain. A 10% reduction in top size reaching the primary crusher can increase crusher throughput by 5-8%. Less oversize means fewer crusher blockages and less unplanned downtime. Better-controlled fines generation means more predictable grinding performance.

One iron ore operation in the Pilbara quantified the total value chain impact of their fragmentation optimisation program at roughly AUD $3.50 per tonne mined. At 30 million tonnes per year, that’s $105 million in annual value — from better blast design informed by rapid ML feedback.

What Comes Next

The logical next step is connecting fragmentation analysis with downstream processing data in real time. If the plant control system knows that the next few hours of feed will have a coarser-than-average size distribution, it can preemptively adjust crusher and mill settings. A few operations are piloting this integrated approach, and the potential gains from coordinating blast design with plant optimisation are substantial.

For drill and blast engineers who’ve spent careers waiting days for fragmentation feedback, this technology is a genuine step change. The feedback loop is closing, and the engineers who learn to work with it are producing consistently better results.