Key Takeaways

  • BWI (Bond Work Index), DWI (Drop Weight Index), and the Abrasion Index all still start from physical rock-breakage tests, run through JKTech’s JK Drop Weight Test / JKSimMet — the industry-standard equipment and simulation software.
  • 2025 research validates cheaper proxies — hyperspectral imaging (HSI) and Leeb rebound hardness — predicted by ML models as low-cost stand-ins for full BWI testing.
  • A 2025 deep neural network approach (published in Minerals Engineering) predicts DWI/BWI from Geopyörä breakage test data trained across ~700 global ore samples, cutting lab turnaround time.
  • The Abrasion Index has no dedicated AI-native product yet — it rides on the same ML grindability-proxy research as BWI/DWI, making it the least mature of the three (Emerging).
  • None of this replaces physical testwork entirely — it reduces how many full tests you need across a deposit by predicting the rest from cheaper measurements.

TL;DR

Instead of running full Bond/drop-weight tests on every sample, run a cheap proxy measurement (hyperspectral scan or Leeb rebound hardness) on everything and a full physical test on a calibration subset, then train a regression model to predict BWI/DWI/Ai across the rest of the deposit.

How Do I Predict Comminution Parameters With AI?

Comminution characterization — how hard ore is to crush and grind — has always been bottlenecked by testwork cost: a full JK Drop Weight Test or Bond Ball Mill test takes real lab time and a meaningful sample mass, so most deposits only have testwork coverage at a handful of locations. JKTech’s JKDWT and JKSimMet remain the industry-standard equipment and simulation software for running and using these tests, and that doesn’t change — what’s changing is how many physical tests you actually need to run.

The active research direction (2025 papers in Minerals Engineering and related venues) trains machine learning regressors — deep neural networks, Random Forest — on a combination of cheap, fast measurements: hyperspectral imaging (HSI) of drill core or samples, and Leeb rebound hardness (a handheld, near-instant hardness reading). One notable DNN approach was trained on Geopyörä breakage test data across roughly 700 global ore samples and predicts DWI/BWI values that correlate well with the full physical test, without running the full test on every sample. The practical workflow: run your cheap proxy measurement (HSI scan or Leeb hardness) across your full sample set, run the expensive physical test on a calibration subset that spans your geological/alteration domains, train a regression model on that subset, then apply it to predict BWI/DWI across everything you didn’t physically test.

The Abrasion Index (AI test, measured via Cerchar Abrasivity Index rigs) is the least mature of the three — there’s no dedicated commercial AI product built specifically for it yet. It’s currently treated as another output of the same grindability-proxy ML pipeline used for BWI/DWI, rather than having its own validated approach, so treat any Ai prediction as more provisional than a BWI/DWI prediction from the same pipeline.

As with any proxy model, validate against a representative held-out set of physical tests before trusting predictions in domains you haven’t calibrated against — a model trained on one alteration type will not reliably extrapolate to a mineralogically distinct zone.

Try It With Geocluster

Building a defensible proxy model means tracking which samples have physical tests, which have proxy measurements, and how those relate across your geological domains — exactly the kind of structured research task Geocluster is built to handle. If you’re setting up a comminution proxy pipeline, Geocluster can help you organize the calibration and validation work behind it.