Predicting Comminution Parameters (BWI, DWI, Ai) with Machine Learning

Key Takeaways

• Comminution parameters, namely BWI (Bond Work Index), DWI (Drop Weight Index) and the Abrasion Index, still start from physical rock-breakage tests. JKTech’s JK Drop Weight Test and JKSimMet remain the industry-standard equipment and simulation software.
• 2025 research validates cheaper proxies. Hyperspectral imaging (HSI) and Leeb rebound hardness, fed into ML models, act as low-cost stand-ins for full BWI testing.
• A 2025 deep neural network approach (published in Minerals Engineering) predicts DWI and BWI from Geopyörä breakage test data trained across about 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 and 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

To predict comminution parameters with AI, run a cheap proxy measurement (hyperspectral scan or Leeb hardness) on every sample and a full physical test on a calibration subset. Train a regression model on that subset. Then apply it to estimate BWI, DWI and Ai across the rest of the deposit.

What Are Comminution Parameters?

Comminution parameters describe how hard ore is to crush and grind. The Bond Work Index (BWI) gives the energy, in kWh per tonne, needed to grind ore in a ball mill. The Drop Weight Index (DWI) describes how rock breaks under impact, which matters for SAG mills and crushers. The Abrasion Index (Ai) measures how quickly ore wears steel liners, media and crusher parts.

Together they drive mill sizing, throughput forecasts and energy budgets. The problem is cost. A full JK Drop Weight Test or Bond Ball Mill test takes real lab time and a meaningful sample mass. As a result, most deposits only have testwork at a handful of locations, even though hardness can vary a lot between domains.

How Do I Predict Comminution Parameters With AI?

The workflow has three steps: measure cheap proxies everywhere, test a calibration subset, then train and apply the model.

Measure cheap proxies on every sample

Start with fast measurements you can afford on the full sample set. The active research direction uses two. First, hyperspectral imaging (HSI) of drill core or samples, often from hyperspectral point or scan systems. Second, Leeb rebound hardness, a handheld, near-instant hardness reading.

Run physical tests on a calibration subset

Next, run the expensive physical tests on a calibration subset. That subset must span your geological and alteration domains. JKTech’s JKDWT and JKSimMet remain the industry standard for running and using these tests, and that doesn’t change. What changes is how many physical tests you actually need.

Train a model to predict comminution parameters

Then train a machine learning regressor on the paired data. Research published in 2025 in Minerals Engineering and related venues uses deep neural networks and Random Forest. One notable DNN approach was trained on Geopyörä breakage test data across roughly 700 global ore samples. Its DWI and BWI predictions correlate well with the full physical test, without running that test on every sample.

Finally, apply the model to every sample you didn’t physically test. This is the same pattern behind wider geometallurgical proxy models.

Limits of Predicted Comminution Parameters

The Abrasion Index is the least mature of the three. Ai is measured with the Bond abrasion test, which weighs the loss from a steel paddle tumbled with ore. There’s no dedicated commercial AI product built specifically for it yet. Instead, it is treated as another output of the same grindability-proxy ML pipeline used for BWI and DWI. Therefore, treat any Ai prediction as more provisional than a BWI or DWI prediction from the same pipeline.

As with any proxy model, validate against a representative held-out set of physical tests. Do this 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.

Predicted values pay off downstream. They feed hardness and recovery models in the block model, and those in turn inform grinding circuit optimization at the plant.

FAQs

What are the main comminution parameters?

The main comminution parameters are the Bond Work Index (BWI), the Drop Weight Index (DWI) and the Abrasion Index (Ai). BWI covers grinding energy, DWI covers impact breakage and Ai covers wear on steel. Together they size mills and forecast throughput.

Can machine learning replace comminution testwork?

No. ML reduces how many full tests you need, but it still needs physical tests to learn from and to validate against. A calibration subset across every geological domain remains essential.

Which cheap measurements predict comminution parameters best?

Current research focuses on hyperspectral imaging and Leeb rebound hardness. Both are fast and cheap enough to run on every sample. Their predictive power depends on how well your calibration subset covers the deposit’s mineralogy.

How reliable are Abrasion Index predictions?

They are the least reliable of the three. The Abrasion Index has no dedicated AI approach and rides on models built for BWI and DWI. Treat predicted Ai values as provisional until you validate them with physical tests.

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. That structured research task is what 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.