Sustainable Data
Published in

Sustainable Data

Machine learning in the mining industry — a case study

Recently we attended the Unearthed Data Science event in Melbourne. A gold mining company — Newcrest Mining — provided operating data for a number of its plants, with the aim that some of the teams attending could provide useful solutions grounded in Data Science.

One particular system caught our eye — the autoclaves. What’s an autoclave? Take a look at this:

Cutaway image of gold ore processing autoclave (source: Unearthed Challenge-Newcrest Mining)

Newcrest extracts gold from ore at their Lihir Gold operation in Papua New Guinea. This ore is rich is sulphide minerals (sulfide if you’re American) such as iron pyrite (FeS2) (aka “Fool’s Gold”). Yes, there’s real gold among the fool’s stuff.

Sulphides inhibit the processing techniques used to extract gold from ores, so it’s ideal if you can get rid of them.

That’s where the autoclaves come in. An autoclave is a type of chemical reactor that provides the right physical and chemical conditions for certain chemical reactions to occur. From the image above, you can see it is a long cylindrical vessel divided into sections by internal walls called baffles. Each section has its own mixer to ensure good contact for the chemical reactions, and an entry point at the bottom to let in oxygen gas and steam.

In Newcrest’s case, they use autoclaves to oxidise the sulphur minerals of the ore using a combination of heat, pressure and oxygen. The ore is crushed and mixed with acids to form a slurry. It is this slurry that mixes with oxygen gas inside the autoclave.

The sulphide minerals chemically react with oxygen to form other compounds that can be easily removed. The remaining solids are much richer in gold than the raw ore, enabling easier leaching of gold downstream of the autoclaves.

Using air directly as an oxygen source isn’t suitable — air is mostly nitrogen which is inert and would slow down the reaction of oxygen and sulphides. This would require a much larger autoclave to do the same job.

To keep autoclave sizes and capital costs down, Newcrest’s autoclaves instead rely on purified oxygen, provided by an air separation unit (ASU). Oxygen is injected at the bottom of the autoclave into each chamber divided by baffles (internal walls). The mixers ensure good contact between the oxygen bubbles and the ore slurry.

Air separation units are heavy energy consumers. Although they obtain oxygen from the air, which is free, the use of electricity to drive the ASU means that purified oxygen is quite expensive in energy terms, and as a result is linked to significant greenhouse gas emissions and operating costs as well.

As this is at a remote site, fuel supplies for electricity generation are quite expensive, so anything that can reduce energy demand — such as reducing autoclave oxygen requirements — would be of economic and environmental value.

Basic chemistry determines a minimum amount of oxygen required to oxidise the sulphides. In practice though, a so-called excess of oxygen is required to ensure the reactions are completed. This presents an opportunity to minimise the excess oxygen and therefore reduce ASU electricity consumption — saving money and reducing GHG emissions.

Machine learning model of oxygen consumption

In this instance, we wanted to model the total flow of oxygen gas to one of the autoclaves at Lihir.

By using the data set provided (operating data for every “tag” (measurement point) every 5 minutes for one calendar year) we developed a neural network-based machine learning model based on plant operating variables around the autoclave. The variables included temperature measurements, ore flow rate, and operating pressure.

The chart below shows the actual oxygen consumption for one of the Lihir autoclaves and the predicted oxygen consumption from the machine learning model.

To protect Newcrest’s production data, we have standardised the oxygen flowrates. A value of one (1) represents the maximum oxgyen flow rate for the year, and zero represents minimum flow.

Autoclave standardised oxygen flow — Red=actual flow; Blue=machine learning model result

The chart above shows that our machine learning model is predicting oxygen flow (blue) as a function of many other operating variables like temperatures, pressures and (non-oxygen) flows. It very closely matches the real measured oxygen flow (red).

The correlation coefficient between actual and modelled oxygen flow is over 0.99 — a very close modelling result.

It’s also useful to examine the times it doesn’t line up so well — it’s mainly when there are rapid changes in the operation of the autoclave. Because there is some dynamic (time dependent) behaviour, which a machine learning model will struggle to capture, the model will be at its best when the autoclave is running at steady state — that is, when all its operating variables are steady with time. This does not invalidate the model but it does highlight that you have to use machine learning models with care.

Why this model is useful

For starters, we can use our new model to predict what oxygen consumption will be for many different sets of operating conditions. To build a model like this from first principles (using engineering techniques like mass and energy balances) would take much longer and be significantly more challenging to get good agreement with real operating data.

We can repeat the machine learning process for any other variables we’d like to be able to predict — electricity consumption, waste flow, water consumption, emissions — they’re all good candidates for this modelling.

Since the model incorporates many operating variables, we can apply optimisation techiques on the model to see what set of operating conditions can minimise excess oxygen use per tonne of ore processed.

This enables tuning of the operation of the autoclave to minimise oxygen consumption, helping to save fuel costs and emissions for the site. All without any need to spend on new capital equipment — just through better operation of the equipment already on site.

Broader implications

The above example is fairly simple. The broader implications of machine learning are much more exciting. For example, we could build a machine learning model to predict total energy use for all the autoclaves at the plant. This model could be optimised to find the right combination of temperatures, flows, pressures and other parameters that minimise all oxygen use for the site — not just for a single autoclave.

We’re looking forward to bringing more modelling and efficiency benefits of machine learning to industry. Get in touch today — we’d love to help you improve your energy consumption and reduce your emissions and waste.

David Kearns is cofounder of Sustainable Data. We use the power of Big Data and Machine Learning to help industrial businesses save energy, reduce emissions and save money.

Sign up to our newsletter to keep up to date.



Machine learning and big data techniques to increase industrial energy efficiency, reduce emissions and improve productivity

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store
David T. Kearns PhD

#cleantech #carboncapture #ccs #ccus #energy #industrialtransformation #machinelearning #energyefficiency #emissions #carbon #sustainability.