91�����

When we last checked in with , it was just kicking off and researchers from FCAI, Metso, and other partners were considering how to infuse artificial intelligence into the mining industry. Now, a few years into the project, results are starting to emerge. Developing data visualization and machine learning methods for optimization and forecasting at different points in the mining pipeline has been challenging, say the researchers, as they aim to be at the forefront of applying cutting-edge research to this complex industry.

Boosting recovery of minerals from mined ore was one of the project’s initial aims. Mineral processing involves flotation, explains doctoral researcher Sahel Iqbal. Ore is crushed, ground and passed into tanks with water, air and reagents to separate minerals from slurry. These flotation tanks have controls to set the level of froth and rate of air that go in, while lots of data comes out that allows the state of the process to be minutely tracked. “With these data, if we know what is going on now, we can predict how much gold, for example, we will get out,” says Iqbal. With fellow doctoral researcher Mahdi Nasiri Abarbekouh, they have managed to significantly improve these predictions with a neural network-based surrogate model of the process.

Metso uses an existing physics-based simulator of processes like flotation, says Cesar Araujo, who specializes in simulation and process modeling at the industrial giant. “It shows how process variables interact and how they impact performance, but it’s computationally intensive. We wanted a surrogate model that is accurate and fast”, which is where the expertise of members of FCAI professor Simo Särkkä’s research group, like Iqbal and Nasiri, was able to offer a solution.

Applying physics-informed machine learning to the mining industry is a new inroad, explains Nasiri, since froth flotation demonstrates complex dynamics affected by many variables and interactions that aren’t even completely known. “Traditional models for data science often don’t take into account the physics of a problem, which might be fine for some applications, but in process modeling we can’t accept large errors,” adds Araujo. “Other models look at data without giving meaning to it, but here we are trying to give meaning to the data.” The inputs and outputs of minerals flotation must respect mass and energy balance, for example, which are not trivial and are an inherent part of the Metso simulation platform.

A faster simulator with an understanding of physics and accurate input-output predictions is one thing, but hooking up the surrogate to live data is a final goal before the AIMODE project comes to an end in 2025. “Right now, all the training and predictions happen with history data collected from the case plant or synthetic data generated with the simulator, but we’re just few steps away from deploying models to the real digital twin platform,” says senior scientist and project manager Joonas Linnosmaa of project partner VTT. “To be able to optimize with a digital twin, we first need to be able to forecast, accurately and fast, but multivariate, multi-step forecasting of a whole process is not simple. I’m happy with the results, and the cooperation with Metso and Aalto to solve problems together. Visualizing the forecast, especially the uncertainties, is an important step.” 

For data visualization systems manufacturer LightningChart, participating in AIMODE was a chance to advance data visualization technology. “The charting tools we developed in this collaboration and the user insights evolved into a commercial product, a high-performance interactive dataviz add-on for Python,” says LightningChart CEO Pasi Tuomainen.

Read more in these papers:

Above image: Data provided by the ‘Uncertainty_Visualization_dataset’ project, showing time-lagged CNN prediction errors (0 to 6 lead times). Visualized with LightningChart® Python. Code by Soroush Sohrabian and Elmeri Keinänen, 2024.