Dynamic process optimization using data-driven surrogate models: Application to mineral processing

The Business Finland AIMODE project addresses critical challenges in the mining and mineral processing industry—such as raw material variability, energy efficiency, and sustainability—through the application of artificial intelligence and machine learning (AI/ML). Within this project, we have developed several time-series surrogate models, including LSTM, GRU, and CNN architectures, to forecast flotation process performance 30 minutes ahead using both a dynamic process simulator (digital twin) and plant operational data. Anticipating process behavior enables proactive decision-making and more effective operational planning.

In this study, we employ a multistep-ahead LSTM surrogate model to identify future control sequences that improve flotation performance. The model integrates historical process inputs, control variables, and outputs with candidate future controls to efficiently predict multistep outcomes. By simulating alternative control trajectories, the surrogate model facilitates the search for control strategies that maximize predicted average performance, which can then be recommended to operators via the digital twin. Given the large search space—12 level- and air-flow control variables across six future time steps—exhaustive search is computationally infeasible, even when using the surrogate model. To overcome this challenge, we evaluated three optimization strategies: random search, Bayesian optimization with Gaussian processes, and Bayesian optimization with a tree-structured Parzen estimator (TPE).

Our results show that Bayesian optimization with TPE, combined with accurate and computationally efficient surrogate models, can rapidly identify promising control strategies that enhance flotation process performance. These optimized control sequences should be validated in the digital twin and, ultimately, in the operational environment before deployment.