Machine Learning (ML) and Artificial Intelligence (AI) are being applied to key metallurgical process units within the HAlMan project to improve process understanding, support operational decision-making, increase efficiency, and identify optimized operating conditions. The work combines experimental data, thermodynamic modelling, advanced machine learning techniques, and multi-objective optimization to address both pyrometallurgical and hydrometallurgical processes.
A central objective is to move beyond conventional process analysis and develop predictive tools capable of identifying relationships between operating conditions, material characteristics, process performance, and energy requirements. The resulting models can support researchers and industrial operators in evaluating different production scenarios and selecting operating conditions according to priorities such as throughput, energy consumption, and operating cost.
Optimizing hydrogen-based pre-reduction of manganese ores
A major area of development concerns the hydrogen-based pre-reduction of manganese ores. Replacing carbon with hydrogen offers an important pathway towards more sustainable ferromanganese production, as hydrogen reduction produces water vapor rather than the direct carbon dioxide emissions associated with conventional carbon-based reduction.
Optimizing this process is challenging because reduction behaviour depends on several interacting factors, including ore mineralogy, temperature, gas composition, and operating conditions. There is also an important trade-off between achieving rapid reduction, which increases productivity, and limiting energy consumption.
To address this challenge, a comprehensive modelling framework has been developed using thermogravimetric analysis (TGA) experiments covering four manganese ore types: Nchwaning, Comilog, Zambian, and UMK. The experimental database has grown substantially and now contains 77 TGA experiments, providing a stronger basis for developing models that can capture differences between ore feedstocks.
The modelling workflow consists of three closely connected stages: kinetic feature engineering, supervised machine learning, and multi-objective optimization.
In the first stage, experimental reduction curves are transformed into meaningful kinetic features using the Johnson-Mehl-Avrami-Kolmogorov (JMAK) framework. Particular attention has been given to data quality and artifact correction. For example, transient formation of water vapor during hydrogen reduction can appear as temporary weight increases in TGA measurements. Systematic correction procedures were therefore introduced before extracting kinetic parameters and training predictive models.
The resulting features are then used to train ML algorithms capable of predicting the time required to reach specific levels of ore reduction. Several machine learning approaches were evaluated using Leave-One-Out Cross-Validation. Among the tested methods, XGBoost and Gradient Boosting provided the strongest predictive performance, reaching R² values of approximately 0.81 for predicting the time to 50% conversion and 0.74 for 90% conversion.
From prediction to process optimization
The ML models are also used as the foundation for determining operating conditions that balance competing industrial priorities.
A multi-objective optimization approach based on the NSGA-II algorithm was applied to simultaneously consider reduction time and energy consumption. Rather than producing a single theoretical optimum, the optimization generated 526 Pareto-optimal operating conditions. Each represents a different compromise between faster processing and lower energy demand.
Knee-point analysis was subsequently used to identify a smaller group of particularly attractive solutions from the Pareto front. These solutions translate the modelling results into practical operating scenarios.
For example, the analysis identified conditions suited to energy-sensitive production, including Nchwaning ore at around 480°C, just below the experimentally covered temperature range, with an estimated energy requirement of 22.81 kJ. For production environments where throughput is the main priority, Zambian ore at 800°C can reach 50% reduction in approximately 1.14 minutes. A more balanced scenario for Zambian ore achieves 90% reduction at 550°C in approximately 3.75 minutes with an energy requirement of about 36.35 kJ.
These examples demonstrate how ML can move beyond prediction and become a practical decision-support mechanism for metallurgical operations.
Interactive decision support
To make the optimisation results accessible to process researchers and operators, prototype interactive web-based dashboards have been developed. These tools allow users to explore optimal operating conditions without having to interact directly with the underlying ML algorithms.
Users can investigate how recommended conditions change according to ore type, reduction target, energy requirements, and operational priorities. Economic parameters such as electricity and labour costs can also be considered, allowing technically optimal solutions to be assessed from an operational and economic perspective.
This creates a bridge between advanced data-driven modelling and practical process decision-making. Instead of providing a fixed operating recommendation, the tools enable users to explore alternative scenarios and understand the trade-offs associated with each choice.
Machine learning for hydrometallurgical processing
ML modelling has also been developed for the hydrometallurgical route, focusing on the carbonation step used to recover alumina from HAlMan slags. Carbon dioxide is bubbled through the sodium aluminate solution obtained by leaching the slag, neutralizing the free caustic and precipitating aluminium hydroxide.
The central operating question is when to stop adding CO₂. Carbonating further recovers more product, but beyond a certain point the chemistry turns and the desired aluminium hydroxide begins to convert into dawsonite, a sodium-aluminium carbonate that locks up both the aluminium and the sodium. The favourable operating window therefore has a definite end, and the most valuable place to operate is close to it.
The difficulty is that the composition of the solution, which is what defines where that window ends, cannot be measured continuously inside an operating reactor, whereas pH, electrical conductivity, and density can be measured cheaply and without interrupting the process.
The work therefore developed a soft sensor: a model that infers the concentrations of the four main dissolved species (NaOH, NaHCO₃, Na₂CO₃, and NaAl(OH)₄) in real time from those three routine signals. Thermodynamic simulation was used to build a dataset of 16,582 solution compositions, which both trains the model and defines the validated composition envelope within which its predictions are considered reliable. Gaussian Process and ensemble approaches were compared, and a Gaussian Process was selected, reproducing held-out compositions from this dataset with an overall R² of 0.9974.
Accuracy alone did not determine that choice. Because the difficulty of the inversion varies by nearly three orders of magnitude across the operating region, a single global accuracy figure would misrepresent the real reliability, so the selected model reports its own confidence at every individual prediction.
Anchoring the models in measurement and chemistry
The link between sensor readings and solution composition is the foundation of the whole chain, because every later step inherits the quality of that calibration. It was therefore anchored in physical measurement: 40 solutions of known composition were prepared in the laboratory, 37 of which carry the paired sensor readings used for calibration. pH and density calibrate with R² of 0.98 and 0.94 respectively, while conductivity, the most non-linear of the three channels, was improved from 0.86 to 0.97 by combining a physically based description of ionic strength with a machine-learned residual correction. It is this calibration, rather than the inference step, that sets the measurement-anchored accuracy limit of the complete chain.
A second safeguard comes from chemistry rather than from data. Sodium hydroxide and sodium bicarbonate cannot coexist in solution, yet an unconstrained statistical model has no way of knowing this and can return compositions that cannot exist. Imposing the rule as a mass-conserving correction on the model output eliminates every such prediction, at a negligible cost in accuracy, with no tuning parameter and independently of which model is deployed. For an exact constraint of this kind, applying it to the model output proved simpler and more dependable than expressing it as a penalty during training.
From inference to operating decisions
Run retrospectively against data from laboratory carbonation runs, the complete chain executes end to end and indicates the approaching transition well before it is reached, a margin comfortably compatible with the operating decision. A dedicated validation campaign is the next step towards quantifying its accuracy under process conditions.
The same framework also generates operating recommendations. A total of 4,000 candidate operating conditions were screened against three competing objectives: maximizing aluminate yield, minimizing the risk of overshooting the window, and limiting carbonate accumulation. This produced 21 Pareto-optimal conditions, every one verified to lie inside the validated composition envelope rather than resting on extrapolation. Among them, a balanced regime offers around 28% higher expected aluminate yield than the most conservative one, making the trade-off explicit rather than a matter of judgement.
The chemistry behind the indicator was checked independently of the statistical models. The same reaction system was implemented twice, once as a kinetic description and once as a rigorous equilibrium calculation, sharing neither code nor thermodynamic database. The two implementations agree with each other and follow the observed course of the reaction. Agreement between two independent implementations gives confidence in the modelled chemistry itself, not only in the statistics. All these results come from a single reproducible pipeline in which every reported value is re-derived automatically from the underlying data.
Combining machine learning with physical knowledge
Alongside conventional data-driven ML methods, Physics-Informed Neural Networks (PINNs) are being investigated for reduction processes. Unlike purely data-driven models, PINNs incorporate physical knowledge such as fundamental mass-balance constraints directly into the modelling framework.
This approach can be particularly valuable in metallurgical applications where experimental datasets are limited but the governing physical and chemical principles are well understood. Combining these constraints with neural networks can improve physical consistency and potentially increase model reliability outside the exact conditions represented in the training data.
The modelling activities are also connected with the exploration of innovative process concepts, including Flash Smelting for manganese ores. This concept aims to integrate gaseous pre-reduction and metallothermic reduction within a single reactor, opening opportunities for advanced modelling and optimization of highly integrated production routes.
Towards data-driven and sustainable metallurgical processes
The ML modelling work demonstrates how experimental research, thermodynamic simulation, physical process knowledge, machine learning, and optimization can be combined into an integrated framework for metallurgical process development.
For hydrogen-based manganese ore reduction, the approach has progressed from kinetic modelling to validated ML prediction, multi-objective optimization, and interactive decision-support tools. For hydrometallurgical processing, computed thermodynamic data have enabled a real-time composition soft sensor, laboratory-prepared solutions have anchored its calibration in physical measurement, and laboratory carbonation runs have demonstrated the complete chain from sensor signal to stage indication, with independent mechanistic modelling supporting the underlying chemistry. The modelling has also defined its own next step: a targeted experimental campaign to measure how quickly dawsonite actually forms, and a computational extension of the validated composition envelope.
Together, these developments illustrate a broader transition towards intelligent metallurgical processing. Rather than using ML as an isolated analytical technique, the approach integrates data-driven models with physical understanding and process optimization to support lower energy consumption, improved productivity, more informed operational decisions, and ultimately more sustainable production.