Conference Agenda

The Gardanne plant, cradle of the Bayer process, transformed itself in 2022 by suspending digestion of bauxite after nearly 130 years of operation, in response to local environmental constraints. As the focus continues to shift towards specialty alumina, controlling precursor quality remains a major challenge, leading the plant to maintain a variation of the Bayer process as hydrate dissolution/re-precipitation process to control quality, known as the UODP (Unit Operation of Dissolution and Precipitation) process. On the other hand, the variety and specificity of our markets, which can lead to significant upsets in volume throughout the year, demand great flexibility in terms of production rates and quality control.

In the second half of 2023, a sharp drop in demand forced a drastic reduction in production, and whereas the high price of energy, ALTEO Gardanne decided to stop production of the UODP process for several weeks on two occasions to limit stocks while minimizing energy consumption. This paper reviews the preparation and anticipation of risks, the operating conditions of these shutdowns and their consequences on the quality and the operations, which were generally very well controlled. These two shutdowns demonstrate that, under certain conditions, a Bayer-type process can be shut down without any major risk to hydrate quality.

This paper presents a pioneering approach to optimizing the agglomeration process at Hydro's Alunorte alumina refinery through the implementation of a digital twin. This advanced system leverages first principles and data science techniques to provide real-time forecasts and recommendations for critical quality variables, enabling a shift from reactive to proactive process management. The core advancements of this project include the development of precise predictive models, a what-if analysis tool for scenario simulation, and a recommendation tool for operational adjustments. The anticipated benefits of this technology include enhanced product quality, reduced variability, and increased operational efficiency, setting new benchmarks in the alumina refining industry.

UC RUSAL is currently testing a system for optimized precipitation control, which enables stabilization of gibbsite particle size distribution and improved liquor productivity. The current version features a predictive mathematical model represented by an artificial neural network. The neural network has been trained using the historical data and functions well under normal conditions; however, optimizer application requires better adaptability from the model while maintaining the same accuracy. A population balance model (PBM) of gibbsite mass crystallization allows for the development of an adaptable model with strict dependencies between the process parameters and explainable predictions, but this model works slowly and requires manual adjusting. On the other hand, a neural network model performs almost immediate predictions but the results are not always explainable. A surrogate model can be a compromise solution between these options. A PBE-based prototype surrogate model of gibbsite mass crystallization has been developed in similar way as a physics-informed neural network. Laboratory test results show that the model can be used to describe agglomeration and growth of gibbsite crystals in pregnant liquor.

Alumina extraction from boehmite in bauxite, such as in Central India, poses a unique challenge due to its inherently high energy requirements. Global refineries employing such energy-intensive processes must have an acute focus on energy efficiency to ensure competitive production costs. A pivotal efficiency metric is precipitation productivity or yield, indicative of refinery operational efficiency. Traditionally, maintaining optimum yield involves manual analysis of process sample results, where technical analysts meticulously correlate deviations in process conditions with lab values to identify influential factors. This approach is sensitive and heavily relies on human experience and expertise, underscoring the necessity for automated and intelligent solutions.

In response to this challenge, this article presents a pioneering endeavor to develop a mathematical model utilizing machine learning (ML) algorithms, coupled with Bayer process principles and noise factor considerations, to optimize yield in alumina precipitation. Notably, the model partitions the precipitation process into three segments: agglomeration, new growth, and existing growth. This facilitates a granular understanding and targeted intervention. Crucially, our model integrates constraints within each segment, ensuring that the quality of the resultant alumina remains uncompromised while optimizing yield. A transformative shift towards data-driven precipitation control is anticipated by leveraging the power of ML algorithms to optimize productivity while consistently delivering quality alumina. This intrinsic balance between productivity and product quality constitutes a hallmark of our approach, which not only enhances operational excellence but also underscores the disruptive potential of AI-driven solutions in alumina refining.