Attention-based Deep Learning Prediction of Silica Concentrate for Quality Optimization in Mineral Processing

Mohammad, Fardeen (2025) Attention-based Deep Learning Prediction of Silica Concentrate for Quality Optimization in Mineral Processing. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (480kB) | Preview
Preview	PDF (Configuration Manual) Download (431kB) | Preview

Abstract

The prediction of silica and impurity concentration is essential for controlling mineral processing operations, where fluctuations in feed composition and process conditions significantly affect product quality and energy efficiency. Traditional laboratory-based measurement methods are accurate but slow, creating delays that hinder real-time decision-making and operational responsiveness. Although machine learning and deep learning models have been increasingly applied to address these challenges, many existing approaches rely on predefined feature sets or dimensionality reduction techniques, limiting their ability to capture high-order and dynamic feature interactions. Furthermore, conventional models often lack transparency and struggle to generalize across variable industrial conditions. This study introduces an attention-based AutoInt framework designed to overcome these limitations by automatically learning low- and high-order interactions among process variables using multi-head self-attention. The leveraging approach transforms each operational parameter such as reagent dosage, pulp density, air flows, and flotation column indicators into learnable embeddings, enabling adaptive modelling of complex industrial behaviour. In addition, the research a comprehensive pipeline that compares AutoInt with traditional machine learning models and a deep learning architecture using real iron ore flotation plant data. Experimental results show that AutoInt significantly outperforms all baseline models, achieving the lowest MAE value 0.3601, MSE value 0.2774, and RMSE value 0.5267, along with the highest R² score of 0.7777. These findings confirm AutoInt’s superior ability to capture nonlinear and high-order dependencies, making it more robust under fluctuating industrial conditions. The model’s attention mechanisms also provide, revealing which variables most strongly influence silica concentration.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Tomer, Vikas UNSPECIFIED
Subjects:	G Geography. Anthropology. Recreation > GE Environmental Sciences > Earth sciences Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	08 Sep 2026 10:44
Last Modified:	08 Sep 2026 10:44
URI:	https://norma.ncirl.ie/id/eprint/9887

Actions (login required)

View Item