Machine learning-driven strength prediction for solid waste-based backfill materials using infrared radiation indices

Abstract

The resilience of underground mining systems is deeply associated with the unconfined compressive strength (UCS) of backfill components, which is key to securing safety, operational performance, and ecological balance. This investigation presents a data-centric methodology that amalgamates machine learning (ML) techniques with infrared radiation (IR) indices, specifically Average Infrared Radiation Temperature (AIRT) and Variance of Infrared Radiation Temperature (VIRT), aimed at improving the precision of UCS predictions. A thorough dataset of 180 samples focused on backfill materials taken from solid waste underwent a careful analytical methodology using progressive machine learning models, including Light Gradient Boosting Machine (LightGBM), Artificial Neural Network (ANN), Random Forest Regression (RFR), and Least Squares Support Vector Machine (LSSVM), where LightGBM yielded stellar outcomes, reaching an R² score of 0.9276 and reflecting notably small error margins, as underlined by a mean absolute error (MAE) of 0.5618 and a root mean square error (RMSE) of 0.8803. The variance accounted for (VAF) was calculated at 0.9342, with the relative standard residual (RSR) recorded at 1.002. The prediction interval (PI) was defined as [2.0178, 5.9030], alongside a normal mean bias error (NMBE) of −0.0327. The employed methodology incorporated stringent experimental protocols, encompassing sample preparation, curing processes, UCS evaluations, and thermal imaging analyses to capture the dynamic responses elicited during loading. A user-centric graphical user interface (GUI) was engineered employing Tkinter, facilitating real-time UCS forecasts and expeditious decision-making within mining contexts. This study accentuates the transformative capacity of synergizing IR metrics with ML, presenting a replicable framework for integrating empirical testing with computational modeling. It emphasizes the necessity for larger, more heterogeneous datasets to bolster model robustness while fostering sustainable backfill design and environmentally conscientious mining practices.