Development of an Artificial Intelligence System for Optimizing the High-Temperature Metallurgical Properties of Coke

Abstract

This paper proposes an artificial-intelligence-based decision-support approach for selecting coal-blend grade compositions used in metallurgical coke production. The method combines a previously developed artificial neural network, which predicts the coke reactivity index (CRI) from the percentage shares of ten coal grades, with a genetic algorithm that solves the inverse composition-property problem. Candidate blends are generated within specified technological and model-validity bounds, normalized to preserve the total charge balance of 100%, and improved by elitist selection, single-point crossover, additive mutation, clipping, and repeated normalization. The target CRI value was set to 23.0%. Computational experiments showed that the optimized blends decreased the predicted CRI from the baseline value of 33.50% to 26.44-29.53% and increased the calculated coke strength after reaction (CSR) from 44.90% to 51.65-56.90%. Three of the four independent runs converged to different feasible compositions with predicted CRI values that cannot be reliably distinguished within the predictive resolution of the model, indicating reproducible convergence to the same quality region rather than an accidental single solution. For a representative coke rate of 500 kg per ton of hot metal, the numerically best candidate corresponds to an estimated reduction in specific coke consumption of 45.0 kg/t, within a range of 30.0-60.0 kg/t reflecting the reported variation in the CSR sensitivity coefficient. The results demonstrate that neural prediction and evolutionary search can be combined to solve the inverse coal-blend design problem directly in the space of controllable coal-grade proportions and provide a practical tool for preliminary optimization before laboratory or pilot-oven validation.