Quality prediction of iron ore pellets based on process data and a metallurgy-oriented large language model

Abstract

Compressive strength is a key quality index of iron ore pellets, reflecting the combined effects of pellet integrity, material characteristics, thermal history, and operating conditions during grate–kiln pelletizing. Reliable early prediction of pellet compressive strength is therefore essential for burden-quality control and stable downstream ironmaking. A physically constrained calibration framework, termed Hierarchical Physically Constrained Large Language Model (HyPC-LLM), is developed for pellet compressive strength prediction using real industrial pelletizing data. HyPC-LLM integrates multi-criteria feature selection, a hybrid regression pool (HRP), and structured LoRA calibration. A numerical-to-language bridge encodes branch predictions, recent target history, and process summaries as structured context, while shrinkage and ordered interval projections regulate the calibrated output. On 1587 chronological test records, HyPC-LLM achieves an MAE of 7.8812, an RMSE of 13.5355, and an R 2 of 0.8747, reducing MAE by 3.54% relative to HRP. Compared with an MLP residual calibrator, HyPC-LLM reduces MAE by 1.30%, although the temporal-block confidence intervals for this difference include zero. Ablation results indicate that the main performance gain is associated with LoRA-adapted structured calibration, whereas prompt guidance and empirical clipping have smaller aggregate effects. These results indicate the potential of integrating numerical process prediction with context-aware LLM calibration for pellet compressive strength prediction.