Raman Spectroscopy and Generative Variational Autoencoders for Metallurgical Coke Quality Prediction

ABSTRACT Metallurgical coke is a critical raw material in blast furnace ironmaking, where its mechanical strength and reactivity directly govern process stability, energy consumption, and productivity. These properties are conventionally assessed through the coke strength after reaction (CSR) and drum index (DI) tests, which are destructive, time‐consuming, and incompatible with real‐time process control. Raman spectroscopy offers a fast, nondestructive alternative, yet the nonlinear relationship between spectra and these quality parameters, combined with the scarcity of labeled industrial samples, limits the development of robust predictive models. This work proposes an integrated methodology combining a standardized sample preparation protocol, crushed coke homogenized in distilled water, with spectral preprocessing, synergy vector selection, and synthetic spectrum generation via conditional variational autoencoders (CVAE) and supervised cariational autoencoders (SVAE), evaluated across datasets of 77, 146, and 180 real samples. The CVAE reproduced spectral morphology with reconstruction RMSE below 0.021, and synergy vector selection was the intervention of greatest impact, reducing multicollinearity and stabilizing the latent space. Synthetic augmentation proved beneficial only above a minimum real‐sample threshold, below which it degraded all regressors. PLS achieved the most consistent CSR performance, with RMSE of 1.43–1.51, whereas kNN with the synergy vector achieved the best DI performance, with RMSE of 0.41–0.51, and the narrow dynamic range of DI remained the main prediction challenge. The findings of this study can help advance Raman‐based soft sensors for steelmaking quality control, reducing dependence on normative assays and supporting more responsive, data‐driven blast furnace operations.