Proximate analysis-based prediction of gross calorific value of Indian coals using a Geochemical Feature Embedding Network (GFEN) and XGBoost regressor

Gross Calorific Value (GCV) is a key indicator of coal quality and a major determinant of coal grading and pricing. Measurement of GCV through bomb calorimetry is accurate but requires dedicated instrumentation and careful sample preparation. This study demonstrates that GCV of Indian coals can be predicted directly from proximate analysis using a hybrid Geochemical Feature Embedding Network (GFEN) and XGBoost regressor. Using only ash%, moisture%, volatile matter%, and fixed carbon% as inputs, the proposed GFEN–XGBoost framework achieves MAE values of 83.07–125.32, RMSE values of 111.66–175.24, and R2 values of 0.9753–0.9881 across four conditioning-state datasets derived from 219 samples from Central Indian coalfields and 331 samples from Southern Indian coalfields. The GFEN–XGBoost framework is the primary predictive model in this study, while representative baseline models are used only for comparative benchmarking. Its main contributions are the introduction of a domain-informed Serial feature to represent ash-linked quality regimes, the use of GFEN to learn compact latent representations from proximate variables, and the integration of these embeddings with XGBoost for final predic-tion. The model is intended as a rapid screening and decision-support tool that can reduce the number of samples requiring full calorimetric testing once proximate-analysis data are available.