A multi-model approach for estimation of ash yield in coal using Fourier transform infrared spectroscopy

The ash yield resulting from the alteration of inorganic elements during the processes of combustion and gasification of coal stands as a crucial quality indicator for coal. Ash yield, along with calorific value, determines the commercial rating, ranking, and industrial usage of coal. Traditional methods of determining the ash yield in coal as per proximate analysis protocols are tedious and time-consuming as they involve the combustion of coal samples. A novel approach that uses mid-infrared Fourier Transform Infrared spectroscopy (FTIR) (optical technique) data in the range of 1450–350 cm-1 to identify spectrally sensitive zones (fourteen selective absorption bands) and to predict the ash yield in coal samples is presented. Multiple algorithms, including piecewise linear regression (PLR), artificial neural networks (ANN), partial least squares regression (PLSR), support vector regression (SVR), and random forest (RF), were utilized to predict the ash yield in coal. The present study suggests a multi-model estimation (MME) approach, using the average of the best three models (PLR, PLSR, and ANN) to achieve greater accuracy and robustness. This method outperforms individual models with a coefficient of determination – R-squared (R2) of 0.883, Root Mean Square Error (RMSE) of 3.059 wt%, RMSE in percentage (RMSE%) of 30.080, Mean Bias Error in percentage (MBE%) of 3.694, and Mean Absolute Error (MAE) of 2.249 wt%. The two-tailed t-test and F-test for mean and variance (99% Confidence Interval, CI) show no significant difference between the proximate analysis-derived ash yield and the multi-model estimated ash yield using FTIR data. FTIR spectroscopy data can accurately predict the ash yield in coal and perform well for coal samples from Johilla Coalfield, Umaria, Madhya Pradesh. The present model using FTIR analysis is a potential industrial tool for the quick determination of ash yield in coal and can be further improved by including data from other basins worldwide. © The Author(s) 2025.