Biochar Stability Revealed by FTIR and Machine Learning

Abstract

Biochar is a carbon-rich and environmentally recalcitrant material, with strong potential for climate change mitigation. There is a need for rapid and accessible estimations of biochar stability, the resistance to biotic and abiotic degradation in soil. This study builds on previous work by integrating Fourier-transform infrared spectroscopy (FTIR) data with predictive modeling to estimate standard stability indicators: H:C and O:C molar ratios. Lignocellulosic feedstocks were pyrolyzed at highest treatment temperatures (HTT) ranging from 150-700 °C, and all samples achieved H:C < 0.7 and O:C < 0.4 at HTT of 400 °C and above. Several statistical and machine learning models were developed using FTIR spectra. The random forest (RF) models, which incorporated full data preprocessing, yielded the highest accuracy (R ² = 0.96 for both ratios) when tested on an unseen feedstock. Variable importance analysis identified spectral regions linked to aromaticity and inversely correlated to C-O stretches in cellulose and lignin as key predictors. The findings of this study verify that FTIR data can serve as a rapid and accurate tool for estimating biochar stability.

Keywords: H:C; O:C; Random Forest; grass; infrared spectroscopy; modeling; molar ratios; wood.

1

Van Krevelen diagram of both H:C and O:C molar ratios of biomass and biochars produced from the various feedstocks over a range of pyrolysis HTT. The black rectangle represents the ratios determined by EBC and IBI to certify biochar as suitable for soil amendment. ,

2

ATR-FTIR spectra of all biochars and starting materials; each spectrum is an average of triplicates for illustrative simplicity and is displayed on a common, offset y-axis. HTT is denoted by color, and common peak assignments are noted above their respective peaks.

3

Model comparisons of R ² and RMSE for both predicted H:C and O:C ratios on training data. The dot signifies the mean of the cross-validated resamples, and the error bars display the standard deviation. Preprocessing steps are abbreviated to N (normalized), S (scaled), NS (normalized and scaled), and NSP (normalized, scaled, and PCA).

4

Model predictions of H:C and O:C molar ratios vs actual values on unseen test data, which comprises a new feedstock (BS) and temperature treatments. Preprocessing steps are abbreviated to N (normalized), S (scaled), NS (normalized and scaled), and NSP (normalized, scaled, and PCA).

5

Exploration of PCA used in the preprocessing steps of model training. A) Scree plot showing variance explained of each individual PC, B) Cumulative variance of the principal components, C) Loadings of each wavenumber in the fingerprint region in the first 4 PCs, D) Variable importance of each PC for both the H:C and O:C model predictions as measured by increase in node purity, arranged in order of descending importance.