Non-destructive prediction of carbonization indices in biochar derived from underutilized forest biomass using ATR-IR chemometric modeling

Abstract

Biochar has emerged as a promising strategy for carbon sequestration in the context of climate change and carbon neutrality goals. Among various feedstocks, underutilized forest biomass (UFB) holds significant potential for conversion into high-value carbon materials. However, the heterogeneity of UFB and the high cost of conventional analyses highlight the need for rapid prediction techniques for key carbon indicators, such as carbon content, atomic oxygen-to-carbon ratio, and atomic hydrogen-to-carbon ratio. This study proposes a chemometric model that non-destructively predicts the carbonization characteristics of biochar using attenuated total reflectance infrared (ATR-IR) spectroscopy combined with partial least squares regression (PLSR). Twenty biochar samples were produced from UFB at carbonization temperatures of 200 °C, 300 °C, and 400 °C. The ATR-IR spectra were preprocessed using normalization and second-derivative transformation before being used to construct the predictive models. The optimized PLSR models, which were validated through cross-validation and outlier removal, achieved high prediction accuracy for all three carbon indices (R² > 0.94). Variable importance in projection (VIP) analysis further identified the key spectral regions contributing to the model performance. These findings demonstrate that high predictive power and interpretability can be achieved without the use of complex machine learning algorithms, providing a practical analytical tool for assessing the quality of biochar and for the efficient utilization of forest residues.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-37261-z.

Introduction

The escalating global climate crisis has highlighted the urgent need for effective strategies to reduce greenhouse gas emissions and enhance carbon sequestration^1,2. The consequences of climate change—including rising sea levels, floods, droughts, wildfires, and extreme heat—are increasingly manifesting as severe environmental disasters^3,4. These events underscore the importance of sustainable carbon management and the development of circular resource technologies^5,6.

Within this context, biochar has attracted significant attention as a promising solution for achieving both carbon neutrality and resource circularity^7,8. Biochar is a carbon-rich, chemically stable material produced through the pyrolysis of biomass at temperatures between 200 °C and 700 °C under oxygen-limited conditions⁹. Due to its high carbon stability and porous structure, biochar offers potential for a wide range of environmental applications, including soil amendment, pollutant adsorption, energy storage, and catalyst support^10–12. Notably, its capacity to enhance soil water retention and nutrient uptake directly contributes to long-term carbon sequestration in terrestrial ecosystems^13,14.

Forestry operations in the Republic of Korea generate an estimated 1.43 million tons of underutilized forest biomass annually, valued at approximately 10 million USD^15,16. Despite its abundance, most of this biomass remains unused and is commonly disposed of through open burning. The heterogeneous nature of this biomass—stemming from variations in tree species and anatomical components—poses challenges in achieving consistent material properties and optimizing carbonization processes. To address these challenges, reliable quality assessment techniques based on elemental composition are essential.

The elemental composition of biochar, particularly its carbon (C), hydrogen (H), oxygen (O), and nitrogen (N) contents, is a critical determinant of its quality and functionality. Conventionally, these elements are quantified using elemental analyzers, which are destructive, time-consuming, and costly methods^17,18. To overcome these limitations, spectroscopic techniques, such as infrared (IR) and near-infrared (NIR) spectroscopy, have emerged as viable alternatives for the rapid and non-destructive characterization of biochar¹⁹.

IR spectroscopy offers the advantage of rapidly acquiring information about chemical functional groups. When combined with peak interpretation and chemometric modeling, it can effectively predict various biochar properties, including ash content, volatile matter, and fixed carbon^20,21. More recently, machine learning-based regression techniques have enabled the quantitative estimation of additional functional properties, such as specific surface area, pore volume, and metal sorption capacity^22,23. However, most existing studies have focused on biochar produced from agricultural residues or standardized feedstocks, with few predictive models specifically developed for biochar derived from underutilized forest biomass^24,25. In addition, systematic data quality control procedures such as outlier detection and removal have rarely been incorporated, even though they are crucial for improving model robustness and generalizability. Furthermore, there is an increasing demand for interpretable and practically deployable models suitable for field applications.

To address these gaps, the present study proposes a chemometric framework that integrates IR spectroscopy with partial least squares regression (PLSR) for the rapid, non-destructive prediction of carbonization-related indices in biochar derived from underutilized forest biomass. The proposed approach emphasizes model interpretability, robustness through outlier handling, and field applicability, thereby promoting the utilization of underused biomass resources and improving the efficiency and controllability of biochar production processes. These aspects collectively distinguish this study from previous IR-based biochar analyses and demonstrate its novelty and practical significance.

Materials and methods

Feedstock description

The biochar feedstock used in this study was underutilized forest biomass (UFB), supplied by Shinyoung E&P Co., Ltd. (Cheongju, Korea). To ensure sample homogeneity, the raw material was first ground using a grinder (CFM-201XB, Cuckoo Electronics, Yangsan, Korea). It was then sieved through a 50-mesh (≈ 300 μm) screen using a sieve shaker (SS2000, Beijing Grinder Instrument Co., Ltd., China) to obtain a uniformly sized powder. This preprocessing step aims to reduce experimental variability due to heterogeneity of feedstock and ensure reproducible results.

Carbonization procedure

The sieved UFB powder was placed in a carbonization boat and thermally treated in a tube-type furnace (LF-GT660, LK Lab Korea, Namyangju, Korea). During carbonization, nitrogen (N₂) gas was continuously introduced at a flow rate of 5 mL/min to maintain an oxygen-free atmosphere. This flow rate was selected based on previous studies using similar laboratory-scale systems^26,27, which demonstrated that it was sufficient to ensure inert conditions without excessive gas consumption. The carbonization process was performed at target temperatures of 200 °C, 300 °C, and 400 °C, each maintained for 1 h (Fig. 1). Upon reaching the desired temperature, the samples were held isothermally and then allowed to cool naturally. A total of 20 samples were prepared, comprising six biochar samples per temperature condition and two uncarbonized, ground UFB controls. The resulting biochar was ground and sieved to produce a homogeneous powder for subsequent physicochemical analyses.

Fig. 1.

Schematic representation of the biochar production process from underutilized forest biomass.

Elemental analysis

The C, H, N, and sulfur (S) contents of the biochar samples were measured using an elemental analyzer (FlashSMART, Thermo Fisher Scientific, Waltham, MA, USA). Before analysis, approximately 2.0 mg of each biochar sample was weighed and sealed in a tin capsule. Elemental contents were determined as weight percentages (wt%) using a combustion-based method. The oxygen content was calculated by subtracting the sum of the C, H, N, and S contents from 100.

ATR-IR spectroscopy

Infrared (IR) spectra of the biochar samples were acquired using an attenuated total reflection infrared (ATR-IR) spectrometer (Alpha-P, Bruker Optics, Billerica, MA, USA). Measurements were performed using dried biochar powder without further pretreatment. Each sample was placed directly on the ATR crystal in contact mode. Spectra were collected across a wavenumber range of 4000–400 cm⁻¹ with a spectral resolution of 4 cm⁻¹. For each sample, 16 scans were averaged to improve the signal-to-noise ratio. To address potential heterogeneity in particle size and composition within each batch, 12 independent spectral measurements were conducted for each sample. Each measurement used powder drawn from a distinct subsample location, thereby improving the representativeness and reliability of the acquired spectra.

Spectral data processing for multivariate analysis

The raw spectral data underwent a series of preprocessing steps to prepare them for predictive modeling. First, L2 normalization was applied to correct differences in absorbance intensity across spectra and enhance their comparability. For each spectrum X = [x₁, x₂, …, x_n], normalization was performed according to Eq. 1:

1

where x_i represents the absorbance at the i-th wavenumber. This transformation scales each spectrum to unit length while preserving the relative shape and peak patterns, thereby facilitating subsequent multivariate analysis²⁸.

Next, second-derivative preprocessing using the Savitzky–Golay filter was applied to reduce baseline variations and enhance the resolution of key absorption bands²⁹. The filter was applied with a window size of 41, a polynomial order of 3, and a derivative order of 2. The window size was empirically determined through visual inspection of representative spectra to prevent noise amplification commonly associated with second-derivative processing while preserving the characteristic peak shape. A third-order polynomial was chosen to balance peak curvature preservation and noise suppression, whereas a second-order derivative was applied to enhance band resolution and remove baseline drift, providing a stable and robust preprocessing foundation for subsequent chemometric modeling. The final normalized second-derivative spectra were used as input variables for the multivariate analysis. All preprocessing, modeling, and data analysis tasks were conducted using the Python programming environment.

Principal component analysis (PCA)

PCA was conducted to explore the structural characteristics of the IR spectral data and visually assess the distribution patterns among samples. PCA is a dimensionality reduction technique that transforms the variance structure of high-dimensional multivariate data into a new orthogonal coordinate system. This transformation retains the primary sources of variation while preserving the correlations among the original variables³⁰.

IR spectra subjected to L2-normalization and second-derivative processing were used as input for PCA. The resulting principal component (PC) scores were used to visualize the relative positions of individual samples, thereby enabling the identification of distributional patterns, potential clustering, and anomalous samples. In addition, PC loadings were examined to identify the spectral regions that contributed the most to the data structure. These loadings provided a basis for interpreting the variables in subsequent analyses.

Partial least squares regression (PLSR) modeling

PLSR was conducted to predict the elemental composition of biochar from its IR spectral data. PLSR is a regression method that provides high predictive accuracy and interpretability, particularly suited to cases where input variables are high-dimensional and exhibit multicollinearity³¹. In this study, the input variables were the IR spectra processed using L2 normalization and second-derivative filtering, while the output variables included elemental indices such as carbon content (%C), atomic oxygen-to-carbon ratio (O/C), and atomic hydrogen-to-carbon ratio (H/C).

To enhance model generalizability, group K-fold cross-validation was employed, with biochar samples from the same batch treated as a group. All 12 replicate spectra per sample were used as individual observations, and replicates from the same sample were grouped to ensure that they were assigned exclusively to either the calibration or prediction set, thereby avoiding inflated model performance.

Model performance was assessed based on the variation in root mean squared error (RMSE) across folds. The optimal number of latent variables (factors) was determined by identifying the point beyond which the additional factor produced negligible reductions in RMSE, thereby balancing the prediction accuracy with model simplicity.

Assessment of residual normality

To validate the predictive performance of the PLSR model and ensure the reliability of residual-based statistical analyses, the distributional characteristics of the prediction residuals were evaluated using the prediction set. Residual normality is a key assumption for inferential procedures, including outlier detection, confidence interval estimation, and statistical interpretation. Therefore, it is essential to confirm whether the residuals follow a normal distribution before applying such methods³².

In this study, residuals were defined as the difference between the predicted and measured values for each sample obtained through cross-validation. The normality of these residuals was assessed using the Shapiro–Wilk test³³, a non-parametric method known for its sensitivity in detecting deviations from normality, even with relatively small sample sizes³⁴.

The Shapiro–Wilk statistic (W) and corresponding p-value were used to guide the selection of appropriate outlier detection methods. When the residuals satisfied the normality assumption, the standardized residual-based approaches were considered suitable. Conversely, if the residuals did not conform to a normal distribution, alternative methods less dependent on normality were adopted.

Outlier detection

To enhance the predictive reliability of the model and prevent distortion of regression coefficients by extreme values, an outlier detection and removal procedure was implemented. This process was preceded by an assessment of the normality assumption for the PLSR residuals. Based on the residual distribution, studentized residuals were used for outlier detection due to their statistical robustness and reduced sensitivity to violations of normality assumptions³². A studentized residual is defined as the residual for an observation standardized by the estimated residual standard deviation obtained by excluding that observation (i.e., leave-one-out estimation). It is calculated using Eq. 2 as follows:

2

where ei is the residual for the i-th observation, s_(i) is the residual standard deviation estimated without the i-th observation, and h_ii is the leverage of that observation. This statistic enables a quantitative assessment of how far each residual deviates from the expected behavior of the model, providing a more reliable basis for identifying outliers than raw residuals, as it partially accounts for leverage and variance structure.

Observations with absolute studentized residuals greater than 3.0 (|t_i| > 3.0) were considered outliers following a widely accepted threshold^32,35. Identified outliers were excluded from the modeling process, and the remaining data were used to construct a revised PLSR model. Model performance was then compared before and after outlier removal to evaluate the impact of the regression coefficient on predictive accuracy and interpretability.

Measures of variable importance

To identify the key spectral regions that contribute to the predictive performance of the PLSR model and evaluate the relative importance of each variable, variable importance in projection (VIP) scores were computed³⁶. The VIP score is a quantitative index that reflects the contribution of each independent variable to the overall predictive ability of the model, and is widely used for variable selection and interpretation in prediction-oriented analyses. It is defined by Eq. 3:

3

where p is the total number of predictor variables, A is the number of selected latent variables (factors), SSₐ(Y) represents the variance in the response variable Y explained by the a-th latent variable, SSₜₒₜₐₗ(Y) is the total variance of Y, and is the weight of the j-th variable in the a-th component.

By definition, VIP scores are normalized such that their average across all variables equals 1.0. Variables with VIP scores greater than 1.0 are interpreted as contributing more than average to the predictive performance of the model. Therefore, a threshold of VIP > 1.0 was adopted to determine the importance of the variable. Spectral regions exceeding this threshold were significant contributors and were subsequently prioritized in the interpretation of IR spectra and the analysis of the chemical characteristics of biochar.

Results and discussion

Elemental composition of biochar

The elemental composition of biochar—particularly carbon, hydrogen, and oxygen—provides key insights into its chemical structure and degree of carbonization^37,38. The contents of these elements were quantitatively analyzed, and their trends were examined in relation to the carbonization temperature.

Table 1 shows that each elemental component exhibited distinct trends depending on the carbonization temperature. Biochar produced at 200 °C contained relatively high amounts of hydrogen and oxygen. As the carbonization temperature increased, the carbon content gradually increased, whereas the hydrogen and oxygen contents declined. This behavior is attributed to the promotion of dehydrogenation and deoxygenation reactions at elevated temperatures. At 400 °C, the carbon content peaked at approximately 67.3%, and the atomic O/C and H/C ratios decreased to 0.32 and 0.64, respectively. These changes indicate that the biochar structure undergoes progressive aromatization and improved thermal stability with increasing temperature.

Table 1.

Elemental composition of biochar samples produced at different carbonization temperatures.

Sample ID	Carbonization temperature (℃)	Elemental composition (wt%)					Atomic ratio (mol/mol)
		C	H	O1)	N	S	O/C	H/C
Control		45.00	5.57	49.06	0.37	0.00	0.82	1.48
B200	200	47.21	5.59	46.81	0.39	0.00	0.75	1.41
B300	300	53.67	5.07	40.77	0.49	0.00	0.57	1.13
B400	400	67.30	3.62	28.41	0.67	0.00	0.32	0.64

¹⁾Oxygen (O) content was calculated by difference: O (wt%) = 100 – (C + H + N + S)

Figure 2 presents a Van Krevelen diagram illustrating the changes in the carbon characteristics of biochar as a function of the atomic O/C and H/C ratios with increasing carbonization temperature. These ratios represent the relative proportions of oxygen and hydrogen and are closely related to the structural properties and stability of biochar³⁷.

Fig. 2.

Compositional changes of biochar samples in Van Krevelen space with increasing carbonization temperature, along with the reference biomass-to-coal trajectory.

As the carbonization temperature increases, the data points shift from the upper right to the lower left of the diagram. This shift reflects the ongoing loss of hydrogen and oxygen due to the reactions of thermal degradation. These compositional changes indicate an increase in aromaticity and thermal stability, suggesting that stable aromatic ring systems predominate in the molecular structure of biochar^37,38. Notably, biochar produced at 400 °C exhibits O/C and H/C ratios of approximately 0.3 and 0.5, respectively. These values position the sample within the Van Krevelen region characteristic of lignite or low-rank coal, indicating that biochar undergoes structural transformations similar to those of coalification, even at moderate temperatures. Consequently, the material achieves enhanced structural integrity and long-term durability.

Spectral characteristics

As illustrated in Fig. 3, which presents the averaged spectra for each carbonization condition, the normalized and second-derivative IR spectra reveal clear temperature-dependent trends. Overall, distinct changes in spectral intensity and band patterns were observed as the carbonization temperature increased. Raw spectra prior to normalization are provided in Supplementary Figure S1 for reference, and additional enlarged views of the normalized and second-derivative spectra covering the major regions are presented in Supplementary Figures S2 and S3. Table 2 summarizes the major absorption bands and their corresponding functional groups.

Fig. 3.

ATR-IR spectra of biochar samples produced at different carbonization temperatures: (a) normalized spectra and (b) second-derivative spectra. Each spectrum represents the average of repeated measurements across all samples for each carbonization condition.

Table 2.

Major ATR-IR absorption bands and their peak intensities at different carbonization temperatures.

Wavenumber (cm− 1)	Functional group assignment	Control	200 ℃	300 ℃	400 ℃
3320	O–H stretching (hydrogen-bonded OH)39,40	0.035 (0.002)	0.030 (0.002)	0.026 (0.002)	0.016 (0.002)
2920	C–H stretching41,42	0.026 (0.002)	0.023 (0.001)	0.020 (0.001)	0.019 (0.002)
1728	C = O stretching (hemicellulose)42,43	0.015 (0.002)	0.015 (0.002)	0.014 (0.003)	0.009 (0.003)
1595	Aromatic C = C stretching (lignin)42,43	0.023 (0.001)	0.022 (0.001)	0.027 (0.004)	0.052 (0.002)
1429	CH2 bending (crystalline cellulose)41–43	0.012 (0.001)	0.013 (0.001)	0.021 (0.003)	0.036 (0.001)
1315	Aromatic skeletal vibration (lignin)41–44	0.014 (0.001)	0.018 (0.001)	0.028 (0.003)	0.044 (0.001)

Values represent mean peak intensities of normalized ATR-IR spectra. Numbers in parentheses indicate standard deviation.

In the normalized spectra (Fig. 3a), a broad absorption band centered around 3600–3000 cm⁻¹ progressively weakens with increasing temperature, corresponding to a reduction in the number of hydroxyl groups. Similarly, the C–H stretching bands in the 3000–2800 cm⁻¹ region decrease, reflecting the thermal degradation of aliphatic hydrocarbons⁴⁵. Near 1600 cm⁻¹, the intensities of absorption bands associated with C = O or C = C groups decrease; however, minor contributions from nitrogen-containing functionalities (e.g., C = N stretching or N–H bending) cannot be excluded, considering the increase in nitrogen content with carbonization temperature (Table 1)⁴⁶. At higher temperatures, the observed decrease of N–H stretching around 3300 cm⁻¹ is consistent with the formation of aromatic nitrogen species such as pyrrolic and pyridinic N. Spectral features in the 1200–1000 cm⁻¹ region become less complex at higher temperatures, showing a reduction in the number and intensity of absorption bands. This trend is consistent with structural rearrangements and a loss of oxygenated functional groups during carbonization⁴⁷.

The second-derivative spectra (Fig. 3b) improved the band resolution and revealed subtle changes in the overlapping peaks, enabling a more detailed analysis of the functional group evolution. Notably, the region from 1600 to 1000 cm⁻¹ exhibits variations in both peak position and intensity with increasing temperature. The peak at 1595 cm⁻¹ corresponds to C = C stretching vibrations in lignin aromatic rings^43,48, and its subtle shifts indicate progressive condensation of lignin-derived aromatic structures at elevated temperatures.

Peaks observed at 1462 cm⁻¹ and 1429 cm⁻¹ in the normalized spectra were attributed to CH₂ bending vibrations in amorphous and crystalline cellulose, respectively^41–43. These peaks disappeared as the temperature increased, indicating the degradation of the cellulose backbone and the collapse of crystalline domains. The spectral variations observed in the 1300–1000 cm⁻¹ range are primarily associated with the C–O stretching and C–H bending modes, indicating the presence of oxygen-containing functional groups^49,50. Collectively, these spectral changes provide evidence of functional group loss, molecular rearrangement, and increased aromaticity during thermal treatment. The progressive aromatization and condensation of carbon structures imply a higher degree of carbonization and improved structural stability of the resulting biochar⁵¹.

PCA

To quantitatively assess the compositional variation in IR spectra as a function of carbonization temperature, PCA was applied separately to the normalized and second-derivative spectra. The results are presented in Fig. 4, where the upper panels (Fig. 4a and b) show the score plots, representing sample distribution patterns, and the lower panels (Fig. 4c and d) show the loading plots that illustrate the spectral variables contributing to principal component separation.

Fig. 4.

PCA score plots (a, b) and loading plots (c, d) derived from normalized (a, c) and second-derivative (b, d) ATR-IR spectra of biochar samples produced at different carbonization temperatures. All plots are based on the first two principal components.

The first two principal components, PC1 and PC2, account for approximately 79% of the total variance in the score plot based on the normalized spectra (Fig. 4a), with PC1 contributing 61.6% and PC2 contributing 17.1%. Samples display a clear progression along the PC1 axis in the following order: control, 200 °C, 300 °C, and 400 °C, reflecting compositional shifts associated with increasing carbonization temperatures. This distribution suggests that the dominant spectral variations are effectively captured along PC1.

The corresponding loading plot (Fig. 4c) indicates that the variables in the 1800–1000 cm⁻¹ region exhibit strong loadings on PC1, driving much of the differentiation. Although a portion of variables in the < 1000 cm⁻¹ region lie near zero, a large number deviate clearly from the origin along PC1, indicating a substantial contribution. These regions encompass absorption bands related to C = O, C = C, C–O, aromatic structures, and inorganic functional groups—all of which are sensitive to thermal treatment. Conversely, the high-wavenumber region (> 2800 cm⁻¹), associated with O–H and C–H stretching shows relatively low loadings, indicating limited structural information beyond moisture and alkyl group content. Thus, PC1 primarily reflects changes in oxygenated and aromatic functional groups during the carbonization process.

The score plot based on the second-derivative spectra (Fig. 4b) explains a lower cumulative variance (PC1: 30.1%, PC2: 10.0%; total: 40.1%), yet the clustering of samples appears more compact and distinct. In particular, the biochar produced at 400 °C is separated from the other groups, which can be attributed to the enhanced spectral resolution and peak separation provided by the second-derivative processing. This approach emphasizes subtle differences in functional groups that may be evident in normalized spectra.

The associated loading plot (Fig. 4d) shows that the variables below 1800 cm⁻¹ are widely distributed along the outer ellipse, indicating strong contributions to PC1 and PC2. These regions correspond to functional groups, such as C = O, C–O, aromatic systems, and out-of-plane vibrations, which are sensitive to carbonization. In contrast, variables above 2800 cm⁻¹ are clustered near the center of the ellipse, implying minimal or consistent influence on structural variation. Therefore, the primary sources of variation in PC1 and PC2 are the oxygen-containing and aromatic functional groups.

These findings demonstrate that PCA applied to IR spectral data effectively captures the compositional evolution of biochar with increasing carbonization temperature. While normalized spectra better reflect gradual compositional changes across temperatures, second-derivative spectra provide enhanced resolution of subtle differences in functional groups among samples. Analysis of the loading plots identified key spectral regions responsible for structural variation, offering insight into the chemical properties of biochar and supporting the identification of reactive domains.

PLSR analysis

Performance of the PLSR models

Figure 5; Table 3 present the results of the PLSR analysis for predicting the carbonization characteristics of biochar. Figure 5a shows the RMSE profile used to determine the optimal number of latent factors, with nine factors selected for the final model predicting carbon content. Figure 5b and d depict the relationships between measured and predicted values for %C, O/C, and H/C, respectively, demonstrating high coefficients of determination (R²) and strong agreement in both calibration and prediction sets. These results confirm the robustness and reliability of the models.

Fig. 5.

PLSR results using normalized ATR-IR spectra: (a) RMSE profile used to determine the optimal number of latent components for carbon content prediction, with the selected value indicated by a red marker; (b–d) plots of measured vs. predicted values for carbon content, atomic O/C ratio, and atomic H/C ratio, respectively.

Table 3.

Predictive performance of the PLSR models for estimating the carbonization characteristics (%C, O/C, and H/C) of Biochar based on normalized and second-derivative ATR-IR spectra.

Response	Preprocessing	#Factors	R 2 C	RMSEC	R 2 P	RMSEP
%C	Normalized	9	0.989	0.942	0.971	1.508
	2nd-derivative	2	0.971	1.502	0.956	1.847
O/C	Normalized	8	0.980	0.027	0.952	0.042
	2nd-derivative	2	0.961	0.038	0.944	0.046
H/C	Normalized	10	0.996	0.021	0.986	0.039
	2nd-derivative	6	0.997	0.019	0.986	0.038

#Factors: Number of latent variables used in the PLSR model; R²_C/R²_P: Coefficient of determination for calibration/prediction; RMSEC/RMSEP: Root mean square error for calibration/prediction.

Two types of preprocessed spectra—normalized and second derivative—were used as input variables. In all cases, the models achieved high predictive accuracy, with strong R² values and low RMSE values across both calibration and prediction datasets (Table 3). Notably, models based on normalized spectra consistently outperformed those using second-derivative spectra, yielding R² values of 0.971, 0.952, and 0.986 for %C, O/C, and H/C, respectively.

In contrast, models developed using second-derivative spectra required fewer latent factors but exhibited slightly lower or comparable predictive performance. This suggests that while second-derivative preprocessing improves peak resolution and enhances sensitivity to subtle spectral features, it may also increase noise susceptibility and reduce data stability. This effect may partly reflect the use of a relatively large smoothing window, which effectively reduced noise but could have attenuated fine spectral details relevant to model sensitivity.

Overall, the PLSR models developed in this study demonstrated high accuracy and reproducibility for the non-destructive quantitative prediction of biochar carbonization characteristics. The comparison between preprocessing methods highlights the influence of spectral treatment on model performance and underscores the importance of selecting appropriate preprocessing techniques in chemometric modeling.

Outlier detection

To evaluate the reliability of the constructed PLSR models, residual analysis was conducted. Figure 6a shows that the Shapiro–Wilk test for the carbon content prediction model based on normalized spectra produced a W value of 0.962 (p < 0.001), indicating a statistically significant violation of the normality assumption. This suggests the presence of asymmetry or outliers in the residual distribution. Consequently, outlier detection was performed using Studentized residuals.

Fig. 6.

Residual analysis of the PLSR model based on normalized ATR-IR spectra for predicting carbon content: (a) quantile–quantile plot used to assess the normality of residuals (Shapiro–Wilk test, W = 0.962, p < 0.001); (b) outlier identification using Studentized residuals. Detected outliers are annotated with sample IDs (Temp–Sample–Replicate).

Using an outlier threshold of ± 3, several samples were identified as outliers. Outliers were observed in the model based on normalized spectra among the 300 °C and 200 °C samples (Fig. 6b). In contrast, an outlier was detected from the control (untreated) group for the model developed using second-derivative spectra. This may be attributed to the characteristics of second-derivative preprocessing, which enhances subtle spectral differences and may increase variability in samples with minimal chemical change, such as untreated controls.

Among all models, only the residuals of the H/C prediction model based on normalized spectra satisfied the normality assumption. All other models exhibited statistically significant deviations, suggesting that both the preprocessing technique and the specific target variable influence the residual distribution. These differences should be considered during further model refinement or outlier adjustment.

The analysis suggests that the identified outliers are not merely numerical anomalies, but may also reflect complex or nonlinear interactions between spectral features and carbonization behavior. Potential contributing factors include uneven thermal treatment, compositional heterogeneity, measurement noise, and instrumental variability. These considerations are essential for improving the generalizability and robustness of predictive models in future applications.

Impact of outlier removal on model performance

To assess the influence of outliers on model performance, PLSR models were reconstructed after removing the identified outliers using the same modeling procedure, and their performance was compared (Table 4). Overall, the removal of outliers had a positive impact on the predictive accuracy of most models.

Table 4.

Comparison of PLSR model performance before and after outlier removal based on studentized residuals.

Response	Preprocessing	w/Outliers			w/o Outliers
		#Factor	R2P	RMSEP	#Factor	R2P	RMSEP
%C	Normalized	9	0.971	1.508	5	0.983	1.206
	2nd-derivative	2	0.956	1.847	5	0.965	1.168
O/C	Normalized	8	0.952	0.042	6	0.979	0.029
	2nd-derivative	2	0.944	0.046	6	0.946	0.043
H/C	Normalized	10	0.986	0.039	10	0.989	0.036
	2nd-derivative	6	0.986	0.038	No outliers

#Factors: Number of latent variables used in the PLSR model; R²_P: Coefficient of determination for prediction; RMSEP: Root mean square error for prediction.

The original model employed nine latent factors for predicting carbon content based on normalized spectra, achieving an R²_P of 0.971 with an RMSEP of 1.508. Following outlier removal, the number of latent factors decreased to five, R²_P improved to 0.983, and RMSEP was reduced to 1.206. This indicates that outliers adversely affected model performance, and their exclusion led to both model simplification and enhancement. Similarly, the model based on second-derivative spectra also improved after outlier removal, despite requiring more latent factors; RMSEP decreased from 1.847 to 1.168. This suggests that outlier removal improved prediction consistency regardless of the preprocessing method used.

For the O/C and H/C ratio predictions, the model performance generally improved after removing outliers. In the O/C prediction model using normalized spectra, R²_P increased from 0.952 to 0.979 and RMSEP decreased from 0.042 to 0.029, indicating a substantial improvement. However, the second derivative-based model showed only minor gains, suggesting that the impact of outliers was limited in this case. For H/C prediction, slight improvements were observed in the normalized model, whereas no outliers were identified in the second-derivative model, and its structure remained unchanged. These results demonstrate that the benefits of outlier removal depend on both the preprocessing technique and the target variable.

Collectively, these findings confirm that outliers can distort model training and generalization performance. Statistically guided detection and removal of outliers are critical preprocessing steps in enhancing multivariate prediction models. As the impact of outlier removal varies with preprocessing strategy and prediction target, a flexible approach tailored to variable sensitivity and dataset characteristics is essential for constructing robust and generalizable models. However, in real-world applications with higher sample variability and measurement noise, strict outlier exclusion strategies may have limited effectiveness. Incorporating robust modeling or adaptive preprocessing approaches can help maintain model stability under such variable conditions.

Analysis of variable importance

The VIP scores were analyzed to identify the key spectral regions contributing to the predictive performance of the PLSR models for the characteristics of biochar carbonization (Fig. 7). Spectral variables with VIP scores greater than 1.0 are generally considered important because they contribute more than the average to the overall explanatory power of the model^31,52.

Fig. 7.

Variable importance in projection (VIP) scores from PLSR models used to predict biochar carbonization characteristics (%C, O/C, and H/C) based on normalized ATR-IR spectra.

The VIP profiles presented in Fig. 7 indicate that spectral regions below 1800 cm⁻¹ generally exhibit higher importance, with the most prominent peak observed at approximately 890 cm⁻¹ across all three models. This region corresponds to the C–O–C valence vibration of β-glycosidic linkages⁴¹. The progressive reduction in the intensity of this peak with increasing carbonization temperature reflects the decomposition of glycosidic bonds and the breakdown of carbohydrate backbones (Fig. 3), leading to the formation of condensed carbon structures—a hallmark of biochar transformation^37,53.

The %C prediction model exhibited consistently high VIP values across the spectral range, with a particularly strong contribution near 870 cm⁻¹. The H/C model highlighted the influence of C = O and C = C functional groups in the 1800–1600 cm⁻¹ region, and also showed notable importance in the 3000–2800 cm⁻¹ region. This latter region corresponds to aliphatic C–H stretching bands⁵⁴, whose intensity decreases with increasing temperature, indicating the degradation of long-chain aliphatic structures and their conversion into aromatic rings. These transitions reflect hydrogen loss and increased aromaticity, contributing to the observed reduction in the H/C ratio.

In contrast, the O/C model exhibited relatively low and flat VIP scores, with slightly elevated contributions primarily below 1000 cm⁻¹. These results suggest that oxygen-containing functional groups, such as C–O and O–H, contribute broadly across the lower wavenumber region, and their progressive removal during carbonization leads to changes in the spectral pattern relevant to O/C ratio prediction.

These findings align with the key spectral regions identified in the PCA loading plots (Fig. 4c and d), particularly in the 1600–1000 cm⁻¹ and < 1000 cm⁻¹ ranges. This overlap highlights the significance of these regions in determining the chemical structure and degree of carbonization in biochar. Furthermore, the integrated use of VIP and PCA loading analyses provides a practical strategy for selecting informative spectral bands to enhance model accuracy, supporting applications in biochar quality assessment, feedstock differentiation, and process optimization.

Potential and limitations of the predictive model

The PLSR models demonstrated strong potential for predicting the key carbonization characteristics of biochar using ATR-IR spectral data. Notably, the use of normalized spectra alone enabled the accurate quantitative estimation of the elemental composition without the need for conventional elemental analysis. This highlights the practical value of the model as a rapid, non-destructive tool for assessing biochar quality and suggests its potential extension to real-time, inline monitoring systems.

However, several considerations must be addressed before applying predictive models in broader contexts. The current models were developed from a dataset based on a specific type of biomass under limited carbonization conditions. Consequently, predictive accuracy and model transferability may be limited when different feedstocks or processing parameters are used, owing to inherent compositional differences and variations in carbonization behavior. A more practical and widely adopted strategy is to develop expanded models trained on multi-feedstock datasets⁵⁵, which can better capture compositional diversity and improve predictive robustness across a broader range of biomass types and carbonization setups.

It is also important to acknowledge that the high predictive performance reported in this study was based on cross-validation within a controlled experimental setting. In practical applications, prediction errors may increase due to feedstock heterogeneity, `environmental variability during measurement, or extrapolation to conditions that are not directly observed. Therefore, model predictions should be supplemented with preliminary validations and additional analyses to avoid overreliance in field conditions.

Moreover, the quality and diversity of the dataset play a critical role in the performance of the model. Model complexity, outlier sensitivity, and variable importance are all influenced by the composition of the dataset. Sufficient data accumulation and repeated measurements are crucial for enhancing predictive robustness. Under such conditions, the IR-based predictive modeling framework proposed in this study could serve as a reliable and efficient analytical platform for biochar characterization and quality control.

Conclusions

This study demonstrated that ATR-IR spectroscopy combined with PLSR enables accurate prediction of carbonization characteristics of biochar derived from underutilized forest biomass. The modeling framework effectively captured temperature-dependent compositional and structural transformations, identifying key spectral regions associated with carbonization. These findings highlight the potential of IR-based multivariate modeling as a rapid and non-destructive tool for evaluating biochar quality and monitoring the carbonization process. Expanding this approach to a wider range of biomass feedstocks and carbonization conditions could further enhance its robustness and industrial applicability.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

YK was the primary contributor to this study and drafted the manuscript. CH and HS conducted the preparation and characterization of the biochar samples. SWH led the predictive modeling and data interpretation, while BK was primarily responsible for the chemical analysis. The original concept was jointly developed by SWH and BK. All authors reviewed and approved the final version of the manuscript.

Funding

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2025-00560264).

Data availability

The datasets used or analyzed during the current study are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.