From Leaf to Blend: CNN-Enhanced Multi-source Feature Fusion Enables Threshold-Driven Style Control in Digital Tobacco Formulation

doi:10.21203/rs.3.rs-6468408/v1

Background:This study establishes a computational framework for predictive style modeling in tobacco formulation design, resolving the critical disconnect between empirical approaches and blended system complexity. A convolutional neural network (CNN) framework was developed to integrate conventional chemical indicators with thermogravimetric analysis derived features from 434 geographically authenticated tobacco leaf samples. By implementing regionally constrained Monte Carlo sampling of composition ratios, 304,800 formulation datasets simulating real-world blending constraints were generated to enable robust model training.

Results: The leaf-centric CNN demonstrated remarkable region-style classification accuracy (99.54% via five-fold cross-validation), outperforming conventional machine learning models and revealing thermal-chemical complementarity in regional style characterization. However, direct application to blended formulations revealed a critical limitation: only 50.91% of blended formulations maintained stylistic consistency with their primary source leaves, underscoring the inadequacy of single-material models for blended systems. To overcome this, a hybrid learning model through multi-source data fusion of leaf and formulation representations waas engineered, achieving dual breakthroughs: 90.09% regional style identification accuracy and 87.90% leaf-to-blend style consistency. Mechanistic analysis identified a nonlinear threshold effect, showing that primary source leaves maintained 99.91% stylistic dominance when their composition exceeded 90%, decreasing to 67.90% at 30% composition. Significant formulation style deviation risks emerged when contribution gaps between principal and secondary source leaves composition narrowed below 10%.

Conclusions: Building on these insights, a probabilistic style modulation strategy was proposed and validated through case applications. This data-driven framework advances tobacco engineering from empirical practices to predictive digital transformation, providing a template for agricultural product manufacturing systems with similar formulation challenges.

Multi-Source Feature Fusion

Convolutional Neural Network (CNN)

Regional Style Classification

Formulation Style Control

The geographic variation in tobacco cultivation regions produces distinct stylistic signatures, rendering both the quantitative characterization of these region-specific attributes and the systematic optimization of blended formulations barriers to intelligent manufacturing transformation in the tobacco industry. Current methodologies predominantly depend on artificial experience for multimodal assessment of morphological and organoleptic properties [1, 2], yet encounter persistent limitations including substantial operator-dependent variability and inadequate metrological foundations.

Recent advancements in analytical instrumentation have catalyzed novel approaches for digital style profiling: Systematic investigations into chemical constituents including total sugars and nicotine concentrations have elucidated region-specific distribution patterns of chemical quality markers [3, 4]; Near-infrared spectroscopy has empowered geographical provenance determination through machine learning-driven spectral fingerprint analysis [5, 6]; Thermogravimetric analysis has elucidated dynamics through pyrolysis kinetic modeling that correlates with geographical origins and grades [7, 8]. In morphological analysis, the integration of hyperspectral imaging systems with advanced machine learning architectures has enabled automated quantification of morphometric descriptors including chromatic signatures and surface topology [9–11]. Three critical limitations persist in current research paradigms: First, predominant reliance on unimodal data streams without establishing cross-domain feature synergy frameworks; Second, most models target single-grade leaves, lacking quantitative modeling of style superposition and deviation effects in complex blending formulations; Third, conventional small-scale samples that inadequately capture hierarchical influence networks between primary and secondary source leaves in blending systems.

This investigation undertakes a systematic approach to address these challenges through the development of a multisource deep learning framework that synthesizes heterogeneous data streams. A curated dataset comprising 434 authenticated tobacco leaf samples was established, integrating conventional chemical indicators with thermogravimetric analysis derived features to construct a bivariate feature space spanning tobacco composition and pyrolysis characteristics. The implementation of Monte Carlo stochastic sampling methodology generated 304,800 unique blend permutations through parametric variation of geographical provenance ratios, effectively mitigating data paucity of conventional experimental designs. A deep convolutional neural network architecture was systematically constructed for single-origin leaf regional style prediction. Multimodal feature fusion at the latent representation level enabled cross-modal information synergy, preceding rigorous canonical correlation analysis between blend stylistic profiles and constituent leaf attributes. This investigation uncovered the limitation in direct extrapolation of single-source models to multi-component blending systems. A dual-scale modeling paradigm was therefore developed, implementing origin-specific contribution to quantify impacts of geographical provenance ratios on overall leaf blend style characteristics. This computational framework establishes mechanistic understanding of blending system, providing a highly interpretable theoretical foundation for intelligent tobacco formulation optimization.

1.1 Materials

Tobacco leaf samples (n=434) were collected from eight major production regions (A-H) in China, representing distinct agro-climatic zones (Table 1). Each region contributed 36-78 samples, covering top, middle, and lower leaf positions, ensuring diversity in both growing regions and leaf positions. Samples were equilibrated referring to YC/T 31[12] at 295 K and 60% RH for 48 h, pulverized through an 80-100 mesh sieve, and stored at 295 K in amber vials to minimize oxidative degradation.

Table 1. Distribution of Tobacco Leaf Samples

Production Region	A	B	C	D	E	F	G	H	Total
Sample Size	54	78	62	61	45	38	36	60	434

1.2 Thermogravimetric Experimental Methods

Thermogravimetric (TG) experiments were performed on a NETZSCH STA 449 F3 analyzer equipped with an alumina crucible. Approximately 10±0.5 mg of tobacco powder was heated from 298 K to 873 K at 10 K/min under ultrapure nitrogen (100 mL/min, ≥99.999%). The TG data were discretely sampled at 0.5 K intervals within 400–800 K, a range covering pyrolysis stages. Raw data were baseline-corrected, smoothed via Savitzky-Golay filtering, and processed with NETZSCH Proteus software to derive Thermogravimetric derivative (DTG) curves.

1.3 Conventional Chemical Composition Detection Methods

In this investigation, six key chemical parameters (total nitrogen, total plant alkaloids, chloride, potassium, water-soluble total sugar, and reducing sugar) were quantified using a continuous flow analyzer (ALLIANCE- Futura). Total nitrogen content was determined following the standardized protocol YC/T 161[13], while total plant alkaloids were quantified according to YC/T 160[14]. Chloride and potassium concentrations were measured using YC/T 162[15] and YC/T 217[16], respectively, with water-soluble sugars and reducing sugar analyzed per YC/T 159[17]. The acquired datasets served as the basis for calculating four critical quality indicators: (i) potassium-to-chloride (K/Cl) ratio, (ii) nitrogen-to-alkaloid (N/Alk) ratio, (iii) total sugar-to-alkaloid (TS/Alk) ratio, and (iv) reducing sugar-to-total sugar (RS/TS) ratio.

2.1 Thermogravimetric Analysis Derived Feature Point Data

DTG curves of tobacco were analyzed to elucidate pyrolysis-kinetic behavior through mass loss rate as a function of temperature. Pioneering work by Baker and Bishop [18] established that DTG curves acquired under inert atmosphere with controlled heating rates exhibit inter-laboratory reproducibility. Divergence in pyrolysis behaviors originates from chemotype-specific distributions of key biopolymers: (i) volatiles (nicotine, sugar), (ii) structural carbohydrates (cellulose, hemicellulose), and (iii) aromatic macromolecules (lignin) [19-23].

However, the raw DTG dataset (n=800 data points per curve) introduced high-dimensional complexity, predisposing predictive models to the curse of dimensionality. Preliminary feature extraction using seven points (A1: onset, A2: termination, B1-B5: inflection) resulted in substantial fitting residuals (Fig. 1a), suggesting inadequate representation of kinetic processes. To overcome these limitations, a three-tiered feature engineering framework was implemented:

Half-peak point enhancement: Augmentation with eight half-peak points (C1-C4 and D1-D4) to parameterize peak asymmetry and shoulder intensity.
Kinetic parameter construction: Calculation of temperature differences (ΔT1–ΔT4) and mass loss rate differences (ΔW1–ΔW4) between half-peak points to quantify pyrolysis rate heterogeneity.
The optimized feature set ensures high consistency between fitted and actual DTG curves (Fig. 1b), significantly improving the characterization accuracy of pyrolysis kinetics. The 38-dimensional feature set (Table 2) underwent Z-score normalization to eliminate dimensional discrepancies.

Table 2. Parameters corresponding to thermogravimetric feature codes

Feature Point	Code	Feature Parameters
Feature Point	Code	T (K)	W (%/K)	∆T (K)	∆W (%/K)
Initial Point	A1	T_A1	W_A1	--	--
Inflection Points	B1、B2、B3、B4、B5	T_B1~T_B5	W_B1~W_B5	--	--
Half-Peak Points	C1、C2、C3、C4、 D1、D2、D3、D4	T_C1~T_C4 T_D1~T_D4	W_C1~W_C4 W_D1~W_D4	∆T₁=T_C1-T_D1 ∆T₂=T_D2-T_C2 ∆T₃=T_D3-T_C3 ∆T₄=T_C4-T_D4	∆W₁=W_C1-T_B1 ∆W₂=T_C2-T_B3 ∆W₃=T_C3-T_B4 ∆W₄=T_C4-T_B5
Terminal Point	A2	T_A2	W_A2	--	--

2.2 Conventional Chemical indicators

The conventional chemical composition characteristics of tobacco leaves show significant correlations with their regional styles. Major Chinese tobacco-producing regions can be effectively differentiated using specific carbohydrate components and mineral element indicators [24]. Research indicates [25] that the characteristic aroma types of tobacco origins are closely linked to differences in their chemical composition.

Total nitrogen, as a representative indicator of total nitrogen-containing compounds, exhibits dual effects in regulating tobacco sensory quality: when the concentration falls within the optimal range of 1.5%–2.8%, it effectively promotes aroma substances, while exceeding this threshold leads to significant changes in sensory quality [26-28]. Nicotine content also shows nonlinear correlations with key sensory indicators such as aroma quality, aroma intensity, irritation, and aftertaste. A nicotine concentration range of 2.0%–3.5% achieves an optimal quality balance [26].

Total sugar (including both reducing and non-reducing sugars) varies significantly among different production regions (P<0.01) [29]. Studies suggest that the highest quality is achieved when total sugar content is 15%-28% and reducing sugar content ranges between 15%-25% [26]. Deviations from these ranges may result in sensory quality degradation.

Among mineral elements, potassium ions show strong correlations with mainstream smoke components. Correlation coefficients indicate significant negative relationships with tar (−0.89), total particulate matter (−0.86), moisture (−0.74), and CO (−0.84) [30]. Chloride ions exhibit significant positive correlations with mainstream smoke components such as tar (0.77), total particulate matter (0.69), moisture (0.55), and CO (0.72) [30]. The relationships between sensory quality and ratios including potassium-to-chloride, nitrogen-to-alkaloid, and sugar-to-alkaloid depend on tobacco leaf origin and stalk position [31,32].

2.3 Dataset

To establish mechanistic linkages between pyrolytic reactivity and chemotypic determinants, a hybrid dataset was architected through multi-source data fusion of thermogravimetric analysis data and conventional chemical composition data.

2.3.1 Tobacco Leaf Sample Dataset

The tobacco leaf samples were structured in the format of "Leaf Identifier-Feature Indicators-Production Region" to construct the dataset. Each sample corresponds to 38 thermogravimetric feature parameters and 10 chemical indicators. To ensure algorithm generalization capability, each of the 8 production regions contains at least 35 distinct samples (see Table 1).

2.3.2 Blended leaf Formulation Dataset

To analyze multi-region combination effects, a large-scale dataset comprising 304,800 blended leaf formulations was constructed, with the following design principles:

Regional combinations: 2-8 regions per blend with fixed total mass (100%).
Proportion gradients: 10% increments within 10%-90% ranges.
Dominance constraints: Single-region maximum proportion per blend.

Data generation involved:

Monte Carlo sampling: Random selection of regional leaves according to blend proportions, with 20 replicates per proportion category.
Feature weighting: Weighted calculation of blend characteristics based on mass fractions [33,34]. All formulations underwent thermogravimetric feature extraction (Section 2.1) and chemical indicators calculation (Section 1.3).

This study conducts algorithmic benchmarking comparing a convolutional neural network (CNN) with three classical machine learning models (Random Forest, K-Nearest Neighbors, Linear Discriminant Analysis) for tobacco leaf origin style classification.

3.1 Convolutional Neural Network Model

A CNN-based style discrimination model is first constructed in this work, with its network architecture illustrated in Figure 2. The network primarily consists of convolutional layers, pooling layers, and fully connected layers. The input layer receives feature vectors representing tobacco leaf samples. When thermogravimetric features are combined with routine chemical components, the input tensor dimension is batch size × 1 × 48. After sliding operations using 1D convolutional kernels of size 3, the feature dimensionality increases. Feature downsampling is achieved through max-pooling layers with a kernel size of 2 and a stride of 2, converting the multi-dimensional feature tensor into a flattened 1D vector for input to the fully connected layer. The fully connected layer outputs a predictive probability distribution corresponding to the 8 candidate production regions, and the region label with the highest probability is selected as the style evaluation result.

3.2 Comparative Models

Random Forest (RF) [35] employs a decision tree ensemble strategy with bagging. During tree node splitting, a feature random selection mechanism is adopted, where a fixed-size subset of features is randomly selected from the full feature set to compute optimal split points. This strategy reduces overall variance by increasing base model diversity, thereby suppressing overfitting. Final predictions are generated by majority voting of base classifiers.

K-Nearest Neighbors (KNN) [36] utilizes a distance-based instance reasoning mechanism, determining class membership by calculating similarity between the target sample and training samples, selecting k nearest neighbors, and assigning the majority class.

Linear Discriminant Analysis (LDA) [37] assumes that all class data follow normal distributions with identical covariance structures. It constructs a low-dimensional discriminant subspace through linear projection, optimizing the goal of maximizing the ratio of inter-class dispersion to intra-class dispersion. Projection directions are determined via generalized eigenvalue decomposition, ensuring concentrated projections for samples of the same class and maximally separated projection centers for different classes.

4.1 Synergistic Effects of Multi-Source Data Fusion and Algorithm Performance Comparison

This study evaluated five data combinations (conventional chemical indicators, thermogravimetric curves, thermogravimetric features, thermogravimetric curves combined with conventional chemical indicators, and thermogravimetric features integrated with conventional chemical indicators) across four algorithms (CNN, RF, KNN, LDA). The enhancement of multi-source data fusion on leaf regional style classification was systematically assessed via five-fold cross-validation, as summarized in Table 3.

Table 3. Regional Style Prediction Performance of Four Algorithms Across Feature Categories

CNN	K1	K2	K3	K4	K5	Average
Chemical Indicators	89.66%	81.61%	85.06%	80.46%	80.23%	83.40%
DTG Curves	97.70%	89.66%	97.70%	89.66%	89.53%	92.85%
DTG Features	83.91%	83.91%	90.80%	83.91%	84.88%	85.48%
DTG Curves + Chemical	98.85%	90.80%	95.40%	88.51%	89.53%	92.62%
DTG Features + Chemical	94.25%	91.95%	97.70%	90.80%	90.70%	93.08%
RF	K1	K2	K3	K4	K5	Average
Chemical Indicators	90.80%	83.91%	72.41%	80.46%	80.23%	81.56%
DTG Curves	88.51%	90.80%	80.46%	90.80%	82.56%	86.63%
DTG Features	87.36%	89.66%	89.66%	87.36%	80.23%	86.85%
DTG Curves + Chemical	90.80%	93.10%	80.46%	91.95%	83.72%	88.01%
DTG Features + Chemical	91.95%	94.25%	91.95%	83.72%	88.37%	90.05%
KNN	K1	K2	K3	K4	K5	Average
Chemical Indicators	63.22%	64.37%	59.77%	59.77%	65.12%	62.45%
DTG Curves	86.21%	82.76%	78.16%	89.66%	81.40%	83.64%
DTG Features	77.01%	73.56%	75.86%	67.82%	66.28%	72.11%
DTG Curves + Chemical	64.37%	64.37%	60.92%	62.07%	63.95%	63.14%
DTG Features + Chemical	79.31%	77.01%	77.01%	73.56%	67.44%	74.87%
LDA	K1	K2	K3	K4	K5	Average
Chemical Indicators	80.46%	80.46%	87.36%	81.61%	86.05%	83.19%
DTG Curves	80.46%	83.91%	73.56%	77.01%	83.72%	79.73%
DTG Features	78.16%	79.31%	82.76%	71.26%	75.58%	77.42%
DTG Curves + Chemical	88.51%	88.51%	86.21%	89.66%	86.05%	87.78%
DTG Features + Chemical	93.10%	88.51%	94.25%	87.36%	95.35%	91.71%

The performance of the four algorithms in regional style prediction across different feature categories indicates: (1) Feature complementarity: CNN, RF, and LDA all show improved prediction performance after fusing thermogravimetric features with chemical indicators, confirming the complementary effects between pyrolysis kinetics and chemical composition, while demonstrating that dimensionality reduction reduces data redundancy. (2) CNN dominance: After combining thermogravimetric features with chemical indicators, CNN achieves an average accuracy of 93.08%, significantly outperforming traditional algorithms (LDA: 91.71%, RF: 90.05%, KNN: 74.87%).

In the approach combining thermogravimetric features with chemical indicators, the model stability (range values) under five-fold cross-validation ranks as CNN (7%) < LDA (8%) < RF (11%) < KNN (12%), validating the robustness of the deep learning model in cross-validation. The accuracy of KNN decreases when integrating high-dimensional DTG curves, reflecting the curse of dimensionality.

This study selects the multi-source data and CNN-based model as the leaf regional style evaluation model for subsequent analysis. It performs exceptionally well in validation on 434 tobacco leaves, as shown in Table 4, achieving an overall accuracy of 99.54%, with region D showing slight misclassification (2/61 samples).

Table 4. Prediction Results of Leaf-centric CNN Model on 434 Tobacco Samples

True Production Region	Sample Size	Predicted Production Region								Errors	Accuracy
True Production Region	Sample Size	A	B	C	D	E	F	G	H	Errors	Accuracy
A	54	54	--	--	--	--	--	--	--	0	100.00%
B	78	--	78	--	--	--	--	--	--	0	100.00%
C	62	--	--	62	--	--	--	--	--	0	100.00%
D	61	--	--	--	59	1	--	1	--	2	96.72%
E	45	--	--	--	--	45	--	--	--	0	100.00%
F	38	--	--	--	--	--	38	--	--	0	100.00%
G	36	--	--	--	--	--	--	36	--	0	100.00%
H	60	--	--	--	--	--	--	--	60	0	100.00%
Total	434	--	--	--	--	--	--	--	--	2	99.54%

4.2 Dominant Style Mechanism Analysis in Blended leaf Formulations

To elucidate the dynamic interplay between the style characteristics of blended leaf formulations and the proportion of primary source leaves, the leaf-centric CNN model was first used to predict the styles of all 304,800 blends. Initial evaluation revealed a 50.91% consistency rate between predicted formulation styles and primary source leaf styles (Table 5), prompting investigation into two potential hypotheses: (1) genuine style transformation arising from multi-regional leaf interactions, or (2) feature extraction limitations in single-leaf style modeling.

Table 5. Consistency Rate Between Formulation Styles Predicted by Leaf-centric CNN and Primary Source Leaf Styles in Blends

Maximum Proportion	Number of Blends	A	B	C	D	E	F	G	H	Average
90%	140	100.00%	100.00%	100.00%	100.00%	100.00%	100.00%	100.00%	99.29%	99.91%
80%	560	99.82%	100.00%	100.00%	100.00%	100.00%	99.64%	100.00%	94.64%	99.26%
70%	1680	99.40%	98.27%	99.76%	98.63%	98.81%	87.44%	98.81%	57.14%	92.28%
60%	4200	92.14%	91.88%	98.21%	92.00%	93.90%	61.52%	88.10%	20.74%	79.81%
50%	9100	67.14%	74.13%	93.43%	74.42%	79.76%	35.03%	65.01%	4.99%	61.74%
40%	14560	37.11%	42.03%	89.06%	47.81%	61.25%	15.65%	43.02%	0.49%	42.05%
30%	7860	15.41%	12.39%	88.49%	19.39%	43.65%	4.40%	25.64%	0.00%	26.17%
Total	38100	49.77%	52.62%	91.67%	56.37%	68.01%	27.73%	53.16%	7.94%	50.91%

To discriminate between these hypotheses, a proportional regression model based on CNN was constructed by adding an additional convolutional layer to the baseline architecture (Figure 2), reducing the output dimension of the fully connected layer to 1, and removing activation functions. Regional proportion prediction across all blends demonstrated robust performance with mean accuracy of 82.17%, and prediction errors remained within ±10% (Table 6). Lower accuracy was observed for Region C (75.83%) and Region F (75.48%), while higher accuracy was achieved for Region B (90.74%) and Region H (89.68%), indicating differences in style retention across regions. Additionally, the integrated regression model achieved 84.75% style consistency between predicted formulation styles and primary source leaf styles. These results confirm that the primary source leaves maintain stylistic dominance despite blending interactions, otherwise the accurate prediction of regional proportions and identification of primary source leaves in blends would not be achievable.

Table 6. Performance of Proportional Regression Model (Within ±10% Error Range)

Designed Proportion	Predicted Proportion Range	Percentage of Blends within Predicted Range
Designed Proportion	Predicted Proportion Range	A	B	C	D	E	F	G	H
90%	80%-100%	97.86%	97.86%	80.71%	98.57%	93.57%	85.00%	92.14%	98.57%
80%	70%-90%	96.79%	96.25%	81.07%	95.00%	92.50%	81.25%	86.79%	99.64%
70%	60%-80%	93.15%	92.56%	72.86%	88.27%	87.38%	76.49%	83.75%	97.98%
60%	50%-70%	89.38%	90.24%	69.90%	84.64%	82.17%	73.50%	78.19%	94.81%
50%	40%-60%	86.85%	88.03%	67.53%	82.35%	79.25%	73.07%	75.52%	90.36%
40%	30%-50%	84.26%	86.27%	66.47%	80.80%	76.45%	72.45%	73.55%	87.01%
30%	20%-40%	77.46%	83.36%	66.16%	76.96%	71.55%	69.95%	68.44%	82.43%
20%	10%-30%	73.05%	84.46%	71.34%	74.82%	69.81%	70.64%	69.59%	79.88%
10%	0-20%	73.47%	95.06%	93.17%	78.43%	74.90%	85.04%	84.01%	84.21%
0%	0-10%	66.28%	93.34%	89.11%	71.38%	67.14%	67.40%	80.03%	81.94%
Average		83.85%	90.74%	75.83%	83.12%	79.47%	75.48%	79.20%	89.68%
Total average		82.17%

4.3 Development and Validation of Fusion Style Model

To overcome the low accuracy (50.91%) of the leaf-centric CNN regional style model in formulation prediction, we developed a fusion style model unifying feature representation across 434 tobacco leaves and 304,800 blend formulations. The model architecture operationalizes the empirically validated hypothesis that blend styles inherit dominance from their primary source leaf origin characteristics, as established in Section 4.2.

When applying this fusion model to region styles prediction for 434 tobacco leaf samples (see Table 7), the average prediction accuracy reaches 90.09%, confirming robust feature extraction capabilities in leaf regional style.

Table 7. Prediction Results of Fusion Style Model on 434 Tobacco Samples

True Production Region	Sample Size	Predicted Production Region								Errors	Accuracy
True Production Region	Sample Size	A	B	C	D	E	F	G	H	Errors	Accuracy
A	54	50	3	--	--	--	--	--	1	4	92.59%
B	78	--	76	--	1	--	1	--	--	2	97.44%
C	62	--	--	54	--	4	3	1	--	8	87.10%
D	61	--	--	--	58	1	--	--	2	3	95.08%
E	45	--	1	--	1	41	1	--	1	4	91.11%
F	38	--	2	2	1	--	33	--	--	5	86.84%
G	36	--	7	--	--	--	1	25	3	11	69.44%
H	60	2	--	--	--	--	--	4	54	6	90.00%
Total	434	52	89	56	61	46	39	30	61	43	90.09%

The integrated style model was applied to predict blended leaf formulation styles, as shown in Table 8. The results demonstrate that the predicted formulation styles strongly align with the primary source leaf styles, achieving an average consistency rate of 87.90%. This significantly outperforms the leaf-centric style model (50.91%) and slightly exceeds the regression model (84.75%). Key findings include:

(1) Threshold Effect: When the primary source leaf proportion ≥40%, the average style consistency exceeds 87.50%.

(2) Region-Specific Retention: Regions B and H exhibit superior style preservation capabilities, maintaining >93% consistency even at a 40% proportion, significantly outperforming other regions.

(3) Nonlinear Decay: As the primary source leaf proportion decreases from 90% to 30%, average consistency nonlinearly declines from 99.91% to 67.90%.

Table 8. Consistency Rate Between Formulation Styles Predicted by Fusion Model and Primary Source Leaf Styles in Blends

Maximum Proportion	Number of Blends	A	B	C	D	E	F	G	H	Average
90%	140	100.00%	100.00%	100.00%	100.00%	100.00%	100.00%	100.00%	99.29%	99.91%
80%	560	100.00%	100.00%	100.00%	100.00%	100.00%	100.00%	100.00%	100.00%	100.00%
70%	1680	100.00%	100.00%	99.94%	100.00%	100.00%	100.00%	100.00%	100.00%	99.99%
60%	4200	99.88%	99.95%	99.31%	99.90%	99.57%	99.81%	99.88%	100.00%	99.79%
50%	9100	98.20%	99.47%	94.41%	97.97%	94.20%	96.33%	97.41%	99.52%	97.19%
40%	14560	88.97%	94.38%	82.21%	88.79%	79.44%	85.46%	87.54%	93.19%	87.50%
30%	7860	66.56%	73.69%	66.93%	66.34%	61.48%	67.42%	70.66%	70.13%	67.90%
Total	38100	88.44%	92.29%	84.97%	88.28%	82.76%	86.82%	88.55%	91.12%	87.90%

The fusion model’s predictive superiority validates its capacity to resolve the feature representation between leaf and blended systems. These results systematically confirm that blend styles constitute weighted integrations rather than emergent properties, with primary source leaves governing stylistic trajectories through quantifiable composition thresholds.

4.4 Style Modulation Strategy in Blended Leaf Formulations

Analysis of the 36,867 style mismatch cases (12.1% of total formulations) revealed critical patterns in compositional-stylistic relationships (Table 9). Detailed analysis of these mismatched cases revealed that 87.5% (32,246 blends) occurred when the proportion difference between primary and secondary source regions fell within 10%. Furthermore, 86.5% of errors (31,902 blends) involved misclassification into secondary source region styles, highlighting the sensitivity of style outcomes to compositional balance.

Table 9. Data of Mismatches Between Predicted and Designed Primary Source Regions in Blends

Proportion Difference Between Primary and Secondary Sources	Total Blends	Mismatched Blends	Mismatch Rate	Classified as Secondary Source
80%	1120	1	0.09%	0
70%	3360	0	0.00%	0
60%	6720	0	0.00%	0
50%	12320	1	0.01%	1
40%	24640	11	0.04%	8
30%	47040	269	0.57%	239
20%	80800	4339	5.37%	3716
10%	128800	32246	25.04%	27938
Total	304800	36867	12.10%	31902

These findings inform a dual-path regulatory framework for precision blend design:

A. Primary Source Style Preservation Protocol

(1) Composition Differential Threshold: Maintain ≥10% proportion advantage over secondary sources.

(2) Risk Mitigation: Reduces style deviation probability to <5%.

(3) Dominance Assurance: Blends with ≥30% proportion difference exhibit high style fidelity.

B. Dual-Source Style Retention Strategy

(1) Controlled Synergy: Restrict primary-secondary proportion differences to <10%.

(2) Hybrid Expression: Enables balanced integration of regional characteristics.

The framework demonstrates particular efficacy in addressing region-specific interactions. Regions B and H exhibit superior style retention, while regions C and F demonstrate reduced preservation capacity attributable to enhanced cross-component interactions in their phytochemical profiles. By quantifying these relationships, the study establishes a predictive foundation for converting compositional parameters into predefined style outputs, transforming blend design from empirical experimentation to computational optimization. This advances the field toward precision engineering of tobacco formulations with controlled stylistic outcomes.

5.1 Design Requirements

Facing supply chain constraints for “Characteristic Tobacco leaf 1” from Region D, this study developed a sustainable modular blend replicating its stylistic profile under three constraints: 1) Mandatory 30% composition from Region B; 2) Limited use of scarce regions (A, E, H); 3) Region D content capped at 50%.

5.2 Implementation Framework

A three-tiered optimization protocol was executed:

1. Style Anchoring: The fusion model identified Tobacco 1 as a definitive Region D style exemplar (probability peak=1.0), establishing baseline characteristics.

2. Constraint-Based Optimization: Candidate blends were filtered through:

a) Style fidelity threshold (>0.99 Region D probability)

b) Compositional requirements (B=30%, D=40%, Δprimary-secondary=10%, minimizing the primary-secondary region proportion difference to reduce reliance on Region D)

c) Chemical compliance (±0.5% total sugar variance from Tobacco 1,ensuring comparable quality grades )

3. Experimental Verification: Final candidate validation through analytical and sensory evaluation

5.3 Optimization Outcome

The system identified Blend No. 25,703 (D=40%, B=30%, C=10%, F=10%, G=10%) as meeting all constraints. This blend exhibited Region D’s characteristic style with a probability peak value of 0.999.

5.4 Analytical and Sensory Validation

The design formula for Blend No. 25,703 was retrieved, and samples were prepared accordingly. Analytical validation demonstrated strong equivalence between the modular blend and Tobacco 1, with thermal gravimetric analysis (DTG) curves showing a normalized root mean square error (NRMSE) of 2.6%, well within the natural variation range observed for Region D leaves (2.5%-4.9%), as shown in Figure 3. Chemical profiling revealed minimal deviations in critical components all remained within equivalent quality specifications, as shown in Table 10. Complementing these findings, the fusion model recalculated a 0.990 probability of Region D style alignment, confirming sustained stylistic fidelity despite compositional modifications.

Sensory evaluation results indicated that the modular blend exhibits style characteristics highly similar to Tobacco Leaf 1. Additionally, certain quality defects observed in Tobacco Leaf 1 were balanced in the modular blend, achieving improved aftertaste and mouthfeel characteristics.

This multi-method verification establishes that the modular blend preserves both the physicochemical signature and regional stylistic essence of the target material.

Table 10. Comparison of Conventional Chemical Components Among Tobacco Leaves in Region D and Modular Blend

Sample Information	Total Nitrogen (%)	Nicotine (%)	Total Sugars (%)	Reducing Sugars (%)	Potassium (%)	Chlorine (%)
Tobacco Leaf 1	1.50	2.14	26.18	23.64	2.38	0.13
Tobacco Leaf 2	1.62	2.15	25.87	23.73	2.45	0.20
Tobacco Leaf 3	1.69	1.97	25.22	22.78	2.60	0.15
Modular Blend	1.67	2.12	26.82	23.91	2.50	0.18

5.5 Application Outcomes

The modular blend was applied to a commercial tobacco formulation to replace Tobacco Leaf 1 in the original blend. Post-replacement analyses demonstrated stability in conventional chemical components and mainstream smoke composition. Sensory smoking tests confirmed that the product met all replacement requirements.

This case establishes a replicable template for targeted material substitution through computational style engineering, successfully decoupling product characteristics from specific geographic dependencies while maintaining commercial viability.

This study establishes a computational framework for leaf and blended formulation styles characterization through multi-source data fusion and deep learning architecture optimization. The main conclusions are as follows:

The complementary nature of thermogravimetric features and chemical indicators enables unprecedented classification accuracy (99.54%) in leaf region style identification when processed through enhanced CNN architecture. This data synergy proves the synergistic role of multi-source data in style evaluation.

Mechanistic analysis reveals hierarchical style inheritance governed by composition thresholds. The developed fusion model successfully bridges the representational gap between single-source and blended materials, achieving 90.09% regional style prediction accuracy while maintaining 87.90% style consistency across formulations.

Practically, the established 10% composition differential threshold enables predictive blend engineering, demonstrated through successful development of a Region D-style equivalent using alternative materials. This modular approach reduced dependency on scarce geographical sources by 60% while maintaining thermal behavior (2.6% NRMSE) and chemical profile equivalency.

While the current framework focuses on thermo-chemical data streams, future extensions could incorporate spectroscopic and sensometric dimensions to create a fully integrated style engineering system. The demonstrated success in balancing material constraints with stylistic fidelity provides a template for data-driven transformation of tobacco processing industries, particularly in addressing supply chain vulnerabilities through computational material substitution strategies.

Acknowledgments

The authors acknowledge China Tobacco Fujian Industrial Co., Ltd. for providing samples and data support in this study.

Authors’ Contributions

Di Wu designed the research, performed data analysis, and drafted the manuscript. Zhongli Ye conceived the study and revised the manuscript. Hui Liang, Qian Gu, Guohua Cai, Yan Lin collected tobacco samples and provided data support. Yiling Chen prepared figures and contributed to manuscript writing. Zechun Liu,Wei Xie, Di Wang, supervised the research and oversaw the project. Qiaoling Li developed the idea for the study, set up the methodology and edited the manuscript. All authors reviewed and approved the final manuscript.

Funding

This work was supported by the Major Science and Technology Program of China National Tobacco Corporation (No. 110202201054(SJ-04)) and the Science and Technology Project of China Tobacco Fujian Industrial Co., Ltd (No. FJZYKJJH2023ZD003)

Availability of Data and Materials

All data generated or analyzed during this study are included in this published article.

Ethics Approval and Consent to Participate

Not applicable.

Consent for Publication

Not applicable.

Competing Interests

The authors declare no competing financial or non-financial interests.

Author Details

¹ Technology Center, China Tobacco Fujian Industrial Co., Ltd., Xiamen 361021, Fujian, China

² Zhengzhou Tobacco Research Institute of CNTC, Zhengzhou, 450001, Henan, China

Xueyi Q , Bing W , Yuping M A ,et al.Method developed for sensory evaluation of quality and style characteristics of flue-cured tobacco leaf[J].Tobacco ence & Technology, 2014, 50(1):74-79.Available from: http://dx.doi.org/10.3969/J.ISSN.1002-0861.2014.09.001.
Changyun Z, Junxiong S, Heng Z, et al. Efect of Guizhou Mountainous Bouquet＆Sweet Tobacco leaves on Quality of Cigarette Brands Style[J]. Chinese Tobacco Science, 2015, 36(2):6.Available from: http://dx.doi.org/10.13496/j.issn.1007-5119.2015.02.003.
Jialei J, Yang L, Yan S, et al. Chemical characteristics and usability of flue-cured tobacco growing in main tobacco producing areas in China[J]. Acta Tabacaria Sinica,2017,23(2):13-27.Available from: http://dx.doi.org/10.16472/j.chinatobacco.2016.123.
Jinfeng H, Xiuxiu Y, Xiaohui X, et al. Chemical quality differences in leaf tobacco of same variety grown in different ecological regions[J]. Acta Tabacaria Sinica, 2016, 22(5):9. Available from: http://dx.doi.org/10.16472/j.chinatobacco.2016.170.
Jianqiang Z, Weijuan L, HouY, et al. Sparse Representation Classification of Tobacco Leaves Using Near-Infrared Spectroscopy and a Deep Learning Algorithm[J].Analytical Letters, 2018, 51(7a9).Available from: http://dx.doi.org/10.1080/00032719.2017.1365882.
Ruidong L, Wenyong H, Guanlan S, et al. Rapid Recognizing the Producing Area of a Tobacco Leaf Using Near-Infrared Technology and a Multi-Layer Extreme Learning Machine Algorithm[J]. Journal of the Brazilian Chemical Society, 2022, 33(3):251-259. Available from: http://dx.doi.org/10.21577/0103-5053.20210143.
Yin C, Deng X, Yu Z. et al. Auto-classification of biomass through characterization of their pyrolysis behaviors using thermogravimetric analysis with support vector machine algorithm: case study for tobacco[J]. Biotechnol Biofuels,2021,14:106. Available from: https://doi.org/10.1186/s13068-021-01942-w.
Heng G, Le W, Jiancai Q, et al. Macro thermogravimetric analysis and its application in identifyingpyrolysis characteristics of different tobacco[J]. Tobacco Science & Technology, 2022, 55(4):58-69. Available from: http://dx.doi.org/10.16135/j.issn1002-0861.2021.0732.
Guru D S, Mallikarjuna P B, Manjunath S, et al. Machine Vision Based Classification Of Tobacco Leaves For Automatic Harvesting [J]. Intelligent Automation & Soft Computing, 2012, 18(5): 581–590.Available from: http://dx.doi.org/10.1080/10798587.2012.10643267.
Kim E-H, Wang Z, Zong H, et al. Design of Tobacco Leaves Classifier Through Fuzzy Clustering-Based Neural Networks With Multiple Histogram Analyses of Images [J]. IEEE Transactions on Industrial Informatics, 2024, 20(3): 4698-4709. Available from: http://dx.doi.org/10.1109/TII.2023.3326533.
Wang D, Li G, Zhen H, et al. Intelligent Classification of Tobacco Leaves based on Residual Network [C], 2022 IEEE International Conference on Artificial Intelligence and Computer Applications, Dalian, China, 2022:156-159. Available from: http://dx.doi.org/10.1109/ICAICA54878.2022.9844494.
State Tobacco Monopoly Administration. Tobacco and Tobacco Products-Preparation of Test Samples and Determination of Moisture Content-Oven Method: YC/T 31-1996 [Standard]. Beijing: China Standards Press, 1996.
State Tobacco Monopoly Administration. Tobacco and Tobacco Products-Determination of Total Nitrogen-Continuous Flow Method: YC/T 161-2002 [Standard]. Beijing: China Standards Press, 2002.
State Tobacco Monopoly Administration. Tobacco and Tobacco Products-Determination of Total Alkaloids-Continuous Flow Method: YC/T 160-2002 [Standard]. Beijing: China Standards Press, 2002.
State Tobacco Monopoly Administration. Tobacco and Tobacco Products-Determination of Chlorine-Continuous Flow Method: YC/T 162-2011 [Standard]. Beijing: China Standards Press, 2011.
State Tobacco Monopoly Administration. Tobacco and Tobacco Products-Determination of Potassium-Continuous Flow Method: YC/T 217-2007 [Standard]. Beijing: China Standards Press, 2007.
State Tobacco Monopoly Administration. Tobacco and Tobacco Products-Determination of Water-Soluble Sugars-Continuous Flow Method: YC/T 159-2019 [Standard]. Beijing: China Standards Press, 2019.
Baker R R, Bishop L J. The pyrolysis of tobacco Ingredients [J]. Journal of Analytical and Applied Pyrolysis, 2004, 71(1): 223–311.Available from: http://dx.doi.org/10.1016/S0165-2370(03)00090-1.
Sung Y J, Seo Y B. Thermogravimetric study on stem biomass of Nicotiana tabacum[J].Thermochimica Acta, 2009, 486(1):1-4.Available from: http://dx.doi.org/10.1016/j.tca.2008.12.010.
Oja V , Hajaligol M R , Waymack B E .The vaporization of semi-volatile compounds during tobacco pyrolysis[J].Journal of Analytical & Applied Pyrolysis, 2006, 76(1/2):117-123.Available from: http://dx.doi.org/10.1016/j.jaap.2005.08.005.
Guo G , Liu X , Li R ,et al.Characterization of tobacco stalk lignin using nuclear magnetic resonance spectrometry and its pyrolysis behavior at different temperatures[J].Journal of Analytical and Applied Pyrolysis, 2019, 142:104665.Available from: http://dx.doi.org/10.1016/j.jaap.2019.104665.
Wu W X, Mei Y F, Zhang L, et al. Kinetics and reaction chemistry of pyrolysis and combustion of tobacco waste[J]. Fuel, 2015, 156: 71-80. Available from: http://dx.doi.org/10.1016/j.fuel.2015.04.016
Yang H , Yan R , Chen H ,et al.Characteristics of Hemicellulose, Cellulose and Lignin Pyrolysis[J].Fuel, 2007, 86(12-13):1781-1788.Available from: http://dx.doi.org/10.1016/j.fuel.2006.12.013.
Yi Z, Qiang L, Zheng W, et al. A classification and regression decision tree method-based model for identifying tobacco growing areas[J]. Acta Tabacaria Sinica,2014,6:28-33. Available from: http://dx.doi.org/10.3969/j.issn.1004-5708.2014.06.005.
Song Q, Zhengyin W, Junxiong S. Study on Quality Character of Leaf for Different Types of Aroma in Guizhou Province[J]. Scientia Agricultura Sinica, 2006, 39(11):2319-2326.Available from: http://dx.doi.org/10.3321/j.issn:0578-1752.2006.11.023.
Jianjun H, Guangcai L, Yaoguang L, et al. Analysis of Nonlinear Relationships Between Routine Chemical Components and Sensory Evaluation Indexes of Flue-cured Tobacco Based on Generalized Additive Models[J].Tobacco Science & Technology,2014,329(12):36-42.Available from: http://dx.doi.org/10.3969/j.issn.1002-0861.2014.12.008.
Lin X, Qifa Z, Xuejun J, et al. Correlation of chemical components with overall sensory quality of tobacco leaves in south Anhui[J].Tobacco Science & Technology,2016,49(11):26-32.Available from: http://dx.doi.org/10.16135/j.issn1002-0861.2015.0641.
Jianwei W, Hailun L, Weidong D, et al. Characteristics of main chemical components in flue-cured tobacco leaves from areas around Qinling mountains and their relationship with cigarette smoking quality[J].Tobacco Science & Technology,2016,49(02):7-13.Available from: http://dx.doi.org/10.16135/j.issn1002-0861.20160202.
Yujun W, Jiheng Z, Qiang L, et al. Chemical Component Usability of Qujing Tobacco Leaves and its Influences on Sensory Quality[J].Tobacco Science & Technology,2014,328(11):67-73.Available from: http://dx.doi.org/10.3969/j.issn.1002-0861.2014.11.014.
Naijie C, Yungui Z, Qingli L, et al.Relationships between routine leaf chemical components in typical medium flavor flue-cured tobacco and mainstream cigarette smoke components by SPSS[J].Tobacco Science & Technology,2017,50(04):31-36.Available from: http://dx.doi.org/10.16135/j.issn1002-0861.2016.0030.
Yuzhen X, Yi W,Dingrong M, et al. Differences of Aroma Note Between Fujian and Yunnan Flue-cured Tobacco and Their Relationship with Chemical Components[J].Tobacco Science & Technology,2015,48(06):68-72.Available from: http://dx.doi.org/10.16135/j.issn1002-0861.20150613.
Houlong J, Honghui M, Yong L, Najia L, Yan Z, Lina G. Inter-annual Variance of Flue-cured Tobacco Quality and Analysis Factors of Sensory Quality in Chongqing Tobacco Districts[J].Chinese Agricultural Science Bulletin. 2014, 30(19): 223-230. Available from: https://doi.org/10.11924/j.issn.1000-6850.2013-3410.
China Tobacco Fujian Industrial Co., Ltd. "Tobacco Component Detection Method and Mixing Uniformity Discrimination Method: CN110441187B" [Patent]. 2022-02-15.
China Tobacco Fujian Industrial Co., Ltd. "Tobacco Matching Method and System: CN107860868B" [Patent]. 2020-09-18.
Breiman, L. Random Forests[J]. Machine Learning.2001, 45:5–32. Available from: https://doi.org/10.1023/A:1010933404324.
Cover T M , Hart P E .Nearest neighbor pattern classification[J].IEEE Trans.inf.theory, 1953, 13(1):21-27.Available from: http://dx.doi.org/10.1109/TIT.1967.1053964.
Haitao Z,Zhihui L,Henry L.Linear Discriminant Analysis[J]. Feature Learning and Understanding, 2020, 4:71-85.Available from: http://dx.doi.org/10.1007/978-3-030-40794-05.

No competing interests reported.

From Leaf to Blend: CNN-Enhanced Multi-source Feature Fusion Enables Threshold-Driven Style Control in Digital Tobacco Formulation

Status:

Journal Publication

Version 1

Abstract

Figures

Background

1 Materials and Methods

2 Data Description

3 Algorithmic Framework

4 Results and Analysis

5. Case Study: Modular Blend Design

6. Conclusions

Declarations

References

Additional Declarations

Status:

Journal Publication

Version 1