Development of an Improved Ultrasonic-Assisted Extraction Method for Determining Soybean Oil Content

Abstract

This study aimed to develop and validate an improved ultrasonic-assisted extraction (UAE) method for rapid and accurate determination of soybean oil content to address the limitations of conventional methods, including time-consuming operation, low efficiency, and unstable detection accuracy. Two typical soybean varieties (Jinda 117 and Jinda 114) were selected as research materials. On the basis of single-factor experiments and thermogravimetric analysis, response surface methodology (RSM) and a genetic algorithm-optimized artificial neural network (ANN-GA) were employed to model and optimize the UAE extraction process. The predictive performance and reliability of the RSM and ANN-GA models were systematically validated by coefficient of determination (R ²) and absolute average deviation (AAD). Analysis of the experimental dataset demonstrated that the ANN-GA model exhibited higher predictive accuracy and better fitting performance. Under the optimal conditions predicted by the ANN-GA model, the oil yields of Jinda 117 and Jinda 114 reached 22.17% ± 0.23% and 20.91% ± 0.70%, respectively, which were significantly higher than those obtained by RSM optimization, traditional Soxhlet extraction, and maceration methods. Further validation across 19 soybean cultivars verified the good universality and stability of the proposed UAE method. Scanning electron microscopy (SEM) characterization further demonstrated that ultrasonic treatment effectively destroyed soybean cell wall structures, thereby promoting oil release and improving extraction efficiency. Compared with conventional detection approaches, the established UAE strategy possesses the prominent advantages of high efficiency, stable accuracy, and strong cultivar adaptability. This study provides a reliable and efficient technical approach for the rapid determination of soybean oil content and offers technical support for soybean quality evaluation and industrial quality control.

1. Introduction

As a globally vital oil crop, soybean boasts great economic significance yet faces a severe supply-demand disparity. China recorded 20.65 million tons of soybean output in 2024–2025, far below the domestic consumption of 115 million tons. With dry-based oil content merely ranging from 18% to 24%, efficient oil utilization is critical to stabilize edible oil supply and safeguard food security. Accordingly, precise and rapid detection methods for soybean oil content are urgently required to support the related extraction and quality evaluation research.

Technologies for soybean oil extraction and oil content determination have independently advanced toward high efficiency, low cost, and non-destructive detection. Traditional oil extraction approaches mainly include solvent extraction and mechanical pressings. Solvent extraction gains wide popularity due to outstanding extraction efficiency, whereas it suffers from high running cost, thermal deterioration of oil properties, potential food safety risks, and poor performance for materials with oil content lower than 14%. Mechanical pressing effectively avoids chemical residues and meets the demand for healthy edible oil, featuring low labor consumption and satisfactory economic benefits, yet it is restricted by a high residual oil ratio and unsatisfactory oil yield. Common pretreatment means including crushing, grinding, thermal treatment, and aqueous enzymatic treatment, − as well as infrared pretreatment, can improve extraction performance but cannot fundamentally balance oil recovery rate and residual oil content. In recent years, innovative extraction technologies have made progress in the extraction efficiency and finished oil quality. Subcritical extraction elevates extraction yield by regulating temperature, extraction cycles, and duration and has attained favorable effect in passion fruit seed oil extraction, while its practical application on soybean oil is still immature. Supercritical CO₂ extraction exhibits prominent superiority in extracting soybean germ oil; combined with ultrasonic assistance, its extraction recovery can reach 87.24% and preserve fine oil quality. , Nevertheless, expensive equipment and operating expenditures hinder its large-scale industrial popularization.

Classic analytical detection methods are represented by Soxhlet extraction , and nuclear magnetic resonance (NMR). Both methods possess reliable detection accuracy but demand tedious sample pretreatment and chemical reagent consumption, which makes them incompatible with rapid on-site quality inspection for soybean breeding and industrial production. Modern nondestructive rapid detection techniques have become the development trend of oil content measurement. Hyperspectral imaging (HSI) can collect spatial and spectral information simultaneously. Combined with competitive adaptive reweighted sampling, uninformative variable elimination, and partial least-squares modeling, HSI realizes precise prediction of soybean oil content and achieves lower prediction error than traditional near-infrared detection. Even so, this method relies on complex chemometric calculations, and limited instrument portability and weak anti-interference capability restrict its field application. Hence, simpler and more adaptable detection technologies are still required to break the existing technical bottlenecks.

Multiple variables exert influences on soybean oil extraction efficiency, thus, reliable predictive models are indispensable for parameter optimization. Orthogonal array design can reflect general experimental tendency yet cannot realize data prediction. Response surface methodology and artificial neural networks are widely applied to forecast extraction yield and build process models. , RSM suits conventional linear and weak nonlinear regression analysis, while ANN shows superior performance in fitting intricate nonlinear extraction behaviors and delivers higher prediction precision than traditional regression models. ,

Nevertheless, the standard Soxhlet method suffers from cumbersome operation, long testing cycle, and heavy reagent consumption, failing to meet fast screening demands. , A single ANN is prone to underfitting and overfitting defects caused by unreasonable weight and threshold configuration. To remedy these drawbacks, this study adopts ultrasonic-assisted extraction (UAE) combined with an ANN-GA hybrid strategy. Genetic algorithm optimizes initial network parameters to stabilize model performance and elevate prediction accuracy. Compared with routine detection and modeling means, the integrated UAE-ANN-GA method achieves rapid, high-precision oil content evaluation and efficient extraction parameter optimization, offering a more practical technical scheme for soybean oil determination and process improvement. −

This study applies ultrasonic-assisted extraction to improve the oil extraction efficiency. Combined with RSM and ANN-GA, this establishes an accurate quantification method for soybean oil. Thermodynamic and kinetic analyses reveal extraction mechanisms, and validation using 19 soybean cultivars verifies the method’s reliability, offering an efficient technical reference for rapid oil quantification and batch quality inspection of soybean raw materials. The technical flowchart is presented in Figure .

1.

Technical roadmap of experimental design.

2. Materials and Methods

2.1. Materials

Soybean cultivars (Jinda117, Jinda114, 1657, Jinda88, Jinke21, Jinke2, Fendou103, Jindou56, Fendou93, Yundou303, Jinke23, Jinhei2, Jinke15, Ningdou10, 1365, Flat Black Bean, Jidou12, Large-Seeded Black Bean, Huaixiandou7, Zhonghuang13, and Zhonghuang37) were obtained from Shanxi Agricultural University (Taigu, China). Before oil extraction, the samples were cleaned, ground into powder, and passed through 0.180–0.075 mm sieves. All chemicals and reagents used in this study were of an analytical grade or higher. Slow-speed filter paper was employed for solid–liquid separation.

2.2. Pretreatment Optimization for Moisture Control in Soybean Powder

To prevent oil–water separation resulting from excessive moisture in soybean powder, a pretreatment method that maximizes the dehydration rate was required. Accordingly, two pretreatment approaches were investigated:

2.2.1. Natural Air Drying

Freshly ground soybean powder with an initial weight of 0.5 g was evenly spread on a fixed-quality filter paper with a fixed spreading thickness of 0.20–0.30 mm and placed on a clean laboratory bench for natural air drying. The air-drying experiment was conducted under constant laboratory conditions of 25 °C and 30%–40% relative humidity. The weight loss of soybean powder was recorded at predetermined time intervals of 0, 30, 60, 90, 120, 150, 180, 210, 240, and 270 min to investigate the effect of drying time on the dehydration rate of soybean samples.

2.2.2. Oven Drying

Freshly ground soybean powder with a fixed initial mass of 0.50 g was evenly spread into a 6 cm × 3 cm filter paper package with a uniform spreading thickness of 0.20–0.30 mm. The sealed filter paper package was steadily placed in the oven with flat placement loading mode for drying treatment. To explore the effects of drying conditions on sample dehydration rate, the dehydration performance was determined under gradient drying temperatures (35, 50, 65, 80, 95, and 110°C) and drying durations (10, 20, 30, 40, 50, and 60 min), respectively.

2.3. Optimization of Soybean Oil Content Determination

2.3.1. Ultrasonic-Assisted Extraction

Ultrasonic-assisted extraction (UAE) was carried out by using an ultrasonic instrument (model KM-300DE, Kunshan Meimei Ultrasonic Instrument Co., Ltd., China) with a maximum power of 300 W and a fixed ultrasonic frequency of 40 kHz. A stoppered Erlenmeyer flask was used as the extraction container. The volume of the extraction solvent was calculated and added based on the preset solid–liquid ratio. During the whole ultrasonic treatment, the water temperature of the ultrasonic bath was monitored in real-time with a thermometer, and ice cubes were supplemented appropriately to keep the extraction temperature stable.

Equation was used to calculate the oil yield.

$$\mathrm{Oil\; yield}(\mathrm{\%})=\frac{m_1-m_0}{m_2}\times 100$$ 1

where m ₀ represents the mass of the empty receiving flask (g); m ₁ represents the mass of the receiving flask plus fat after constant weight (g), and m ₂ represents the mass of the test sample (g).

2.3.1.1. Single-Factor Experiment

To assess the impact of various factors on soybean oil yield, single-factor experiments were carried out by extraction time (30–70 min), temperature (40–70 °C), ultrasonic power (150–270 W), solid-to-liquid ratio (1:10–1:50 g/mL), and particle size (0.075–0.250 mm). The parameter ranges were established based on findings from the published literature and our preliminary experimental results. , All experiments were performed in triplicate. For each factor, only the parameter of interest was varied while the remaining parameters were held constant. The optimum basic extraction conditions were as follows: for Jinda117, the mass of dried soybean flour,, 0.50 g, temperature 40 °C, time 30 min, ultrasonic power 180 W, solid-to-liquid ratio 1:50 g/mL, and particle diameter 0.100-0.125 mm; for Jinda114, temperature 60 °C, time 50 min, ultrasonic power 210 W, solid-to-liquid ratio 1:30 g/mL, and particle diameter 0.100-0.125 mm.

2.3.1.2. Response Surface Optimization

Design Expert version 12 was employed to analyze the relationship between the experimental variables and soybean oil yield. Regression coefficients of the model were obtained by fitting the experimental data to eq .

$$Y={\beta }_0+\sum _{n=1}^k{\beta }_n{X}_n+\sum _{n=1}^k{\beta }_{nn}{X}_n^2+\sum _m\sum _{m<n}^k{\beta }_{mn}{X}_m{X}_n+e_m$$ 2

where Y represents the oil yield (%); X _m and X _n represent the coded independent variables, with m and n ranging from 1 to k; β₀, β _n , β _nn , and β _mn represent the regression coefficients associated with the intercept, linear, quadratic, and the second-order terms, respectively; The symbol k represents the number of independent variables (k = 4); and e _m represents the experimental error.

Following the single-factor experimental results, a central composite design (CCD) with four factors and three levels was employed to optimize the UAE process. The experimental factors and levels for Jinda117 were: time (X ₁), 20, 30, and 40 min; temperature (X ₂), 30, 40, and 50 °C; ultrasonic power (X ₃), 150, 180, and 210 W; solid-to-liquid ratio (X ₄), 1:40, 1:50, and 1:60 g/mL. For Jinda114, the corresponding factors and levels were: time (X ₁), 40, 50, and 60 min; temperature (X ₂), 50, 60, and 70 °C; ultrasonic power (X ₃), 180, 210, and 240 W; and solid-to-liquid ratio (X ₄), 1:20, 1:30, and 1:40 g/mL. The response variable in all experiments was the oil yield (Y). Based on the CCD, a total of 29 experimental runs were conducted for each variety, as summarized in Table a (Jinda117) and Table b (Jinda114), respectively.

1. Design and Results of the RSM Experiment for Jinda117 and Jinda114.

(a) Jinda117
Run	Time, X 1 (min)	Temperature, X 2 (°C)	Ultrasonic power, X 3 (W)	Solid-to-liquid ratio, X 4 (g/mL)	Soybean oil yield, Y 1 (%)
1	20	30	180	1:50	12.31 ± 0.68
2	40	30	180	1:50	14.48 ± 0.96
3	20	50	180	1:50	17.67 ± 0.89
4	40	50	180	1:50	18.87 ± 0.15
5	30	40	150	1:40	13.69 ± 0.22
6	30	40	210	1:40	19.14 ± 0.35
7	30	40	150	1:60	13.60 ± 1.05
8	30	40	210	1:60	13.06 ± 0.54
9	20	40	180	1:40	17.61 ± 0.28
10	40	40	180	1:40	13.70 ± 0.50
11	20	40	180	1:60	15.75 ± 0.24
12	40	40	180	1:60	18.45 ± 0.49
13	30	30	150	1:50	12.93 ± 0.17
14	30	50	150	1:50	15.54 ± 0.35
15	30	30	210	1:50	13.77 ± 0.45
16	30	50	210	1:50	17.41 ± 0.27
17	20	40	150	1:50	14.35 ± 0.41
18	40	40	150	1:50	13.31 ± 0.34
19	20	40	210	1:50	18.38 ± 0.75
20	40	40	210	1:50	18.99 ± 0.87
21	30	30	180	1:40	11.90 ± 0.72
22	30	50	180	1:40	18.78 ± 0.28
23	30	30	180	1:60	15.26 ± 0.78
24	30	50	180	1:60	13.75 ± 0.09
25	30	40	180	1:50	22.41 ± 0.83
26	30	40	180	1:50	20.51 ± 0.49
27	30	40	180	1:50	20.99 ± 0.78
28	30	40	180	1:50	22.92 ± 0.44
29	30	40	180	1:50	23.03 ± 0.61

(b) Jinda114
Run	Time, X 1 (min)	Temperature, X 2 (°C)	Ultrasonic power, X 3 (W)	Solid-to-liquid ratio, X 4 (g/mL)	Soybean oil yield, Y 2 (%)
1	40	50	210	1:30	12.85 ± 0.54
2	60	60	210	1:40	12.69 ± 0.42
3	50	70	180	1:30	16.97 ± 0.60
4	40	70	210	1:30	13.13 ± 0.50
5	50	60	210	1:30	20.40 ± 0.62
6	50	70	210	1:40	16.34 ± 0.56
7	50	60	210	1:30	20.99 ± 0.71
8	60	60	180	1:30	16.84 ± 0.82
9	50	60	210	1:30	20.98 ± 0.64
10	50	60	240	1:40	16.86 ± 0.77
11	50	50	240	1:30	14.89 ±0.83
12	40	60	180	1:30	13.42 ± 0.90
13	50	60	210	1:30	21.05 ± 0.82
14	50	60	180	1:20	16.78 ± 0.43
15	60	70	210	1:30	17.20 ± 0.51
16	50	50	210	1:40	13.81 ± 0.60
17	40	60	210	1:20	14.00 ± 0.50
18	50	70	210	1:20	16.39 ± 0.80
19	60	60	210	1:20	15.85 ± 0.64
20	40	60	210	1:40	12.72 ± 0.52
21	40	60	240	1:30	15.1 ± 0.82
22	50	60	240	1:20	16.66 ± 0.52
23	50	70	240	1:30	16.50 ± 0.60
24	50	60	210	1:30	20.17 ± 0.34
25	60	60	240	1:30	16.50 ± 0.56
26	50	60	180	1:40	17.39 ± 0.47
27	60	50	210	1:30	11.67 ± 0.90
28	50	50	180	1:30	15.25 ± 0.62
29	50	50	210	1:20	13.08 ± 0.50

2.3.1.3. Experimental Design for Optimized Soybean Oil Extraction Using the ANN Model and GA Optimization ,

Before ANN-based experimental analysis, the extraction parameters were loaded into MATLAB (R2025a, The MathWorks, Inc., Natick, MA, USA) to perform deep learning. Subsequently, a multilayer feedforward ANN was constructed to map the input and output variables and analyze the oil extraction process. The dataset was divided into training (80%), validation (10%), and test (10%) subsets. Model parameters were subsequently determined to evaluate the predictive and optimization performance.

The trained ANN model was coupled with a genetic algorithm to optimize the extraction parameters for the maximum oil yield. The GA was implemented in MATLAB with the following parameters: population size, 50; selection strategy, roulette wheel selection; crossover rate, 0.8; mutation rate, 0.01; maximum iterations, 30. Termination Criteria: The algorithm stopped when the optimal fitness value remained unchanged for 20 consecutive generations.

The predictive performance of the model was assessed using several statistical metrics, including standard error of prediction (SEP), correlation coefficient (R ²), absolute average deviation (AAD), and root mean square error (RMSE). Increased R ² values alongside decreased SEP, AAD, and RMSE indicate greater reliability of the established model. The calculation methods were provided below:

$$R^2=1-\frac{\sum _{i=1}^n{(x_i-x_{ik})}^2}{\sum _{i=1}^n{(x_{ik}-x_z)}^2}$$ 3

$$\mathrm{SD}=\sqrt{\frac{\sum _{i=1}^n{(x_i-x_z)}^2}{n-k}}$$ 4

$$\mathrm{MSE}=\frac{1}{n}\sum _{i=1}^n{(x_i-x_{ik})}^2$$ 5

$$\mathrm{RMSE}=\sqrt{\frac{1}{n}\sum _{i=1}^n{(x_i-x_{ik})}^2}$$ 6

$$\mathrm{ESS}=\frac{1}{n^2}\sum _{i=1}^n{(x_i-x_{ik})}^2$$ 7

$$\mathrm{AIC}=n\ln (SSE)+2p$$ 8

$$\mathrm{ADD}(\%)=\left[\frac{\sum _{i=1}^n(|x_{ik}-x_i|/x_{ik})}{n}\right]\times 100$$ 9

Here, x _i represents the predicted oil yield; x _z represents the mean of the experimental oil yield; and x _ik represents the experimental or actual oil yield. Parameters n and p denote the number of data points and the number of model parameters, respectively.

GA optimization was implemented in MATLAB to improve extraction parameter selection. The ANN was trained using the Levenberg–Marquardt algorithm and subsequently integrated with GA to optimize input variables and determine the corresponding optimal output values. The key parameters of the GA were defined with a population size of 50 and a crossover rate of 0.85.

2.3.2. Comparative Experiment

Specifically, 2 g of dried soybean powder and 160 mL of petroleum ether were applied for both the Soxhlet extraction (SE) and maceration method (MM). The SE experiment was conducted in strict accordance with the national standard method (GB 5009.6-2016) for soybean oil content determination, with continuous extraction for 6–-10 h and a stable siphon frequency of six cycles per hour to achieve thorough oil elution and obtain the reference true value of the soybean oil content. For the MM group, the soybean powder was fully immersed in petroleum ether with a unified solid-to-liquid ratio and incubated in a constant-temperature oscillator with continuous shaking for 24 h to reach complete extraction equilibrium.

2.4. Characterization

2.4.1. Thermogravimetric-Differential Thermogravimetric (TG-DTG) Analysis

TG-DTG analysis was performed using a NETZSCH STA 449F3 instrument (Hitachi, Ltd., Japan) to determine the decomposition temperatures of soybean powder and oil from cultivars Jinda117 and Jinda114. TG-DTG was employed to assess the thermal stability of the samples and to characterize their decomposition behavior. Samples were uniformly cut into discs with a diameter of 3.6 mm using a punch. Approximately 3–4 mg of each sample was placed in a crucible and heated in an air environment from 25 to 800 °C. The gas flow was controlled at 100 mL/min with a heating rate of 10 °C/min. Data were recorded at intervals of 0.5 °C.

2.4.2. Scanning Electron Microscopy (SEM) Analysis

The changes in the cell wall structure of soybean powder before and after different extraction treatments were analyzed using an SEM system (Model JSM6490LV, JEOL, Japan). Observations were performed at an accelerating voltage (HV) of 20.0 kV, a magnification (MAG) of 20 000×, and a working distance (WD) of 7.08 mm.

2.4.3. Mechanical Testing

The mechanical properties of the samples were measured by using a Sansi tensile testing machine in accordance with the GB/T 528-2009 standard. The length, width, and height of the soybean samples are 6.75–11.10 mm, 4.95–8.22 mm, and 3.09–7.46 mm, respectively. Tensile stress curves of samples under different placement conditions were recorded during the test. The mechanical characteristics directly reflect the structural firmness of the raw material matrix, which further affects the mass-transfer efficiency of oil during extraction. Therefore, analyzing the mechanical properties is of great significance for exploring the extraction mechanism and optimizing the extraction parameters.

2.5. Extraction Thermodynamics and Kinetics

The change in Gibbs free energy (ΔG) of the extraction process and the corresponding equilibrium behavior of soybean oil at 298, 303, 308, 313, and 318 K were calculated using eq . The enthalpy change (ΔH) and entropy change (ΔS) of the extraction process were determined from the slope and intercept, respectively, of the linear relationship between ln K and 1/T (eq ).

$$\mathrm{\Delta }G=-RT\ln K$$ 10

$$\ln K=-\frac{\mathrm{\Delta }H}{\mathrm{R}}\frac{1}{T}+\frac{\mathrm{\Delta }S}{\mathrm{R}}$$ 11

where R represents the molar gas constant (R = 8.314 J mol^–1 K^–1), and T represents the absolute temperature (K).

The kinetics of soybean oil extraction under different ultrasonic power levels (180, 210, 240, and 270 W) were analyzed using both the kinetic models of first and second order, as described in eqs and , respectively.

$$c_t=c_{\infty }(1-\exp (-k_{1}t))$$ 12

$$c_t=\frac{c_{\infty }^2\times k_{2}t}{1+c_{\infty }{k}_{2}t}$$ 13

where c _t represents the oil content at time t, c _∞ represents the equilibrium oil content, and k ₁ (min^–1) and k ₂ [(mg g^–1) min^–1] represent the rate constants for the first-order and second-order kinetic models, respectively.

For Jinda 117 and Jinda 114, the oil yields were determined at ultrasonic powers of 210 and 240 W, respectively, at different temperatures (25, 30, 35, 40, and 45 °C) over extraction durations ranging from 5 to 60 min. If the soybean oil extraction follows the second-order kinetic model, the extraction yield at time t conforms to eq .

$$\frac{t}{c_t}=\frac{1}{k{c}_e^2}+\frac{t}{c_e}$$ 14

Linear regression of $\frac{t}{c_t}$ versus t gives a straight line with a slope of $\frac{1}{c_e}$ and an intercept of $\frac{1}{k{c}_e^2}$ , enabling the calculation of extraction rate constants under various temperatures. In accordance with eq , linear fitting of ln k against 1/T is conducted, and the slope of the regression line is adopted to determine the activation energy for oil extraction.

$$\ln k=\ln A-\frac{E_{\mathrm{a}}}{R}\frac{1}{T}$$ 15

2.6. Statistical Treatment and Data Analysis

Statistical analyses were performed using IBM SPSS Statistics version 22. All measurements were carried out in triplicate. Results were presented as the mean ± standard deviation, with all analyses conducted in triplicate. Significant differences among protein ingredients were analysed using Tukey’s HSD and T-test at a significance level of p-value <0.05. Identical letters indicate no significant difference; different letters denote significant differences. Oil content was used as the response variable in single-factor experiments evaluating the effects of ultrasonic duration, ultrasonic temperature, ultrasonic power, and liquid-to-material ratio. Experimental design and multivariate data analysis were conducted using Design-Expert version 12 software. An ANN model for predicting soybean oil content was developed using the ANN fitting module in MATLAB R2024a. Based on the experimental results, statistical analyses were performed to assess the significance of differences in oil content obtained by ultrasonic-assisted, immersion, and Soxhlet extraction methods. OriginPro version 8.5.1 was used for data visualization.

3. Results and Discussion

3.1. Effect of Drying Method on the Moisture Content of Defatted Soybean Flour

Soybean moisture content significantly affects oil extraction yield. Oven drying exhibited a faster moisture removal rate than natural air drying (25 °C) for soybean powder of both Jinda117 (Figure a, b) and Jinda114 (Figure c, d). To avoid the limitation of relying solely on dehydration efficiency for parameter selection, the thermal stability of soybean powder and internal oil was further evaluated based on the thermogravimetric analysis (Figure ) and thermal decomposition temperature data (Table ).

2.

Effect of natural air drying ((a) Jinda117, (c) Jinda114) and oven drying ((b) Jinda117, (d) Jinda114) on the moisture content of soybean flour.

3.

TG-DTG curves of soybean flour (panels (a), (c)) and soybean oil (panels (b), (d)) from Jinda117 (panels (a), (b)) and Jinda114 (panels (c), (d)).

2. Thermal Decomposition Temperatures of Soybean Flour and Soybean Oil from Jinda117 and Jinda114.

	Jinda117		Jinda114
Stage	Decomposition temperature of soybean flour (°C)	Decomposition temperature of soybean oil (°C)	Decomposition temperature of soybean flour (°C)	Decomposition temperature of soybean oil (°C)
1	105	363	114	353
2	228	419	230	421
3	305	452	307	530
4	477	526	483	–
5	–	–	517	–

Table shows that the initial thermal decomposition temperature of soybean lipid components is much higher than 95 °C. Consistent with this result, the thermogravimetric curves in Figure confirm that no thermal weight loss, decomposition, or structural damage of soybean oil occurs at 95 °C. Low-temperature drying fails to completely remove bound water and may induce lipid oxidation, while excessive high-temperature heating causes decomposition of unsaturated fatty acids (soybeans contain over 60% unsaturated fatty acids, and the thermal decomposition temperature of unsaturated fatty acids is above 160°C) and oil loss. In contrast, 95 °C drying thoroughly removes residual moisture while maintaining the original physicochemical properties and structural stability of soybean oil.

Combining the high dehydration efficiency (Figure ) and reliable oil thermal stability (Figure and Table ), the pretreatment condition of 95 °C for 1 h was determined as the optimal parameter. Soybeans were pulverized and oven-dried at 95 °C for 1 h to prepare dry-based soybean powder. This pretreatment balances dehydration effect and oil stability, eliminates experimental interference from residual moisture and lipid deterioration, and guarantees the accuracy and repeatability of subsequent oil extraction experiments.

3.2. Single-Factor Experiments of Process Variables Affecting Oil Content in Jinda117 and Jinda114

TG-DTG analysis of soybean flour and oil from the cultivars Jinda117 and Jinda114 (Figure and Table ) showed minimum decomposition temperatures of 105 °C for soybean flour and 353 °C for soybean oil. These results indicated that soybean oil extraction conducted at temperatures below 90 °C have negligible effects on sample composition.

Figure a–j shows that the ultrasonic time, temperature, power, liquid–solid ratio, and particle size exert consistent first-rise-then-fall effects on the oil yields of Jinda117 and Jinda114, governed by ultrasonic cavitation and interfacial mass transfer. Mild ultrasound breaks cell walls and facilitates lipid diffusion, while excessive ultrasonic intensity induces lipid oxidation; too high liquid-solid ratio weakens cavitation. A particle size of 0.100–0.125 mm achieves optimal extraction via moderate specific surface area. Detailed single-factor optimal parameters and peak yields for both varieties are provided in Figure a–j. Slight differences in optimal conditions and maximum yields are attributed to their distinct microstructures and lipid characteristics.

4.

Single-factor experiments evaluating the effects of ultrasonic time, ultrasonic temperature, ultrasonic power, solid-to-liquid ratio, and particle size of the soybean flour on oil content in Jinda117 (a–e) and Jinda114 (f–j).

3.3. Optimization Using RSM

As shown in Figure , the residuals for both the Jinda117 and Jinda114 models are closely aligned with the reference diagonal line, indicating that they approximately follow a normal distribution. This observation suggests that the assumptions underlying the regression analysis are reasonably satisfied, supporting the validity of the statistical inferences derived from the models.

5.

Normal probability plots of residuals for Jinda117 (a) and Jinda114 (b).

Based on the results of the single-factor experiments, ultrasonic time, ultrasonic temperature, ultrasonic power, and solid-to-liquid ratio were selected as independent variables for the CCD in the RSM analysis (Table a and b). Quadratic regression models were established and evaluated by using ANOVA to determine their significance and adequacy. The resulting regression equations describing the oil yields of Jinda117 (Y ₁) and Jinda114 (Y ₂) are presented below.

$$Y_1=-334.40846+0.405162X_1+4.08793X_2+1.61964X_3+4.62798X_4-0.002425X_1{X}_2+0.001377X_1{X}_3+0.016545X_1{X}_4+0.000863X_2{X}_3-0.020957X_2{X}_4-0.004992X_3{X}_4-0.022817{X_1}^2-0.036808{X_2}^2-0.003883{X_3}^2-0.034288{X_4}^2$$

$$Y_2=-287.17342+3.77353X_1+3.41300X_2+0.723055X_3+1.99538X_4+0.013113X_1{X}_2-0.001721X_1{X}_3-0.004688X_1{X}_4-0.000094X_2{X}_3-0.001961X_2{X}_4-0.000346X_3{X}_4-0.039792{X_1}^2-0.032211{X_2}^2-0.001479{X_3}^2-0.026585{X_4}^2$$

3. (a) ANOVA Results for the RSM Model of Jinda117; (b) ANOVA Results for the RSM Model of Jinda114.

(a) ANOVA Results for the RSM Model of Jinda117
Source	Sum of Squares	df	Mean Square	F-value	p-value
Model	285.46	14	20.39	10.18	<0.0001	significant
time, X 1	0.2469	1	0.2469	0.1233	0.7307
temperature, X 2	38.04	1	38.04	19	0.0007
power, X 3	25.01	1	25.01	12.49	0.0033
liquid-to-material ratio, X 4	2.04	1	2.04	1.02	0.3304
X 1 X 2	0.2352	1	0.2352	0.1175	0.7369
X 1 X 3	0.6829	1	0.6829	0.3411	0.5685
X 1 X 4	10.95	1	10.95	5.47	0.0347
X 2 X 3	0.2681	1	0.2681	0.1339	0.7199
X 2 X 4	17.57	1	17.57	8.77	0.0103
X 3 X 4	8.97	1	8.97	4.48	0.0527
X 1 2	33.77	1	33.77	16.87	0.0011
X 2 2	87.88	1	87.88	43.89	<0.0001
X 3 2	79.21	1	79.21	39.56	<0.0001
X 4 2	76.26	1	76.26	38.09	<0.0001
Residual	28.03	14	2
Lack of Fit	22.73	10	2.27	1.72	0.3178	not significant
Pure Error	5.3	4	1.32
Cor Total	313.49	28
R 2	0.9106
Adjusted R 2	0.8212
C.V. %	8.5
Adeq Precision	10.3906

(b) ANOVA Results for the RSM Model of Jinda114
Source	Sum of Squares	df	Mean Square	F-value	p-value
Model	195.39	14	13.96	21.89	<0.0001	significant
time, X 1	7.49	1	7.49	11.74	0.0041
temperature, X 2	18.65	1	18.65	29.25	<0.0001
X 3-power	0.0013	1	0.0013	0.0021	0.9644
X 4 liquid-to-material ratio	0.7219	1	0.7219	1.13	0.3053
X 1 X 2	6.88	1	6.88	10.79	0.0054
X 1 X 3	1.07	1	1.07	1.67	0.2169
X 1 X 4	0.8793	1	0.8793	1.38	0.2599
X 2 X 3	0.0032	1	0.0032	0.005	0.9445
X 2 X 4	0.1539	1	0.1539	0.2413	0.6309
X 3 X 4	0.0431	1	0.0431	0.0676	0.7986
X 1 2	102.71	1	102.71	161.07	<0.0001
X 2 2	67.3	1	67.3	105.55	<0.0001
X 3 2	11.5	1	11.5	18.03	0.0008
X 4 2	45.85	1	45.85	71.9	<0.0001
Residual	8.93	14	0.6376
Lack of Fit	8.28	10	0.8282	5.13	0.0645	not significant
Pure Error	0.6454	4	0.1613
Cor Total	204.31	28
R 2	0.9563
Adjusted R 2	0.9126
C.V. %	4.96
Adeq Precision	15.6163

ANOVA was adopted to assess the statistical significance of the regression models. Single-variable effects on the oil yield were analyzed using F-values and corresponding p-values. Distinct nonlinear correlations among experimental variables rendered linear models unsuitable, hence, the quadratic model was selected for its optimal tradeoff between fitting accuracy and model conciseness. Low p-values and high R ² values from ANOVA further validated that this quadratic model is statistically significant and possesses excellent data-fitting capacity. Results showed that the models for Jinda117 and Jinda114 were found to be highly significant (p < 0.0001), whereas the lack-of-fit was not significant (p > 0.05), indicating good model fitness. For the oil yield of Jinda117, variables X ₂, X ₂ ², X ₃ ², and X ₄ ² were highly significant (p < 0.001), whereas X ₃, X ₁ X ₄, X ₂ X ₄, and X ₁ ² showed significant effects (p < 0.05). The remaining variables were not statistically significant (p > 0.05). Similarly, for Jinda114, X ₂, X ₁ ², X ₂ ², and X ₄ ² were highly significant (p < 0.001), while X ₁, X ₁ X ₄, and X ₃ ² were significant (p < 0.05). All other terms had no significant effects on oil yield (p > 0.05).

Based on the ANOVA significance and regression coefficient magnitude, the order of the dominant factors affecting the oil yield of Jinda117 was ultrasonic temperature (X ₂) > ultrasonic power (X ₃) > solid-to-liquid ratio (X ₄) > ultrasonic time (X ₁), while that of Jinda114 presented a different order: ultrasonic temperature (X ₂) > ultrasonic time (X ₁) > solid-to-liquid ratio (X ₄) > ultrasonic power (X ₃). Notably, the ultrasonic temperature acted as the most critical and universal influencing factor for both soybean cultivars, exhibiting significant quadratic effects on oil extraction. The obvious differences in factor sensitivity and influence order between the two cultivars were derived from their inherent varietal differences in tissue compactness (see Figure ), cell wall structure (see Figure ) and endogenous lipid distribution. The inherent differences in tissue compactness, cell wall structure, and oil distribution result in varying time sensitivities: Jinda 114 has denser cell walls, requiring a longer sonication time for cell disruption, while Jinda 117 is more susceptible to ultrasound power. These botanical characteristics resulted in discrepant responses of the two soybean varieties to ultrasonic cavitation intensity, thermal diffusion efficiency, and solvent mass transfer behavior, ultimately leading to inconsistent dominant influencing factors and optimal extraction parameters during ultrasonic-assisted extraction.

8.

Stress curves of Jinda117 and Jinda114 under different placement conditions.

7.

SEM images of soybean meal (Jinda117 (panels (a)–(e)); Jinda114: (panels (f)–(j)) after extraction by different methods. Note: Panels (a)–(e) and (f)–(j) represent soybean meal treated under the following conditions: raw bean powder, petroleum ether maceration only, Soxhlet extraction, response surface methodology, and artificial neural network model optimization, respectively.

The adjusted R ² values were highly consistent with the corresponding R ² values for the oil yield regression models of Jinda117 and Jinda114, indicating a good model reliability. Furthermore, the nonsignificant lack of fit verified that the regression models were adequate for predicting oil yield under various combinations of extraction conditions.

Under the optimized conditions for Jinda117 (method A), which were 30 min, 40 °C, 180 W, and a solid-to-liquid ratio of 1:50 g/mL, the highest oil yield was achieved. The theoretical oil content under these conditions was 21.97%, while the experimentally measured value was 21.92% ± 0.89%. For Jinda114, the maximum oil yield was obtained under ultrasonic conditions (method B) of 50 min, 60 °C, 210 W, and a solid-to-liquid ratio of 1:30 g/mL. The theoretical oil content for method B was 20.71%, with an actual measured oil content of 20.73% ± 0.54%.

3.4. ANN-GA Modeling

The relationship between four input variables and the oil yield was modeled using an ANN. The model performance during training, validation, and testing is presented in Figure a and f. After iterative optimization, the optimal validation performance was achieved at epoch 29 for Jinda117 (MSE = 1.77 × 10^–4) and epoch 9 for Jinda114 (MSE = 5.53 × 10^–5). Higher MSE values appeared in the initial training stage (epochs <3), while the training MSE decreased steadily with iteration. Model training was suspended when the validation error increased for ten consecutive iterations, and the optimal model was selected at the epoch corresponding to the minimum validation error. The error distributions for all datasets were centered near zero and remained within acceptable limits (Figure b and g). The training parameters including gradient, mu, and validation failure, further confirmed effective model convergence. The gradient values reached 3.53 × 10^–8 at epoch 30 for Jinda117 and 2.13 × 10^–6 at epoch 15 for Jinda114 (Figure c and h). Post-training regression analysis showed high R ² values between predicted and experimental oil yields for both cultivars. The R ² values of Jinda117 and Jinda114 were 0.9463 and 0.9709 (training), 0.9274 and 0.9653 (validation), 0.9499 and 0.9117 (testing), and 0.9407 and 0.9608 (overall dataset), respectively (Figure d and i). The ANN-GA optimization process continued iteratively until the stopping criteria were met, with convergence achieved at 30 epochs for both models (Figure e and j). Table was used to calculate the ANN parameters.

6.

Results obtained using the Levenberg–Marquardt algorithm, including best validation performance (a, f), error histogram analysis (b, g), training state evolution (c, h), regression plots (d, (i), and GA optimization process (e, j). Panels (a)–(e) correspond to Jinda117, and panels (f)–(j) correspond to Jinda114.

4. Weights and Biases of the Optimal Models for Jinda117 and Jinda114.

	Jinda117					Jinda114
	1.7732	2.2769	2.4463	0.2367		-0.8420	–0.6080	–0.9026	-1.2187
	–0.4379	–2.2177	–2.9640	–3.3986		4.5627	–3.0459	6.4193	1.3149
w 1	0.1007	2.1054	–2.3615	–0.2511		–2.3676	–4.3082	2.2757	0.3359
	0.1523	–1.2833	0.0925	–1.3370		4.4697	1.7375	1.0974	–0.5962
	–1.2567	–0.9531	–0.9985	–0.8618		–0.6407	2.4847	–0.0227	1.0601
b 1	2.0938	–1.3960	2.5607	1.2787	–2.1404	–3.8863	–3.9445	1.2780	2.6594	3.3510
w 2	0.5016	–0.7304	0.7002	0.7415	1.1321	0.7951	–0.4321	0.4302	0.7503	1.2189
b 2	–0.7759					–1.1393

The optimized conditions for Jinda117 (method C) predicted by the ANN-GA model were as follows: ultrasound power, 200 W; extraction temperature, 44 °C; solid-to-liquid ratio, 1:44 g/mL; and extraction time, 20 min, corresponding to a maximum predicted oil yield of 23.02%. Considering practical operational constraints, these parameters were adjusted as follows: ultrasonic power, 210 W; extraction temperature, 44 °C; solid-to-liquid ratio, 1:44 g/mL; and extraction time, 20 min. Under these modified conditions, the experimentally measured oil yield was 22.17% ± 0.23%. Similarly, the ANN-GA model predicted the optimal extraction conditions for Jinda114 (method D) to be as follows: ultrasonic power, 234 W; extraction temperature, 63 °C; solid-to-liquid ratio, 1:20 g/mL; and extraction time, 46 min, yielding a maximum predicted oil content of 21.35%. For practical implementation, the ultrasound power was adjusted to 240 W, while other parameters were kept unchanged. Under these conditions, the experimentally obtained oil yield was 20.91% ± 0.70%.

To assess model reliability, three independent replicate experiments were carried out with extraction parameters optimized by the ANN-GA and RSM. As summarized in Table S2, both models achieved acceptable predictive performance, with all R ² values greater than 0.85 within the present dataset. On the basis of the statistical metrics including R ², AAD, and SD obtained from this experimental dataset, the ANN-GA model displayed preferable predictive behavior for the tested soybean varieties: it yielded higher R ² values (0.903 and 0.952) and lower AAD values (0.85% and 0.82%) relative to the RSM model. Meanwhile, the SDs of oil yield for Jinda117 and Jinda114 under ANN-GA (0.449 and 0.391) were smaller than those calculated from RSM (0.867 and 0.853), reflecting better prediction stability of ANN-GA for this set of data. It should be noted that such a discrepancy in model performance is only valid for the current experimental system, and limited external validation restricts further generalization to other materials. The trend observed in our dataset is consistent with the work of Bu et al. They also recorded comparatively better prediction metrics for ANN-GA versus RSM in their respective research systems.

3.5. Comparative Experiment

Table compares the oil yields of Jinda117 and Jinda114 obtained using different extraction methods, and the extraction performance of the ANN-GA optimized process outperforms that of the RSM optimized process. This optimized process increased oil yields by 4% for Jinda117 and 3% for Jinda114, compared to Soxhlet extraction and maceration, respectively. The results reveal that ANN-GA-optimized conditions produce notably higher oil yields than traditional optimization methods. ANN accurately fits complex nonlinear variable interactions, and the genetic algorithm enables global search to avoid local optima of conventional models. Although ANOVA shows no significant differences in equilibrium oil yields among ANN-GA-UAE, RSM-UAE, Soxhlet extraction, and maceration (p ₁₁₇ = 0.186, p ₁₁₄ = 0.687), confirming that ANN-GA optimized UAE achieves extraction completeness equivalent to standard Soxhlet extraction and meets the requirement for accurate soybean oil quantification, UAE possesses prominent merits in shorter pretreatment time and lower organic solvent consumption via ultrasonic cavitation-enhanced mass transfer.

5. Effect of Different Extraction Methods on the Oil Yield of Jinda117 and Jinda114.

		Jinda117			Jinda114
Method	Extraction conditions	Oil yield (%)	Institutional measured value (%)	Extraction conditions	Oil yield (%)	Institutional measured value (%)
ANN-GA	200 W, 44 °C, 1:44 g/mL, 20 min	22.17 ± 0.23		234 W, 63 °C, 1:20 g/mL, 46 min	20.91 ± 0.70
RSM	180 W, 40 °C, 1:50 g/mL, 30 min	21.85 ± 0.56	21.69	210 W, 60 °C, 1:30 g/mL, 50 min	20.59 ± 0.60	20.34
Soxhlet		21.48 ± 0.41			20.41 ± 0.65
Maceration		21.41 ± 0.44			20.22 ± 0.87

^aInstitutional measured value refers to oil content detection results tested by qualified professional testing institutions.

ANN-GA and RSM optimized solvent-based ultrasound-assisted extraction (UAE) conditions for Jinda114 and Jinda117 oilseeds both produced higher oil yields than the national standard extraction method. Camelina seeds with a raw oil content of 38.5% were subjected to optimized ultrasound pretreatment to boost oil extraction efficiency under 40 kHz frequency, 150 W power, 20 °C and 21.02 min sonication, and this ultrasonic treatment raised the camelina seed oil yield by 20.76%. These results verify that ultrasonic treatment facilitates oil extraction efficiency.

To investigate the performance of the optimal extraction process derived from ANN-GA across different soybean germplasms, the optimized scheme established based on the two varieties above was applied to another 19 soybean varieties, and the oil extraction results of all tested varieties are summarized in Table S3. Comprehensive analysis of all test samples reveals that the ANN-GA optimized process generally achieves higher oil yields than other extraction methods, among which Process D (ultrasonic power = 234 W, temperature = 63 °C, solid-to-liquid ratio = 1:20 g/mL, and time = 46 min) exhibits the maximum extraction efficiency. Further evaluation based on model evaluation metrics including the coefficient of determination (R ²), average absolute deviation (AAD), root mean square error (RMSE) and standard deviation (SD) corresponding to the dataset in Table S3 demonstrates that the ANN-GA model presents superior predictive accuracy within the dataset of this experiment.

3.6. Characterization

3.6.1. SEM

SEM images of soybean powder and meal obtained under different extraction methods are shown in Figure a–e for Jinda117 and Figure f–j for Jinda114, at a magnification of 20 000×. SEM images (see Figure ) show distinct cell wall damage across extraction methods (Figure a–e: Jinda117; Figure f–j: Jinda114). Raw powder (a, f) features intact compact honeycomb cell structures. Static petroleum ether maceration (b, g) only slightly erodes cell walls, and Soxhlet extraction (c, h) induces uneven, incomplete cell damage via prolonged thermal permeation. RSM-optimized UAE (d, (i) produces moderate cell breakage with many intact cavities remaining. Since Jinda114 has denser, harder tissue (see Figure ), its cells resist ultrasonic disruption more strongly, leaving more complete cellular frameworks in Figure i. ANN-GA optimized UAE (Figure e, j) coordinates the proper ultrasonic temperature, power, and time to produce powerful cavitation, severely collapsing and perforating cell walls to form abundant lipid diffusion channels. Overall, the ANN-GA scheme achieves maximum cell destruction, lowers mass transfer resistance and promotes lipid outflow, matching its superior oil yield.

3.6.2. Stress Test

Ultrasonic power, temperature, and other extraction parameters influence the mechanical properties of soybean cell walls, which, in turn, directly affect the ease of cell disruption and oil release. To assess the applicability of the optimized oil extraction conditions for Jinda114 to other soybean varieties, the stress curves of Jinda117 and Jinda114 were measured (Figure ). As shown, Jinda114 exhibited a higher peak stress across all orientations, indicating greater cell wall strength and structural integrity. Notably, the analysis of mechanical properties (see Figure S1, Table S4) and oil extraction yield (see Table ) for the 19 soybean varieties revealed no significant intervarietal differences (p>0.05) in maximum stress, maximum strain, elastic modulus, or ductility. This lack of significant variation in cell wall mechanical properties across varieties indicates that the optimized extraction conditions (developed on the basis of Jinda114) are robust to varietal differences and thus generally applicable for efficient oil extraction from different soybean cultivars. These results showed that method D is generally applicable to the efficient extraction of soybean oil.

To verify the relationship between the oil extraction rate and mechanical properties, a significance analysis was performed (Table S3). The data showed a positive correlation between the cell wall integrity and stress. Under 240 W of ultrasonic treatment, tougher and more intact cell walls impede lipid release, leading to a relatively lower oil-extraction yield.

3.7. Extraction Kinetics and Thermodynamics

The UAE oil extraction from Jinda117 and Jinda114 was investigated through thermodynamic and kinetic analyses. From a thermodynamic perspective, oil extraction is fundamentally a mass transfer process, in which oil diffuses from ruptured cells into the solvent and subsequently into the bulk external phase. Examining the effect of temperature on the equilibrium constant (K) provides insights into the driving force of this transfer. Therefore, the oil extraction process was studied over the temperature range of 298–318 K. Using Van’t Hoff’s method, the linear relationships obtained (Figure a and b) were lnK _Jinda117 = −9428.67 × 1/T + 32.64 (R ² = 0.9972) and lnK _Jinda114 = −4311.10 × 1/T + 15.87 (R ² = 0.9912). The thermodynamic parameters (ΔG, ΔH, and ΔS) calculated from linear least-square analysis (Figure a and d) are presented in Table . Negative ΔG values across all temperatures (298–318 K) indicated that the extraction process was thermodynamically spontaneous, while positive ΔH and TΔS values confirmed its endothermic nature.

9.

Plot of ln K versus 1/T during the oil extraction process for (a) Jinda117 and (d) Jinda114, and first-order kinetics ((b) Jinda117 and (e) Jinda114) and second-order kinetics ((c) Jinda117 and (f) Jinda114) of oil extraction.

6. Thermodynamic and Kinetic Parameters of Oil Extraction from Jinda117 and Jinda114.

Thermodynamic parameters
T (K)	lnK	ΔG (kJ/mol)	TΔS (kJ/mol)	ΔH (kJ/mol)
Jinda117
298	1.01 ± 0.11	–2.5111	77.4204
303	1.49 ± 0.21	–3.7491	78.7194
308	1.97 ± 0.22	–5.0213	80.0184	74.9034
313	2.56 ± 0.54	–6.4360	81.3174
318	3.07 ± 0.48	–7.9360	82.6164

Jinda114
298	1.35 ± 0.18	–3.3343	38.0844
303	1.66 ± 0.28	–4.1650	38.7234
308	1.90 ± 0.35	–4.8001	39.3624	34.6291
313	2.11 ± 0.15	–5.4764	40.0014
318	2.31 ± 0.41	–5.9858	40.6404

Kinetic parameters
The first-order kinetic model
	Power (W)	Oil yield (%)	k 1 (min–1)	R 2
	0	17.71 ± 0.93	0.371 ± 0.041	0.986
Jinda117	180	18.71 ± 1.02	0.349 ± 0.032	0.989
	210	21.63 ± 0.47	0.586 ± 0.063	0.996
	240	20.65 ± 0.69	0.486 ± 0.056	0.992
	0	17.00 ± 0.92	0.405 ± 0.064	0.978
Jinda114	210	17.64 ± 0.76	0.455 ± 0.067	0.985
	240	19.78 ± 0.78	0.545 ± 0.111	0.984
	270	18.81 ± 0.79	0.441 ± 0.056	0.988

The second-order kinetic model
	0	18.57 ± 0.22	0.055 ± 0.012	0.991
Jinda117	180	19.66 ± 0.24	0.049 ± 0.011	0.991
	210	21.99 ± 0.13	0.139 ± 0.040	0.997
	240	21.21 ± 0.19	0.087 ± 0.024	0.994
	0	17.77 ± 0.26	0.061 ± 0.018	0.985
Jinda114	210	18.31 ± 0.19	0.072 ± 0.017	0.992
	240	20.37 ± 0.23	0.081 ± 0.026	0.991
	270	19.56 ± 0.16	0.063 ± 0.011	0.995

Thermodynamic parameters intuitively characterize the practical feasibility and operational characteristics of ultrasonic-assisted soybean oil extraction. Thermodynamic analysis indicates that the enthalpy change (ΔH) of the soybean oil extraction system is 35–74 kJ/mol, and the process is dominated by endothermic mass transfer. Combined with thermogravimetric analysis, raw material pretreatment at 95 °C and constant-temperature extraction at 63 °C can fully provide the heat required for extraction, while preventing the oxidative deterioration of heat-sensitive substances under high temperature and maintaining product quality. The positive entropy change (ΔS, 130–250 J/(mol K)) reflects increased system disorder, verifying that ultrasonic treatment disrupts soybean cell structures and enhances mass transfer via the optimization of ultrasonic power, solid-to-liquid ratio, and extraction time. The negative Gibbs free-energy change (ΔG, −2 to −8 kJ/mol) demonstrates the spontaneous nature of extraction with a weak driving force. Overall, synergistic regulation of thermodynamic conditions and process parameters enables efficient and high-quality soybean oil extraction, validating the reliable industrial applicability of the developed UAE method.

The kinetics of oil extraction from Jinda117 and Jinda114 under different ultrasonic powers are presented in Figure b, c, e, and f. The results indicated that the equilibrium oil yield reached its maximum at 210 W for Jinda117 and 240 W for Jinda114 (Table ). Within the present experimental dataset, the two soybean cultivars exhibited distinct kinetic extraction characteristics: For Jinda117 and Jinda114 soybeans, the second-order kinetic model can better fit their extraction processes with a higher coefficient of determination (R ²). Their maximum equilibrium oil yields are 21.63% ± 0.47% and 20.37% ± 0.23%, respectively. The extraction processes are synergistically governed by concentration-gradient-driven mass transfer and cell structure disruption. Compared with traditional extraction methods, ultrasonication significantly shortened the time required to reach equilibrium. Benefiting from the cavitation and mechanical vibration effects that disrupt soybean cell walls and accelerate lipid release, ultrasonic-assisted extraction significantly shortened the equilibrium time, compared with traditional extraction methods. These identified varietal differences in extraction mechanisms provide theoretical support for the differentiated optimization of ultrasonic parameters and extraction duration, which is conducive to standardizing industrial operation and improving the stability and efficiency of large-scale soybean oil extraction.

The activation energies for oil extraction from Jinda117 and Jinda114 were obtained as 104.78 and 90.86 kJ/mol at ultrasonic powers of 210 and 240 W, respectively. It is evident that the increase in ultrasonic power is more advantageous for lowering the activation energy in the cultivar with higher hardness, which, in turn, facilitates the enhancement of oil extraction efficiency.

3.8. Limitations and Future Work

This study still has several limitations. First, the ANN-GA training dataset included only two soybean cultivars (Jinda117 and Jinda114); the other 19 cultivars were used solely for external validation of the optimized process and were not involved in model training, which constrains model generalizability. More diverse genotypes will be incorporated into the training set in future work to enhance the predictive performance. Second, soybean oil content was calculated from the extraction yield in this work, which introduces minor measurement uncertainty compared with direct standard detection. Third, future work will focus on expanding experimental datasets, conducting systematic method calibration, and performing further validation under practical industrial conditions to promote the industrial application of the ANN-GA optimized soybean oil ultrasonic extraction technology.

4. Conclusions

This study developed and optimized an ultrasonic-assisted extraction (UAE) method for the determination of the soybean oil content using two representative soybean varieties, Jinda117 and Jinda114. An artificial neural network–genetic algorithm (ANN-GA) was employed to determine the optimal UAE conditions, which exhibited superior extraction efficiency and higher oil yield, while significantly reducing the pretreatment time and solvent consumption compared with conventional maceration and standard Soxhlet extraction. Scanning electron microscopy characterization further verified that ultrasonic cavitation effectively disrupted the soybean cell wall structure, reduced oil mass transfer resistance, and significantly enhanced the extraction rate of the intracellular lipids. Thermodynamic and kinetic analyses demonstrated that the oil extraction processes for both varieties were endothermic and spontaneous and conformed to the second-order kinetic model. Furthermore, validation across multiple varieties confirmed that the optimized UAE protocol possesses an excellent stability and universality. The differing extraction activation energies further clarified the distinct extraction mechanisms between the two soybean cultivars. In conclusion, this established UAE method enables rapid, stable, and accurate determination of soybean oil content, providing an efficient and reliable alternative to traditional detection techniques.

Based on these findings, a reliable procedure for determining soybean oil content was established. The method involves crushing dried beans, passing them through a 0.100–0.125 mm sieve, and drying at 95 °C for 1 h. For extraction, dried soybean flour (0.50 g) is treated at 240 W and 63 °C, maintaining a solid-to-liquid ratio of 1:20 g/mL for 46 min. Limited external validation and the quantitative uncertainty of yield-based oil content determination are the main limitations. Further verification and method calibration will be conducted in future work.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.6c04123.

• Additional experimental procedures, Statistical indexes of RSM/ANN prediction performance (Table S1), extraction condition parameters (Table S2), soybean oil yield data of 19 cultivars (Table S3), mechanical stress analysis results­(Table S4), stress curves of bean samples (Figure S1) (PDF)

Xiao-Fang Lu: Writingreview & editing, Writingoriginal draft. Ao-Jie Lu: Visualization, Validation, Methodology, Data curation, Conceptualization. Yi-Jie Zhang: Data curation. Min Sun: Writingreview & editing. Kai-Lin Zhang: Software, Methodology, Investigation. Yan-Ran Qiao: Data curation. Zhao-Yang Song: Data curation. Fu-Gui Wang: Data curation, Algorithm optimization. Ai-Qin Yue: Visualization. Jin-Zhong Zhao: Supervision. Wei-Jun Du: Validation, Resources, Project administration, Funding acquisition, Formal analysis, Conceptualization. All the authors read and approved the final manuscript.

The authors declare no competing financial interest.