Calibration and experimental verification of discrete element simulation parameters of seedling pot based on NSGA-II algorithm
Figures
Abstract
To improve the accuracy of the simulation parameters used in discrete element simulations of transplanting operations and to facilitate further optimization of key transplanter components, this study calibrated and optimized the discrete element model of the seedling pot (16 × 8 holes) based on puncture and shear tests. The contact parameters of the seedling pot were determined through basic tests. The Plackett-Burman test was used to select the initial parameters and determine their value ranges. The shear strength (X2), the static friction coefficient (X4), and the ultimate stress (X10) significantly affected the puncture and shear forces. Based on the central composite design (CCD), a two-objective (puncture force and shear force) regression model was established, and the optimal values were obtained with Design-Expert software. When X2, X4, and X10 were 1.4, 0.885, and 0.115, respectively, the predicted puncture and shear forces were 6.4 N and 11.62 N. Compared with the measured values (6.4 N and 11.3 N), the software solution emphasized the puncture force. The parameters were then substituted into EDEM (2022, Altair Engineering Inc., Troy, MI, USA) for simulation, giving simulation values of 7.2 N and 10.1 N. The relative errors with respect to the measured values were 12.5% and 10.6%, respectively. The calculation has defects. In this study, the multi-objective optimization based on NSGA-II algorithm was applied to the EDEM parameter model. A set of optimized Pareto optimal solutions were selected and simulated. The relative errors of the resulting model were small, namely 6.25% and 7.08%. The results indicate the potential feasibility of the NSGA-II calibration method under the investigated conditions.
Citation: Ji D, Shi J, Zhao Z, Liu F, Guo C, Wang Y, et al. (2026) Calibration and experimental verification of discrete element simulation parameters of seedling pot based on NSGA-II algorithm. PLoS One 21(11): e0361917. https://doi.org/10.1371/journal.pone.0361917
Editor: Rachid Bouharroud, Institut National de la Recherche Agronomique, MOROCCO
Received: April 26, 2026; Accepted: October 16, 2026; Published: November 20, 2026
Copyright: © 2026 Ji et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All data are presented in the manuscript.
Funding: This work was supported by the Liaoning Provincial Department of Education Basic Research Project (grant No. LJ222610157002) in China, awarded to SBT.
Competing interests: The authors have declared that no competing interests exist.
1. Introduction
An agricultural task of transplanting seedlings into the field is called transplanting technology. Using transplanting technology for crop production can increase the land multiple-cropping index and improve crop yield and quality. Mechanized transplanting can further improve the planting quality and reduce the production cost [1]. Seedlings grow on the substrate and absorb nutrients from it. However, during mechanized transplanting, seedlings must pass through feeding, conveying, and planting stages [2–4]. This causes collisions, impacts, and extrusion between seedlings and key machine components. Mechanical damage can seriously affect the survival and quality of seedlings, resulting in economic losses. Therefore, studying the mechanical properties of the seedling pot is of great importance.
With the development of computer numerical simulation technology, discrete element method (DEM) has become an important method of agricultural research. Through DEM, the force and movement of seedlings can be tracked, and the interaction relationship between seedlings and machines can be analyzed. It can be targeted to improve the machine, and reduce the test cycle and cost [5–8]. Due to the error of the simulation model and the complicated motion law between particles, the accuracy of DEM has some problems, so it is necessary to calibrate the parameters of the simulation model. Discrete element simulation test requires many simulation parameters, including material physical parameters and contact parameters. The material physical parameters include shape size, density, Poisson’s ratio, elastic modulus and so on. The contact parameters include collision recovery coefficient, static friction coefficient, rolling friction coefficient and so on. At present, many scholars have conducted in-depth research on DEM in agricultural materials, calibrated the simulation parameters of discrete element models such as crop seeds, soil, powder, straw, potato etc. [9]. Wang et al. established sunflower seed models with different precisions by using multi-sphere and polyhedron filling methods, and verified the accuracy of the model through experiments [10]. Bembenek et al. established a discrete element model for material compaction of roller press, which proves that the discrete element method is more applicable than the finite element method [11]. Wang et al. combined the simulation of corn seed stacking angle with the actual test, and calibrated the contact parameters between corn seed particle models, which provided a theoretical basis for the structural design and parameter optimization of the metering device [12]. Zhang et al. used the coefficient of variation of resting Angle as an index to calibrate rice seeds through cylinder lifting and sliding packing tests and the effect of different radius on the dynamic response of rice seed model was studied [13]. Shi et al. established a flexible model of flax stems through shear test and stacking angle test [14].
The multi-objective optimization model seeks a set of solutions for problems with multiple conflicting objectives, achieving a balance among them. Unlike a traditional single-objective optimization problem, a multi-objective optimization problem involves multiple conflicting objective functions and cannot be simply transformed into a single objective function. As a multi-objective optimization algorithm, NSGA-II can directly solve such problems with good optimization performance and fast convergence [15,16]. NSGA-II has been widely used in reservoir operation scheduling, power systems, aerospace, mechanical design, and resource scheduling [17–21], among other fields. Chang et al. [22] applied the NSGA-II algorithm to the optimal operation of a reservoir group and verified its feasibility and effectiveness for multi-objective reservoir operation. Gong et al. [23] used an NSGA-II algorithm based on the weighted crowding degree to concentrate the search on non-dominated regions and improve optimization efficiency. Zhang et al. [24] applied an improved NSGA-II algorithm based on the weighted crowding degree to the joint optimal scheduling of the Xinjiang reservoir group and obtained a multi-objective optimal scheduling scheme with satisfactory annual power generation and ecological benefits.
In this paper, the tomato seedling pot was taken as the research object, and the basic parameters of the tomato seedling pot were obtained through physical experiments. The discrete element model of the tomato seedling pot was established in EDEM software, and the parameter calibration test of the tomato seedling pot was carried out using the shear force and puncture force as the experimental indices. Finally, the CCD was performed using Design-Expert, and multi-objective optimization based on the NSGA-II algorithm was carried out to obtain the optimal simulation parameter combination for the tomato seedling pot.
2. Materials and methods
2.1 Determination of intrinsic parameters of pot and steel
- (1) Seedling pot parameters
Taking the 16 × 8 plug seedling tray as the research object, and the upper surface size, lower surface size and height of a single seedling pot were 30 × 10 × 36 mm. The volume of a single seedling pot was 20 ml, and the seedling substrate was mixed with peat, vermiculite and perlite at a ratio of 3:1:1. The seedling age was 30 days, and the water content was between 30% and 40%. The parameters of tomato seedling pot include mass, density, puncture force, Poisson’s ratio and elastic modulus. The density () was calculated as follow:
(1)where m is the mass of the seedling pot and v is the volume of the pot (20 mL). Twenty seedling pots were randomly selected, and the density of the pot was 321 kg/m3, measured with a high-precision electronic balance (LP-C3003; Liping Metering Equipment Co., Ltd, Suzhou, China).
The compression test of seedling pot was carried out by TMS-Pro texture analyzer, and the intrinsic reference values reflecting the bonding parameters were obtained. The matrix and the root system of the seedlings are intertwined, so the seedling pots are regarded as a whole, which is convenient for research. Due to the irregular shape of the seedling pots, the seedling pots were pruned to make a cube of 10 × 10 × 10 mm in length, width and height. Poisson ‘s ratio () refers to the ratio of transverse deformation to axial deformation under uniaxial tension or compression. The elastic modulus () of pot seedlings reflects the ability to resist elastic deformation. The Poisson’s ratio () and elastic modulus () can be obtained by compression test, which is tested by the flat indenter of the texture analyzer. The experimental picture is shown in Fig 1. The parameters are calculated as follows:
where, is the Poisson ‘s ratio; is the transverse deformation of the pot (mm); is the axial deformation of the pot (mm); is the width of the pot before compression (mm), is the width of the bowl after compression (mm); is the height before compression (mm), is the height after compression (mm). is the elastic modulus, is the maximum bearing force in the elastic deformation stage of the pot (N); is the initial length of the sample (mm); is the cross-sectional area of the sample (mm2); is the length difference of the sample after compression (mm).
2.2 Determination of contact parameters
- (1) Determination of collision recovery coefficient
The restitution coefficient of collision can measure the ability of the seedling pot to recover its original shape after collision. In this paper, the restitution coefficient of the pot and the steel is measured by the free fall method. When the pot falls onto the steel, the coefficient of restitution () is calculated as follows:
(4)where, is the velocity of the pot after the collision impact (m/s); is the velocity of the pot before the collision impact (m/s).
The pot is affected by gravity during the falling process. The normal velocity of the pot before the collision () and the normal velocity of the pot after the collision () can be obtained by the kinetic energy theorem.
(5)(6)Where, is the falling height before collision (mm); is the maximum height of the bounce after the collision (mm). Combined with the above formulas, the recovery coefficient () can be calculated as follow:
(7)In order to facilitate high-speed camera recording and to ensure the impact and rebound of the pot against the steel, a pot with a well-developed root was selected to determine the collision recovery coefficient. The high-speed camera was placed 1000 mm from the test surface, and the pot was dropped freely onto the steel from a height of 300 mm. After the collision, the pot bounced up, and the whole collision process was recorded by a PLEXLOGGER high-speed camera. The collision displacement curve of the pot and the steel was obtained. The experimental setup is shown in Fig 2. The recovery coefficient between the pot and the steel was calculated to be 0.12–0.18 based on 10 repeated tests.
The collision recovery coefficient between the pots is measured by the suspension method, as shown in Fig 3. The pot (b) is suspended and stationary, and the pot (a) is raised to a fixed height (H1), and then released. By using radial collision, the two pots will continue to swing due to inertia after collision, and the degree of swing reflects the collision recovery characteristics between pots.
H1 is the vertical distance from the highest swing point of pot (a) to the datum line before the collision (mm). H2 is the vertical distance from the highest swing point of pot (a) to the datum line after the collision (mm).
The collision process was captured by the high-speed camera. By using coordinate paper to calculate the distance between the highest point and the datum line after the collision of the pot, the collision recovery coefficient () between the bowls can be calculated. The calculation formula is as follows.
(8)Ten pots were selected for the test, and the collision recovery coefficient () between pots was calculated to be 0.21–0.43.
- (2) Determination of static friction coefficient
The F-D201D inclinometer is placed on a horizontal plane, and crushed pots are put into a cylindrical container. The lower surface of the cylindrical container is bonded to the test plane of the inclinometer. The inclinometer is then rotated until the pot slides on the test plane, as shown in Fig 4. The static friction coefficients between pots and between the pot and the steel plate were calculated as follows. Ten tests were carried out, and the average value was taken as the final value. The static friction coefficient between pots is 0.71–0.93, and that between the pot and the steel is 0.64–0.82.
(a) Static friction coefficient test bench; (b) Pot matrix.
where, is the static friction coefficient; is the indication angle when the pot is about to slide (°).
2.3 Simulation model establishment
- (1) Simulation model selection
The Hertz-Mindlin with JKR model is a contact model for particles with strong inter-particle adhesion, mainly used to simulate wet materials such as crops and soil. The model includes normal and tangential contact forces, and the material model is formed by bonding multiple small particles. Bonding bonds are formed between particles, and the material deforms under load, which leads to fracture of the bonding bonds, as shown in Fig 5.Compared with a standard Hertz-Mindlin model without bonding, which would let particles separate freely and fail to maintain the plug integrity, the bond model directly reproduces the cohesive structure and its fracture under puncture and shear. The JKR model only provides reversible adhesion and cannot represent the rigid bonds and fracture strength of the matrix, while linear spring models are too simplistic. Therefore, the bonded-particle model is the most suitable for this cohesive, fracturable material.
- (2) Establishment of pot simulation model
The pot was inverted and rectangular, the shape is more regular. The pot was established by SolidWorks software, and all parameters were consistent with the physical test. The particle radius was set to 1.5 mm, and the number of pot particles was 1766, and a total of 8338 bonding bonds were generated.
- (3) Puncture and shear experiment of the pot
The stems and leaves of the test seedlings were removed from the top of the pot, and the puncture experiment of the pot was carried out using the conical indenter of the TMS-Pro texture analyzer to obtain the maximum puncture force exerted by the indenter on the pot. The shear experiment of the complete pot was carried out with a shear indenter, and the maximum shear force was obtained. During the test, the puncture forward speed was set to 60 mm/min, the return speed to 45 mm/min, and the insertion depth to 15 mm. The shear experiment used the same parameter settings. The maximum puncture force was 6.4 N and the maximum shear force was 11.3 N. The simulation and physical experiment model diagram is shown in Fig 6. The puncture and shear tests were performed on independently selected seedlings. To reduce systematic error, all samples were randomly coded before testing, and the order of testing was randomized. In this way, the puncture and shear measurements can be treated as independent observations.
2.4 Simulation parameter calibration of the pot
- (1) Plackett-Burman design
The coefficient of restitution between the pots and between the pot and steel was measured by high-speed photography. The static friction coefficients between pots and between the pot and steel were measured with an inclinometer. The Poisson’s ratio and elastic modulus were obtained from the compression test. The simulation test parameters were selected from the above test results, and the remaining parameters are referred to in the references [25–27]. In this study, the maximum value of puncture force (Y1) and shear force (Y2) were selected as the evaluation index. Twelve groups of experiments were carried out, and each group of experiments was repeated three times. The average value was selected as the puncture force of the pot. The Plackett-Burman design scheme is shown in Table 1.
- (2) CCD
Based on the Plackett-Burman test, the three most significant parameters were selected: the shear strength of the pot (X2), the static friction coefficient between the pots (X4) and the ultimate stress (X10). The value ranges are taken as low level (−2) and high level (+2). The relative error between the maximum force of the simulated puncture test and the maximum force of the physical puncture was used as the evaluation index (Y1/%). The relative error of the maximum force of the simulated shear test and the maximum force of the physical puncture was used as the evaluation index (Y2/%). The CCD experiment was designed with these two indexes. In this work, α was set to 2 because the CCD satisfies the orthogonality condition. The levels coding table of the CCD is shown in Table 2.
2.5 Fitting method based on NSGA-II algorithm
The NSGA-II algorithm is mainly used to solve the problem of multiple conflicting objective functions. In multi-objective optimization, there is usually not a single optimal solution, but a series of solutions. These solutions reach a certain equilibrium state between each target, which is the Pareto optimal solution. As an effective multi-objective optimization algorithm, NSGA-II is favored by researchers because of its ability to deal with multi-objective problems efficiently. By combining non-dominated sorting and crowding strategy, NSGA-II can not only find high-quality solutions, but also maintain population diversity, thereby improving the quality of the final solution set.
The NSGA-II process begins by randomly generating an initial population of size N and setting the generation number (Gen) to 1. Selection, crossover, and mutation are then performed, advancing Gen to 2. From the second generation onward, the parent and offspring populations are merged to form a new population. Non-dominated sorting is performed again, and the crowding degree among the non-dominated individuals is calculated. The optimal solutions are then selected based on the crowding degree and the non-domination relationship, forming a new offspring population. Finally, through continuous iteration, the optimal solution is obtained once the termination criterion is met. The process is shown in the Fig 7.
3. Results and discussion
3.1 Simulation test results
- (1) The Plackett-Burman experimental design
The Plackett-Burman experimental design and results are shown in Table 3. Table 4 is the significance analysis of the puncture test, and the order of parameters that most affect the puncture test is X4 > X10 > X2. Table 5 is the significance analysis of the shear test, and the order of parameters that most affect the shear test is X10 > X2 > X4. The CCD was carried out for these three factors.
- (2) CCD
The CCD and its results are shown in Table 6. The P-value analysis results for the puncture and shear tests are shown in Table 7. The regression model for the puncture force had P = 0.0136 (< 0.05), indicating that the relationship between the puncture force and the regression equation was significant. It can be seen from Table 7 that X10 have an extremely significant effect on puncture force. X2, X4 and X2X4 have a significant effect on puncture force, and other factors are not significant.
For the final reduced models, the analysis of variance showed that both models were statistically significant. For the puncture-force model (Y1), the model F value was 7.93 with p = 0.001. For the shear-force model (Y2), the model F value was 38.35 with p < 0.0001. The lack-of-fit was not significant for either model (p = 0.36 for Y1 and p = 0.10 for Y2). The coefficients of determination were R² = 0.74 for Y1 and R² = 0.95 for Y2.
The interaction of each factor is shown in Fig 8. Fig 8a and 8d show the interaction effects of X2X4 on puncture force and shear force respectively when X10 is a fixed value. Fig 8b and 8e show the interaction effects of X2X10 on puncture force and shear force respectively when X4 is a fixed value. Fig 8c and 8f show the interaction effects of X4X10 on puncture force and shear force respectively when X2 is fixed. Although the interaction of factors has a certain rule on the change of puncture force and shear force, it is not a simple ground line relationship that can be described clearly. The regression model established by machine learning is considered for optimal solution.
Figure a, b and c show the influence of interaction factors on the puncture force, and figure d, e and f show the influence of interaction factors on the shear force.
The optimal solution is obtained by inputting the target value (Y1 = 6.4, Y2 = 11.3) in the optimization module of DesignExpert software. When X2, X4, X10 are 1.4, 0.885, 0.115, the predicted puncture force and shear force are 6.40 N and 11.62 N, respectively. The predicted puncture force is equal to its target, whereas the predicted shear force is still 0.32 N (2.8%) above its target of 11.3 N. The returned solution matched the puncture-force target better than the shear-force target. This behaviour comes from the desirability-based numerical optimization: the two responses are combined into a single scalar index, and this index can reach its maximum at a point where one response is fully satisfied while the other is not. When the same combination was verified in EDEM, the puncture force and the shear force were 7.2 N and 10.1 N, giving relative errors of 12.5% and 10.6% with respect to the measured values. Both errors are large, so this combination cannot be used as the final simulation parameters, and multi-objective optimization was therefore performed by the NSGA-II algorithm.
3.2 Multi-objective optimization based on NSGA-II algorithm
A multi-objective optimization model was established in modeFRONTIER software, as shown in Fig 9. In the figure, X2, X4 and X10 are the shear strength of the pot, the static friction coefficient of between the pots and ultimate stress respectively. Y1 and Y2 are the puncture force and shear forces of the tomato seedling pot. The calculation formulas of Y1 and Y2 were obtained from the CCD, as shown in formulas (10) and (11). These regression equations were established by eliminating all non-significant terms. In modeFRONTIER, the NSGA-II algorithm was selected with a mutation probability of 0.1, a crossover probability of 0.9, a population size of 20, and a termination generation of 50.
Fig 10 shows the bubble diagram of the Pareto solution set, representing the distribution of individuals at the 50th generation. The green bubbles represent the Pareto solution set, and the Pareto front is marked with a red contour. As shown in the figure, most designs lie on the Pareto front and represent trade-offs between Y1 and Y2, indicating that the NSGA-II algorithm achieves a good optimization effect.
Fig 11 is the parallel coordinate diagram based on Pareto solution set, and the abscissa represents the optimized parameters (X2, X4 X10) and indexes (Y1 Y2). The red fold lines represent the Pareto front solution set, which can be selected as the optimal solution. Select the optimal design parameters in the Pareto solution set. The predicted parameters and experimental values are shown in the Table 8.
To check whether these settings are sufficient, a convergence analysis and a sensitivity analysis were carried out on the same optimization problem defined by formulas (10) and (11). Two normalized objectives were used, f1 = |Y1 − 6.4|/ 6.4 and f2 = |Y2 − 11.3|/ 11.3, and the combined relative error E = f1 + f2 was used as the decision index. Each setting was repeated 20 times with different random seeds.
In the convergence analysis, the mean best value of E decreased from 2.05% at 10 generations to 0.071% at 50 generations and to 0.009% at 100 generations, and the hypervolume of the Pareto front became stable after about 30 generations (0.05002 at 30 generations and 0.05006 at 50 generations, a change of less than 0.1%). Therefore, 50 generations is located on the converged part of the curve.
In the sensitivity analysis, one setting was changed at a time. When the population size was increased from 10 to 50, the mean best value of E decreased from 0.310% to 0.007%; when the number of generations was increased from 30 to 80, it decreased from 0.278% to 0.015%. The crossover probability (0.7, 0.8, 0.9) and the mutation probability (0.05, 0.1, 0.2) had no clear systematic effect, and the mean best value of E stayed between 0.055% and 0.099% for all of them.The scatter caused by the algorithm settings is therefore much smaller than the error of the calibration itself, which is 6.25% and 7.08% in the verification simulation. This indicates that the settings used in this study give stable and reliable optimization results, as shown in Fig 12.
Fig 12 shows the convergence of the NSGA-II optimization. The hypervolume rose quickly at first and became stable at about 0.050 within 50 generations. The best combined relative error E* dropped from about 8% to nearly 0 in the same range. The results changed little after 50 generations, so 50 generations were used in this study.
Fig 13 shows the sensitivity of E* to the four NSGA-II settings. When the population size increased from 10 to 50, the mean best value of E* dropped from 0.310% to 0.007%, and the error bars became very small at 50. The number of generations showed the same trend: the mean best value of E* fell from 0.278% to 0.015% as the generations increased from 30 to 80. In contrast, the crossover probability (0.7, 0.8, 0.9) and the mutation probability (0.05, 0.1, 0.2) had no clear systematic effect, and the bars were low and almost level, with values staying between 0.055% and 0.099%. The scatter caused by the algorithm settings is therefore much smaller than the error of the calibration itself, which is 6.25% and 7.08% in the verification simulation. This indicates that the settings used in this study give stable and reliable optimization results.
3.3 Optimization results of NSGA-II algorithm
It can be seen from Table 8, when the parameters X2, X4, X10 are 1.476, 0.716 and 0.124 the simulation values (Obj1, Obj2) are respectively 6.53 and 10.66. Taking the parameters into EDEM, the puncture force and shear force are 6.8N and 12.1N, respectively. The relative error is 6.25% and 7.08%. The calculation is shown in formulas 12 and 13. The difference from the actual value is small (Y1 = 6.4, Y2 = 11.3). The results indicate the potential feasibility of the NSGA-II calibration method under the investigated conditions. In this study, the optimal parameters were first obtained from the surrogate models (the reduced quadratic equations) by NSGA-II, and the corresponding force values are reported as surrogate-model predictions in Table 8. To verify these predictions, EDEM simulations were then performed using the optimized parameters, and the results are listed separately as EDEM validation. A comparison shows that the surrogate predictions (6.53N and 10.66N) agree well with the EDEM results (6.8N and 12.1N), with a relative difference of about 6–12%. This discrepancy is mainly attributed to the calibration error of the DEM contact parameters and model extrapolation, and falls within an acceptable range.
(12)(13)4. Discussion
In this study, multi-objective optimization was used to solve the parameter calibration problem of the EDEM model. Taking the puncture force and shear force measured in the physical experiments as the optimization objectives, the factors, the ranges of values and the fitting equations that have a significant influence on these two objectives are selected through a series of experiments (Plackett-Burman design and CCD). Based on the NSGA-II algorithm, the optimal solution is obtained. Compared with the optimization module in Design-Expert, the multi-objective optimization based on NSGA-II can establish a relatively accurate model. Compared with the traditional single-objective verification method (angle of repose), the model with the puncture force and the shear force as the optimization objectives is a new method and can be applied to more complex simulation environments.
In the bonded-particle model, the seedling pot is built from many small particles. These particles are joined by bonds, and the load is carried through these bonds. The ultimate stress (X10) sets the limit at which a bond breaks. When the bonds break, the pot fails. So X10 controls the peak force in both the puncture test and the shear test. X10 therefore represents the bond strength of the matrix. It is not measured directly; instead, it is a DEM bonding parameter whose value is obtained by calibrating the bonded-particle model so that the simulated puncture and shear forces match the experimentally measured values via the central composite design and NSGA-II optimization. The shear strength (X2) controls how hard it is for two bonded particles to slide past each other. In the shear test, the pot fails mainly by sliding along a shear plane. So X2 has a strong effect on the shear force. The static friction coefficient (X4) controls the friction between pots after the bonds have broken. It decides how the broken pieces hold together and how the load moves to other parts. This affects the force after the peak is reached. Other parameters, such as the restitution coefficient, the rolling friction coefficient and Poisson’s ratio, play a smaller role. They mainly affect the elastic stage or the rebound of the particles. They have little effect on the peak force in a slow (quasi-static) test. This is why the Plackett-Burman screening did not find them significant.
In Design-Expert, the two responses are combined into one value by a desirability function. The result then depends on the weight given to each response and on the units of the responses. For this reason, the method often gives more attention to the response that changes faster. This agrees with the result of this study. The Design-Expert solution fitted the puncture force well, but it showed a larger error for the shear force. NSGA-II works in a different way. It does not combine the two responses into one value. It keeps both responses and searches for a set of balanced solutions. The crowding-distance strategy also keeps the solutions well spread, so the search does not stop too early. Because of this, the final solution can be chosen at a point where both responses are close to the measured values.
The proposed framework, screening by the Plackett-Burman design, regression modelling by the central composite design, and multi-objective solving by NSGA-II is not restricted to tomato plug seedlings. It can be transferred to other plug seedlings (e.g., cucumber and melon) and to other bonded-particle DEM models, provided that measurable macroscopic responses are availabl. However, some limitations should be noted. First, the calibrated parameters are optimal only for the responses used in this study. Parameters that do not affect these responses cannot be determined uniquely. This is a well-known equifinality problem in DEM calibration. Second, the result depends on the contact model and the particle size. If the bonding model or the particle size is changed, the parameters must be calibrated again. Third, only two responses were used. For other operations, more indexes may be needed, such as the angle of repose or the compression curve. Finally, the seedling pot is a biological material and shows large differences between individuals. A calibration based on average measured values cannot fully cover this natural variation.
5. Conclusions
- (1) The basic physical parameters of tomato seedling pot were determined by physical experiments. The range of collision recovery coefficient and the range of static friction coefficient between tomato pot seedlings were measured by high-speed camera and slope instrument, which were 0.12–0.18 and 0.71–0.93, respectively. The range of collision recovery coefficient and static friction coefficient between tomato pot seedling and steel plate were 0.12–0.18 and 0.64–0.82, respectively.
- (2) The Plackett-Burman experiment was carried out based on the physical parameters measured by the physical experiment as the basis for the selection of simulation test parameters. Shear strength of the pot, static friction coefficient between the pots and ultimate stress were selected as the parameters that had a significant effect on the puncture force and shear force. Their ranges are 1.04–0.76, 0.71–0.93 and 0.06–0.16, respectively. Using these three parameters to carry out CCD.
- (3) The optimization module of DesignExpert software was used to optimize the parameters, but the software more considered the impact of puncture force on the index, and the shear force was significantly different from the physical experiment. So multi-objective optimization is performed through NSGA-II algorithm to solve the optimal solution.
- (4) The Pareto diagram is obtained by using modeFRONTIER software to construct the optimization model for multi-objective optimization. A set of optimal solutions were selected and brought into EDEM. When the shear strength of the pot (X2), static friction coefficient between the pots (X4) and ultimate stress (X10) are 1.476, 0.716 and 0.124 respectively, the relative error is 6.25% and 7.08%. The difference from the actual value is small. The results indicate the potential feasibility of the NSGA-II calibration method under the investigated conditions.
The Plackett-Burman, CCD, NSGA-II framework established in this study provides a reusable paradigm for the parameter calibration of biological materials, and it can be transferred to other plug seedlings (e.g., cucumber and melon) and to other bonded-particle DEM models. In practice, the calibrated parameters can be directly used in the simulation and optimization of transplanting mechanisms (grasping, ejecting, and dropping seedlings), which helps reduce trial-and-error, lower seedling damage, and support the development of transplanting equipment.
References
- 1. Hwang SJ, Park JH, Lee JY, Shim SB, Nam JS. Optimization of main link lengths of transplanting device of semi-automatic vegetable transplanter. Agronomy. 2020;10:1938.
- 2. Han LH, Kumi F, Mao HP, Hu JP. Design and tests of a multi-pin flexible seedling pick-up gripper for automatic transplanting. Appl Eng Agric. 2019;35:949–57.
- 3. Wang Y, Chen JN, Wu JW, Zhao Y. Mechanics property experiment of broccoli seedling oriented to mechanized planting. Trans Chin Soc Agric Eng. 2014;30:1–10.
- 4. Wen YS, Zhang LA, Huang XM, Yuan T, Zhang JX, Tan YZ, et al. Design of and experiment with seedling selection system for automatic transplanter for vegetable plug seedlings. Agronomy. 2021;11(10):2031.
- 5. Fang HM, Ji CY, Chandio FA, Guo J, Zhang QY, Arslan C. Analysis of soil dynamic behavior during rotary tillage based on distinct element method. Trans Chin Soc Agric Mach. 2016;47:22–8.
- 6. Milkevych V, Munkholm LJ, Chen Y, Nyord T. Modelling approach for soil displacement in tillage using discrete element method. Soil Tillage Res. 2018;183:60–71.
- 7. Murray S, Chen Y. Discrete element modeling of soil displacement resulting from hoe openers. Trans ASABE. 2019;62(2):253–62.
- 8. Zeng ZW, Ma X, Cao XH, Li ZC, Wang XC. Application status and prospect of discrete element method in agricultural engineering research. Trans Chin Soc Agric Mach. 2021;52:1–20.
- 9. Hou ZF, Dai NZ, Chen Z, Qiu Y, Zhang XW. Measurement and calibration of physical property parameters for Agropyron seeds in a discrete element simulation. Trans Chin Soc Agric Eng. 2020;36:46–54.
- 10. Wang S, Yu Z, Zhang W. Study on the modeling method of sunflower seed particles based on the discrete element method. Comput Electron Agric. 2022;198:107012.
- 11. Bembenek M, Buczak M, Baiul K. Modelling of the fine-grained materials briquetting process in a roller press with the discrete element method. Materials (Basel). 2022;15(14):4901. pmid:35888372
- 12. Wang YX, Liang ZJ, Zhang DX, Cui T, Shi S, Li KH, et al. Discrete-based calibration of interspecific contact parameters in maize seed granule models. Trans Chin Soc Agric Eng. 2016;32:36–42.
- 13. Zhang RF, Jiao W, Zhou JL, Qi B, Liu H, Xia QQ. Calibration of discrete element model parameters of rice seeds with different filled particle radii. Trans Chin Soc Agric Mach. 2020;51:227–35.
- 14. Shi RJ, Dai F, Zhao WY, Zhang FW, Shi LR, Guo JH. Establishment of discrete element flexible model of flax stem and experimental verification of contact parameters. Trans Chin Soc Agric Mach. 2022;53:146–55.
- 15.
Deb K, Agrawal S, Pratap A, Meyarivan T. A fast elitist non-dominated sorting genetic algorithm for multi-objective optimization: NSGA-II. Paper presented at the Parallel Problem Solving from Nature PPSN VI: 6th International Conference Paris, France, 2000 Sept 18ق│ô20 Proceedings 6; 2000.
- 16. Deb K, Pratap A, Agarwal S, Meyarivan T. A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE Trans Evol Comput. 2002;6(2):182–97.
- 17. Gulec MO, Ertugrul S. Pareto front generation for integrated drive-train and structural optimisation of a robot manipulator conceptual design via NSGA-II. Adv Mech Eng. 2023;15(3):16878132231163051.
- 18. Qu YY, Song B, Cai SB, Rao PZ, Lin XC. Study on the optimization of Wujiang - water resources by combining the quota method and NSGA-II algorithm. Water. 2024;16:359.
- 19. Sadeghi R, Heidari A, Zahedi F, Khordehbinan MW, Khalilzadeh M. Application of NSGA-II and fuzzy TOPSIS to time–cost–quality trade-off resource leveling for scheduling an agricultural water supply project. Int J Environ Sci Technol. 2023;20(10):10633–60.
- 20. Xu Z, Ning X, Yu Z, Ma Y, Zhao Z, Zhao B. Design optimization of a shell-and-tube heat exchanger with disc-and-doughnut baffles for aero-engine using one hybrid method of NSGA II and MOPSO. Case Stud Therm Eng. 2023;41:102644.
- 21. Zhang ZH, Guo YP, Tao Q. Dynamic multi-objective path-order planning research in nuclear power plant decommissioning based on NSGA-II. Ann Nucl Energy. 2024;199:110369.
- 22.
Chiu CL, John CF. Multi-objective evolutionary algorithm for operating parallel reservoir system. 2009.
- 23. Gong W, Duan QY, Li JD, Wang C, Di ZH, Ye AZ, et al. Multiobjective adaptive surrogate modeling - based optimization for parameter estimation of large, complex geophysical models. Water Resour Res. 2016;52:1984–2008.
- 24. Zhang JW, Wang X, Liu P, Lei XH, Li ZJ, Gong W, et al. Assessing the weighted multi-objective adaptive surrogate model optimization to derive large-scale reservoir operating rules with sensitivity analysis. J Hydrol. 2017;544:613–27.
- 25. Cui YJ, Wei YZ, Ding XT, Cui GP, He Z, Wang MH. Design and experiment of adjustable spacing end-effector based on cylindrical cam. Trans CSAM. 2022;53:104–14.
- 26. Tong JH, Shi HF, Wu CY, Ding YH, Zhao X, Wang RY. Simulation and test of seedling pot grabbing by spade end-effector. Trans CSAM. 2019;50:107–16.
- 27. Gao GH, Xie HF, Wang TB. EDEM simulation and experiment of pullout force on protected vegetable harvester. Trans Chin Soc Agric Eng. 2017;33:24–31.