Forward development and optimization design method for liquid fuel atomization system of engine

Abstract

The invention discloses a forward development and optimization design method for an engine liquid fuel atomization system, and relates to the field of optimization of a spray combustion system.The forward development and optimization design method comprises the steps that a numerical simulation method is adopted to conduct spray simulation, and high-precision transient three-dimensional spray characteristics are calculated; sampling by using a scrambling Holton sampling method, and constructing an initial sample library; constructing an agent model and training to obtain a low-precision agent model; performing multi-objective optimization by using an NSGA-II genetic algorithm based on the model to obtain a Pareto frontier solution; adopting K-means clustering to classify Pareto frontier solutions, select a finite solution set and calculate prediction and simulation errors, if requirements are met, outputting a high-precision proxy model, otherwise, screening potential optimal sample points based on an expected improvement function, and adding the potential optimal sample points into a sample library for retraining; and on the basis of the high-precision agent model, according to the optimization target, structure parameters and injection parameters are reversely deduced. By comprehensively adjusting the structure and working condition parameters, the mixing quality of fuel and air is remarkably improved, the combustion efficiency is improved, and emission is reduced.

Description

Technical Field

[0001] This invention relates to the field of spray combustion system optimization, and more particularly to a forward development and optimization design method for an engine liquid fuel atomization system. Background Technology

[0002] The engine liquid fuel atomization system is a system that disperses liquid fuel into the combustion chamber in the form of fine droplets. It has a significant impact on the engine's combustion efficiency, stability, and emissions. The nozzle, as its core component, directly determines the fuel atomization effect based on its performance.

[0003] The engine's liquid fuel atomization system has a multifaceted impact on emissions through the performance of the nozzles and the atomization effect. Good atomization breaks the fuel into fine droplets, allowing it to mix thoroughly with air, resulting in more complete combustion and improved combustion efficiency. Poor atomization can lead to excessively high or low local fuel concentrations, resulting in incomplete combustion and the generation of pollutants such as carbon monoxide (CO) and hydrocarbons (HC).

[0004] Engines place high demands on nozzle atomization performance. Nozzles with stable atomization reduce fluctuations in the combustion process, resulting in smoother emissions, facilitating overall engine emission control, and lowering peak pollutant emissions. Improper nozzle design can lead to uneven atomization, causing localized combustion abnormalities, and in severe cases, engine malfunction or even stalling. Therefore, nozzle optimization is crucial. Optimized nozzle design breaks down fuel into finer, more uniform droplets, increasing the contact area between fuel and air, promoting rapid mixing, and thus improving combustion efficiency. Increased combustion efficiency also reduces pollutant emissions, helping to minimize engine pollution, meet increasingly stringent environmental regulations and emission standards, and achieve green development.

[0005] While forward development and optimization of nozzles have significant value, they face numerous problems and challenges in practical development. Currently, a comprehensive and accurate theoretical model is lacking for the internal flow and atomization mechanisms of complex engine nozzles. The complex flow phenomena make it difficult for design theories to accurately describe actual conditions, leading to designs based on these theories potentially failing to achieve ideal results. Furthermore, although computer simulation technology provides a powerful tool for nozzle design, the time and economic costs required for optimization remain substantial due to limitations in computing resources. For actual experiments, comprehensive and accurate experimental verification of engine nozzle performance under different operating conditions is challenging due to limitations in experimental conditions and costs. In addition, the large amount of data generated during experiments requires effective analysis and processing to extract valuable information for nozzle design optimization. Moreover, the high cost of building experimental platforms, especially combustion chambers, and the fact that the focus is primarily on changing operating conditions, makes it difficult to optimize the configuration itself.

[0006] Furthermore, nozzle optimization design requires extensive simulation calculations or actual experiments, which is costly and time-consuming. Surrogate models can replace a large number of simulation calculations with a small number of sample points, reducing costs. At the same time, surrogate models can quickly predict nozzle performance under different combinations of structural parameters, enabling the evaluation of a large number of sample designs in a short time during nozzle optimization design, without requiring detailed simulations for each sample. This helps to quickly screen high-potential solutions and refine optimization.

[0007] In summary, optimizing nozzles is of great significance, but traditional optimization methods, including theoretical calculations, simulations, and actual experiments, are insufficient to accurately and efficiently establish an optimization system for nozzles.

[0008] Therefore, those skilled in the art are dedicated to developing a forward development and optimization design method for engine liquid fuel atomization systems. Summary of the Invention

[0009] In view of the above-mentioned deficiencies of the prior art, the technical problem to be solved by the present invention is that there is a lack of a comprehensive and accurate theoretical model for the internal flow and atomization mechanism of complex engine nozzles, and therefore a lack of a simple model that can be directly used to guide the forward development of engine fuel atomization systems.

[0010] To achieve the above objectives, the present invention provides a method for forward development and optimization design of an engine liquid fuel atomization system, the method comprising the following steps:

[0011] S101: Employ numerical simulation methods to simulate engine liquid fuel spray and calculate high-precision transient three-dimensional spray characteristics;

[0012] S103: Use the scrambled Holton sampling method to perform low-correlation sampling on the high-dimensional parameter space to construct an initial sample library;

[0013] S105: Construct a proxy model and train the proxy model using the training set to obtain a low-precision proxy model;

[0014] S107: Based on the trained low-precision surrogate model, the NSGA-II genetic algorithm is used to optimize the optimization objective to obtain the Pareto front solution;

[0015] S109: Use K-means clustering to classify the Pareto front solutions, select a finite set of solutions from the classified Pareto front solutions, carry out high-precision numerical simulation, calculate the error between the prediction results and simulation results of the low-precision surrogate model, and if the error meets the error threshold requirement, output the final high-precision surrogate model and execute step S113; otherwise, execute step S111.

[0016] S111: Filter potential optimal sample points based on the expected improvement function, add the optimal sample points to the sample library, retrain the low-precision surrogate model using the sample library, and execute step S107.

[0017] S113: Based on the finally trained high-precision proxy model, and according to the optimization objective, the structural parameters and injection parameters of the engine liquid fuel atomization system are inversely derived.

[0018] Furthermore, in step S101, when simulating the engine liquid fuel spray, the simulation includes the structural parameters and operating parameters of the engine liquid fuel atomization system. The structural parameters include the swirl chamber tilt angle, nozzle length, swirl channel rise angle, and number of swirl channels. The operating parameters include fuel temperature, injection pressure, and ambient pressure.

[0019] Further, in step S101, computational fluid dynamics is used to perform numerical simulation on the sample, and the simulation data is post-processed to extract atomization characteristic parameters, including oil film thickness, atomization cone angle, and Sottle average diameter.

[0020] Furthermore, in step S105, if the model accuracy of the proxy model does not meet the requirements, Holton sampling expansion is performed, and the proxy model is retrained based on the new sample library to complete the training of the low-precision proxy model.

[0021] The accuracy of the model is calculated using the coefficient of determination R. 2 To evaluate, the coefficient of determination R 2 The following method is used for calculation:

[0022]

[0023] Among them, SS res S is the sum of squared residuals, and SS is the sum of squared differences between observed and predicted values. tot The sum of squares is the total sum of squares, and the sum of squares of the differences between the observed values ​​and the mean of the observed values.

[0024] Furthermore, the Holden sampling expansion involves continuing to use the scrambled Holden sampling method for sampling in the original parameter space, and then recalculating the new sampling points through simulation and adding them to the sample library.

[0025] Further, in step S107, the optimization objective is multi-objective, including oil film thickness, atomization cone angle and Sottle mean diameter. Based on the trained low-precision surrogate model, a large number of samples are taken in the sampling space and input into the low-precision surrogate model to calculate the objective function value. The objective function value is sorted by non-dominated order and crowding is calculated to obtain the elite solution parent. The parent samples are subjected to selection, crossover and mutation operations. After multiple rounds of iteration, the Pareto front solution is finally obtained.

[0026] Furthermore, in step S109, when calculating the prediction result of the low-precision surrogate model, the prediction result is calculated using the following method:

[0027] y0 = f(x; θ)

[0028] Where x is the input vector, y0 is the prediction result, and θ is the model parameter, which makes y0 as close as possible to the true output y.

[0029] Further, in step S111, the desired improvement function is:

[0030]

[0031] Where EI(x) is the expected improvement function, x is the new sample point, σ(x) is the standard deviation of the test point, μ(x) is the predicted mean of the test point, and y best For the optimal observation, Φ(·) is the cumulative distribution function of the standard normal distribution.

[0032] Further, in step S113, the structural parameters and the injection parameters are obtained by inverse derivation based on the Pareto front solution, specifically including the following process:

[0033] S1131: Set different weights for multiple optimization objectives to obtain the weight-objective curve;

[0034] S1132: Solve for the intersection of the weight-target curve and the Pareto front to obtain the corresponding optimal solution;

[0035] S1133: Based on the optimal solution, the structural parameters and the injection parameters are derived in reverse.

[0036] Furthermore, in step S113, the structural parameters and injection parameters obtained by reverse derivation are simulated and calculated to obtain actual calculation results. The actual calculation results are then compared with the expected results to calculate the actual error and complete the closed-loop verification.

[0037] In a preferred embodiment of the present invention, compared with the prior art, the present invention has the following beneficial technical effects:

[0038] 1. This invention targets key structural parameters and operating parameters, using several key indicators affecting spray atomization effect as objective functions to construct a proxy model, thereby achieving a leap from theoretical modeling difficulties to engineering operability for engine fuel atomization systems. This significantly improves the design feasibility and optimization capability of fuel atomization systems, while also possessing data-driven flexibility and scalability.

[0039] 2. This invention is based on Gaussian process regression and uses a few-sample addition technique for dynamic optimization to construct a closed-loop system of "precise sampling-model iteration". While ensuring prediction accuracy, it significantly reduces the sample requirement and can train a high-precision surrogate model with a relatively small sample size. This significantly reduces the sample requirement for training the high-dimensional parameter space surrogate model of the atomization system, while ensuring the accuracy of model prediction and the correctness of the direction of spray feature optimization.

[0040] 3. Based on a trained surrogate model, this invention can quickly reverse-engineer the nozzle structure and operating parameter combination that satisfies the optimal solution of the objective function. The corresponding parameter combination can be searched according to the optimization objective weight, which avoids the waste of experimental or computational costs caused by blindly conducting experiments due to the unknown nature of the spray feature results in the process of optimizing spray feature design using traditional forward optimization design methods.

[0041] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the method steps of a preferred embodiment of the present invention;

[0043] Figure 2 This is an overall flowchart of the optimization method of a preferred embodiment of the present invention;

[0044] Figure 3 This is a schematic diagram of a fuel injector nozzle structure according to a preferred embodiment of the present invention;

[0045] Figure 4 This is a schematic diagram of the nozzle geometry of a preferred embodiment of the present invention;

[0046] Figure 5 This is a schematic diagram of the sampling point distribution results before and after scrambling according to a preferred embodiment of the present invention. The left figure is a schematic diagram before scrambling, and the right figure is a schematic diagram after scrambling.

[0047] Figure 6 This is a schematic diagram of the spatial distribution of Holton sampling points according to a preferred embodiment of the present invention;

[0048] Figure 7 R is a preferred embodiment of the present invention.2 A schematic diagram showing the changes in the number of expansion rounds of Holden;

[0049] Figure 8 This is a schematic diagram of the Pareto front and point selection of a preferred embodiment of the present invention;

[0050] Figure 9 This is a schematic diagram illustrating the variation of the average error with the number of addition rounds in a preferred embodiment of the present invention;

[0051] Figure 10 This is a schematic diagram illustrating how the optimization target improvement amount changes with the number of rounds of point addition in a preferred embodiment of the present invention. Detailed Implementation

[0052] The following description, with reference to the accompanying drawings, illustrates several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0053] In the accompanying drawings, components with the same structure are indicated by the same numerical designation, and components with similar structures or functions are indicated by similar numerical designations. The dimensions and thicknesses of each component shown in the drawings are arbitrary, and the present invention does not limit the dimensions and thicknesses of each component. To make the illustrations clearer, the thickness of some components has been appropriately exaggerated in the drawings.

[0054] To address the difficulty of accurately and efficiently establishing an optimization system for nozzles through existing optimization methods, including theoretical calculations, simulations, and practical experiments, this invention proposes a forward design and optimization method for engine liquid fuel atomization systems that couples a few-shot optimization method with a surrogate model. This method uses multiple performance parameters as optimization objectives, models the spray generation mechanism based on Gaussian process regression, performs multi-objective optimization using the NSGA-II (Non-dominated Sorting Genetic Algorithm II) genetic algorithm, and integrates a few-shot optimization method to train a relatively accurate predictive surrogate model with a very small sample size. Based on the trained surrogate model, the corresponding parameter combinations can be quickly derived in reverse from the optimization results on the Pareto front, achieving a goal-oriented optimization design system.

[0055] like Figure 1 As shown in the figure, an embodiment of the present invention provides a forward development and optimization design method for an engine liquid fuel atomization system, comprising the following steps:

[0056] S101: Numerical simulation is used to simulate engine liquid fuel spray and calculate high-precision transient three-dimensional spray characteristics.

[0057] When simulating engine liquid fuel spray, the simulation parameters include the structural parameters and operating parameters of the engine liquid fuel atomization system. The structural parameters include the swirl chamber tilt angle, nozzle length, swirl channel rise angle, and number of swirl channels. The operating parameters include fuel temperature, injection pressure, and ambient pressure.

[0058] In this embodiment, computational fluid dynamics is used to perform numerical simulation on the sample, and the simulation data is post-processed and calculated to finally extract atomization characteristic parameters, including oil film thickness, atomization cone angle and Sottle average diameter. These characteristic parameters are also the multi-objective optimization parameters of this embodiment.

[0059] S103: Use the scrambled Holton sampling method to perform low-correlation sampling on the high-dimensional parameter space to construct an initial sample library.

[0060] Halton sequences are used for low-correlation sampling in high-dimensional parameter spaces. For fuel atomization systems with complex structures and injection parameters, the correlation between inverse cardinal functions using different cardinalities in different dimensions leads to poor two-dimensional projection distribution. Therefore, scrambled Halton sampling can be used. This method finds the optimal multiplier for the Halton sequence in a linear scrambled space and reorders the breakpoints of the ordinary Halton sequence, avoiding uneven spatial distribution of sample points.

[0061] S105: Construct a proxy model and train the proxy model using the training set to obtain a low-precision proxy model.

[0062] Because a comprehensive and accurate theoretical model for the internal flow and atomization mechanism of complex engine nozzles is still lacking, there is a lack of simple models that can be directly used to guide the forward development of engine fuel atomization systems. In this embodiment, a surrogate model is constructed using several key indicators affecting the spray atomization effect as objective functions, targeting key structural parameters and operating parameters.

[0063] The sample library was divided into training and validation sets. Gaussian process regression (GPR) was used, and kernel functions such as RBF kernel and Mntern kernel were selected to construct a surrogate model.

[0064] Use certain methods to evaluate model accuracy, such as calculating the coefficient of determination (R²). 2 To assess the accuracy of the model, a certain threshold is set as the convergence criterion.

[0065] In this embodiment, model accuracy is calculated using the coefficient of determination R. 2 To evaluate, the coefficient of determination R 2 The following method is used for calculation:

[0066]

[0067] Among them, SS res S is the sum of squared residuals, and SS is the sum of squared differences between observed and predicted values. tot The sum of squares is the total sum of squares, and the sum of squares of the differences between the observed values ​​and the mean of the observed values.

[0068] If the accuracy of the proxy model does not meet the requirements, Holden sampling expansion is performed, and the proxy model is retrained based on the new sample library to complete the training of the low-precision proxy model.

[0069] When expanding the Holden sampling, the scrambled Holden sampling method is used to sample in the original parameter space, and the new sampling points are recalculated through simulation and added to the sample library to continue training the surrogate model.

[0070] S107: Based on the trained low-precision surrogate model, the NSGA-II genetic algorithm is used to optimize the objective and obtain the Pareto front solution.

[0071] In this embodiment, when using the NSGA-II genetic algorithm to optimize the optimization objective, the optimization objective is multi-objective, including oil film thickness, atomization cone angle and Sottle average diameter.

[0072] When using the NSGA-II genetic algorithm for multi-objective optimization, a large number of samples are taken in the sampling space based on a trained low-precision surrogate model and input into the low-precision surrogate model to calculate the objective function value. The objective function value is then sorted by non-dominated order and crowding is calculated to obtain the elite solution parent. The parent samples are then subjected to selection, crossover and mutation operations. After multiple rounds of iteration, the Pareto front solution is finally obtained.

[0073] S109: Use K-means clustering to classify the Pareto front solution, select a finite set of solutions from the classified Pareto front solutions, carry out high-precision numerical simulation, calculate the error between the prediction result and the simulation result of the low-precision surrogate model, if the error meets the error threshold requirement, output the final high-precision surrogate model and execute step S113; otherwise, execute step S111.

[0074] In this embodiment, when the mean error between the prediction result and the actual output of the low-precision proxy model is stable within the threshold, the optimization loop is terminated and the final high-precision proxy model is output; otherwise, the low-precision proxy model needs to be retrained.

[0075] Since the input-output mapping of the surrogate model is essentially based on a data-driven function approximation process, the surrogate model needs to construct an approximate function, and the prediction result is calculated using the approximate function:

[0076] y0 = f(x; θ)

[0077] Where x is the input vector, y0 is the prediction result, and θ is the model parameter, which makes y0 as close as possible to the true output y.

[0078] S111: Filter potential optimal sample points based on the expected improvement function, add the optimal sample points to the sample library, retrain the low-precision surrogate model using the sample library, and execute step S107.

[0079] When the prediction results of the low-precision surrogate model cannot meet the requirements, the sample library needs to be expanded. The expected improvement function is used to screen potential optimal sample points and add the optimal sample points to the sample library. The low-precision surrogate model is retrained using the expanded sample library. Steps S107 to S109 are repeated until the error between the prediction results and the simulation results meets the error threshold, and the final high-precision surrogate model is output.

[0080] In this embodiment, the desired improved function is:

[0081]

[0082] Where EI(x) is the expected improvement function, x is the new sample point, σ(x) is the standard deviation of the test point, μ(x) is the predicted mean of the test point, and y best For the optimal observation, Φ(·) is the cumulative distribution function of the standard normal distribution.

[0083] S113: Based on the final trained high-precision surrogate model, and according to the optimization objective, the structural parameters and injection parameters of the engine liquid fuel atomization system are reverse-engineered.

[0084] The surrogate model is trained using the final sample library. Based on the optimization target results, the structure and injection parameters can be derived in reverse using the surrogate model. Since a well-trained surrogate model establishes a definite mapping relationship between input and output, this mapping is mathematically reversible. Furthermore, the Pareto front solution is the optimal solution set obtained by integrating multiple optimization objectives. Different weights can be set for multiple optimization objectives, and weight-objective curves can be obtained. The intersection of the weight curve and the Pareto front yields the corresponding optimal solution, thus allowing for the inverse derivation of its corresponding parameter combination. Simulation calculations are performed on the inversely derived structure and injection parameters to obtain the actual calculated results of the target parameters under these parameter combinations. These results are compared with the expected results to calculate the actual deviation. The deviation between the atomization performance indicators and the predicted values ​​is confirmed to meet expectations, completing the closed-loop verification of the method.

[0085] Based on the inverse derivation of the Pareto front solution, the structural parameters and injection parameters are obtained, specifically including the following process:

[0086] S1131: Set different weights for multiple optimization objectives to obtain the weight-objective curve;

[0087] S1132: Solve for the intersection of the weight-target curve and the Pareto front to obtain the corresponding optimal solution;

[0088] S1133: Based on the optimal solution, the structural parameters and injection parameters are derived in reverse.

[0089] The structural and injection parameters obtained through reverse derivation are simulated and calculated to obtain the actual calculation results. The actual calculation results are then compared with the expected results to calculate the actual error and complete the closed-loop verification.

[0090] Compared with existing technologies, the forward development and optimization design method for engine liquid fuel atomization system provided in this invention has the following characteristics:

[0091] 1. Addressing the lack of comprehensive and accurate theoretical models for the internal flow and atomization mechanisms of complex engine nozzles in existing technologies, and the absence of simple models directly guiding the forward development of engine fuel atomization systems, this invention constructs a surrogate model targeting key structural and operating parameters, using several key indicators affecting spray atomization performance as objective functions. The surrogate model is trained based on specific data, thus bypassing direct modeling of the complex internal flow and atomization mechanisms of the atomizing nozzle. Through this surrogate model technology, a leap from the difficulty of theoretical modeling to engineering feasibility in engine fuel atomization systems is achieved, significantly improving the design feasibility and optimization capabilities of fuel atomization systems, while also possessing the flexibility and scalability of data-driven approaches.

[0092] 2. Existing technologies using traditional surrogate model training methods for engine fuel atomization systems rely on large-scale training samples to cover the complex nonlinear relationships in the high-dimensional parameter space of the atomization system. This results in high training costs, wasted computational resources, and limited generalization ability. This invention addresses this issue by using Gaussian process regression and a small-sample addition technique for dynamic optimization, constructing a closed-loop system of "precise sampling-model iteration." This significantly reduces the sample requirement while maintaining prediction accuracy, enabling the training of a high-precision surrogate model with a relatively small sample size. This invention proposes an intelligent sample addition strategy based on the Pareto front and Expected Improvement (EI). It calculates the EI value of the Pareto front point and its prediction uncertainty, dynamically selecting the optimal sample points to add to the training library while balancing improvement potential (EI value) and exploration risk (uncertainty). This allows for the precise selection of parameter combinations that significantly contribute to both the surrogate model and the spray feature optimization objective with a very small number of high-quality samples, significantly reducing the sample requirement for training the surrogate model in the high-dimensional parameter space of the atomization system, while ensuring the accuracy of model prediction and the correctness of the spray feature optimization direction.

[0093] 3. Addressing the need for optimized nozzle structure and control parameters in engineering applications based on specific spray characteristics—that is, deriving nozzle structure and operating parameters from the optimal solution of the objective function—this invention utilizes a trained surrogate model to rapidly derive the nozzle structure and operating parameter combination that satisfies the optimal solution of the objective function. The surrogate model establishes a rapid mapping relationship between spray characteristics and design variables. Based on the optimized optimal solution, the parameter combination can be derived inversely, realizing the process of deriving the optimal structure parameters and spray parameter combination from the optimal solution of spray characteristics. The nozzle structure and operating parameter combination is derived inversely from the optimization results, and the corresponding parameter combination can be searched based on the optimization objective weight. This avoids the waste of experimental or computational costs caused by blindly conducting experiments due to the unknown nature of spray characteristic results in the traditional forward optimization design process for spray characteristic design optimization.

[0094] The present invention will now be described in detail with reference to preferred embodiments.

[0095] This invention addresses the high-dimensional optimization problem of engine liquid fuel atomization systems, providing a design and optimization method for engine fuel atomization systems that couples a few-sample optimization approach with a surrogate model. By constructing a dimension-adaptive prime base selection strategy, the correlation of samples in the high-dimensional parameter space is reduced. Combining the uncertainty quantification capability of the Gaussian Process Regression (GPR) surrogate model, a dynamic few-sample addition mechanism is designed to achieve an efficient balance between exploration and utilization in the high-dimensional space. Based on this, the NSGA-II multi-objective optimization algorithm is integrated to establish an intelligent design framework for coupled optimization of fuel atomization parameters. This method not only provides methodological support for the design of high-dimensional, few-sample, and multi-constraint spray combustion system optimization scenarios and for the study and optimization of the influence of injection parameters and structural parameters on the atomization system, but can also be applied to other high-dimensional, few-sample, and multi-objective optimization design needs.

[0096] The method provided in this embodiment of the invention includes the following steps:

[0097] Step 1: Numerical simulation and experimental verification

[0098] Numerical simulation is used to simulate the spray pattern and calculate high-precision transient three-dimensional spray characteristics. When performing numerical simulations on spray parameters, the results should be compared with experimental data to ensure that the error between the simulation and experimental results remains within a certain threshold, thus verifying the reliability of the numerical method.

[0099] In this embodiment, the optimization variables and optimization objectives are set as follows:

[0100] 1) Structural parameters: swirling chamber inclination angle α, nozzle length L0, swirling channel rise angle β, number of swirling channels N;

[0101] 2) Operating parameters: Fuel temperature T f Injection pressure P inj Environmental pressure P env ;

[0102] 3) Optimization objectives: minimize oil film thickness δ, maximize atomization cone angle θ, and minimize Sottle mean diameter SMD.

[0103] This embodiment addresses the optimization of fuel injector atomization characteristics by constructing an initial sample space based on the Latin Hypercube Sampling (LHS) method. The following parameters are selected: swirl chamber inclination angle (α), nozzle length (L0), swirl channel rise angle (β), number of swirl channels (N), and ambient pressure (P). env Injection pressure (P) inj ) and fuel temperature (T) f Seven key design parameters were identified. Computational Fluid Dynamics (CFD) was used to numerically simulate the samples, simulating the dynamic behavior of the fuel jet during atomization. Numerical simulations of fuel spray were conducted using the Fluent simulation software, including mesh generation and boundary condition settings. For example, the Volume of Fluid (VOF) method was used to accurately capture the topological evolution of the phase interface during near-field jet breakup, and the Lagrange Particle Tracking (LPT) method was combined to simulate the dynamic behavior of spray droplets, revealing the distribution characteristics of the far-field spray. The calculated data underwent post-processing to extract atomization characteristic parameters such as oil film thickness (δ), atomization cone angle (θ), and Sottle mean diameter (SMD). The calculation results were compared with experimental or literature data to calculate the error between the actual data and the simulation results, ensuring that the simulation error was within the target range and verifying the reliability of the numerical model.

[0104] Step 2: Initial sampling in high-dimensional space

[0105] The Holden sampling method is used to sample within the parameter space, and a small number of samples are collected for numerical simulation calculations. An initial sample library is then established based on the calculated samples.

[0106] Halton sequences are used for low-correlation sampling in high-dimensional parameter spaces. For the complex structure and injection parameters of fuel atomization systems, the correlation between cardinal inverse functions using different cardinalities in different dimensions leads to poor two-dimensional projection distribution in high-dimensional cases. Therefore, scrambled Halton sampling methods, such as the Mhalthon algorithm, can be used. This sequence is created using a linear digital scrambling method, which can find the optimal multiplier for the Halton sequence in the linear scrambling space and reorder the breakpoints of the ordinary Halton sequence to avoid uneven spatial distribution of sample points. By using high-dimensional Halton sampling methods with the aforementioned independent variable parameters as the sampling space, an initial sample library with a small number of samples is generated.

[0107] Step 3: Training and Validation of the Proxy Model

[0108] Select a portion of the initial sample set as the validation set, and the remainder as the training set. Train the agent model using the training set, and calculate the determination coefficient (R²) of the current agent model. 2 Based on the current coefficient of determination, a decision is made on whether to expand the sample database and what method to use for expansion.

[0109] The sample database was divided into training and validation sets. Gaussian process regression (GPR) was used, and kernel functions such as the RBF kernel and the Mntern kernel were selected to construct a surrogate model. The model accuracy was evaluated using methods such as calculating the coefficient of determination (R²). 2 The accuracy of the model is evaluated, and a certain threshold is set as the convergence criterion. The coefficient of determination is used to assess how well the regression model fits the data. Its value is between 0 and 1 and can be interpreted as the proportion of the variance of the dependent variable explained by the model.

[0110] Coefficient of determination R 2 calculate:

[0111]

[0112] Among them, SS res It is the sum of squared residuals, that is, the sum of the squares of the differences between the observed and predicted values. SS tot This is the total sum of squares, which is the sum of the squares of the differences between the observed values ​​and their mean. Since Holden sampling is a low-correlation sampling method, it is difficult for the trained surrogate model to achieve high accuracy; therefore, this threshold can be set relatively low.

[0113] Step 4: Holden sampling expansion

[0114] If the current surrogate model's determination coefficient is too low, below the preset low-precision surrogate model threshold, it indicates that the basic sample size is too small and needs to be expanded based on the Holden sampling method. Holden sampling is used to continue the previous sampling, and after simulation calculation, the samples are added to the sample library as training set samples. The surrogate model is then trained based on the new sample library, and the surrogate model's determination coefficient is recalculated. This process of expanding the sample size in small increments is repeated until the determination coefficient reaches the expected value.

[0115] If the proxy model R 2 The value did not reach the set threshold. Since Holden sampling is a sequential sampling method, the sampling points in the fixed parameter space are fixed points. Based on this characteristic, Holden sampling can be used to continue sampling in the original parameter space. The sampling points are then recalculated through simulation and added to the sample library. The surrogate model is retrained based on the new sample library, and this process is repeated until the surrogate model R... 2 Once the threshold requirements are met, the low-precision surrogate model training is completed. Based on the surrogate model that has achieved the basic accuracy requirements, the NSGA-II genetic algorithm can be used to iteratively optimize the atomization system and carry out spray feature optimization design. Additionally, a few-shot method can be used for targeted sampling to expand the sample library and further improve the accuracy of the surrogate model.

[0116] Step 5: Multi-objective optimization

[0117] The NSGA-II genetic algorithm was used to optimize the objective and obtain the Pareto front solution.

[0118] Based on the NSGA-II algorithm, multi-objective optimization is performed on key parameters of spray characteristics. The multi-objective optimization process involves using a pre-trained surrogate model to sample extensively in the sampling space, inputting the samples into the surrogate model to calculate the objective function value, sorting these results by non-dominated order, and calculating crowding to obtain elite parent solutions. Parent samples undergo selection, crossover, and mutation operations, iterating through multiple rounds to finally obtain the optimal solution that balances optimization degree and diversity, i.e., the Pareto front solution. The optimization objective can be flexibly set, such as using goal programming to make the optimization objective approach the target value, or setting the optimization objective to approach the maximum value, and obtaining the optimized Pareto front solution.

[0119] Step 6: Add points to a small sample size

[0120] A few-sample addition process is implemented, clustering the Pareto front solutions and selecting the point with the largest EI value for integrated multi-objectives within each cluster as a new sample. This new sample is then used for simulation calculations and added to the sample library, which is used to train the surrogate model. Based on the new surrogate model, the NSGA-II algorithm is used to optimize the objective function.

[0121] In this embodiment, K-means clustering is used to classify the frontier solution, and potential optimal sample points are selected based on the expected improvement function (EI). The expected improvement function is used to balance the exploration and utilization of sample points, that is, based on the current model prediction, to evaluate the degree and possibility of improvement of the surrogate model by newly added sample points, and to quantify the potential benefits of adding sample points to the sample pool. Assume that the current optimal observation value is y. best For a new sample point x, if its predicted value follows a normal distribution f(x)~N(μ(x),σ2(x)), then the formula for calculating the EI value is:

[0122]

[0123] Where σ(x) is the standard deviation of the test points, μ(x) is the predicted mean of the test points, and Φ(·) is the cumulative distribution function of the standard normal distribution.

[0124] The selected samples are simulated and verified to calculate the relative error between the predicted and simulated values. If the error does not meet the target threshold, the sample is added to the training set and the surrogate model is retrained. Based on the new surrogate model, the NSGA-II algorithm is used to optimize the three objective functions of oil film thickness (δ), atomization cone angle (θ), and Sottle mean diameter (SMD). Steps 5 and 6 are repeated until the error remains within the target threshold.

[0125] Step 7: Convergence Determination and Reverse Design

[0126] Repeat the above process until the relative error between the simulation results and the prediction results of the new sample remains within the set threshold, and output a high-precision surrogate model.

[0127] When the mean error stabilizes within the threshold, the optimization loop terminates, and the final high-precision surrogate model is output. This is because the input-output mapping of the surrogate model is essentially based on a data-driven function approximation process.

[0128] Assuming the input is a vector x and the output is y0, the surrogate model needs to construct an approximate function:

[0129] y0 = f(x; θ),

[0130] Here, θ is the model parameter, which makes y0 as close as possible to the true output y.

[0131] Because the model has learned the mapping relationship between input parameters and output results from historical data during the training phase, and has achieved sufficient accuracy through multiple updates, the final surrogate model can guarantee the prediction of key spray characteristics for atomization systems with given structures and spray parameters. Furthermore, the optimized Pareto front solution, through accurate mapping by the high-precision surrogate model, can not only achieve accurate prediction of the optimization objective, but also achieve significant improvements in the optimization objective due to the use of a genetic algorithm for iterative optimization.

[0132] Step 8: Reverse derivation and loop closure verification

[0133] Based on the final trained high-precision surrogate model, the structure and injection parameters are inferred using the surrogate model according to the optimization objective. Simulation calculations are then performed on the inferred structure and injection parameters to obtain the actual calculated results of the target parameters under these parameter combinations. These results are compared with the expected results to calculate the actual error and complete the closed-loop verification.

[0134] The surrogate model is trained using the final sample library. Based on the optimization target results, the structure and injection parameters can be derived in reverse using the surrogate model. Since a well-trained surrogate model establishes a definite mapping relationship between input and output, this mapping is mathematically reversible. Furthermore, the Pareto front solution is the optimal solution set obtained by integrating multiple optimization objectives. Different weights can be set for multiple optimization objectives, and weight-objective curves can be obtained. The intersection of the weight curve and the Pareto front yields the corresponding optimal solution, thus allowing for the inverse derivation of its corresponding parameter combination. Simulation calculations are performed on the inversely derived structure and injection parameters to obtain the actual calculated results of the target parameters under these parameter combinations. These results are compared with the expected results to calculate the actual deviation. The deviation between the atomization performance indicators and the predicted values ​​is confirmed to meet expectations, completing the closed-loop verification of the method.

[0135] Compared with existing technologies, the forward development and optimization design method for engine liquid fuel atomization systems provided in this invention can achieve high-efficiency, low-emission engine performance optimization through comprehensive adjustment of nozzle parameters and injection condition parameters, and has the following characteristics:

[0136] I. Technological Advantages

[0137] 1. High computational efficiency: A dynamic point-addition strategy based on the Pareto front and Expected Improvement (EI) is adopted. This strategy intelligently selects parameter combinations (such as points with high EI values) that significantly contribute to both the accuracy of the surrogate model and the optimization objective, thereby significantly reducing invalid sampling in the high-dimensional parameter space. Experiments show that, compared to traditional uniform sampling methods, this method requires only a significantly smaller sample size to achieve the same accuracy, while reducing computational resource consumption.

[0138] 2. Optimize Goal Orientation: By constructing a direct mapping relationship between spray characteristics and design variables through Gaussian process regression (GPR), the limitations of complex mechanism modeling are avoided. Combined with the NSGA-II multi-objective optimization algorithm, a closed loop is achieved from forward prediction of "parameter-goal" to inverse design of "goal-parameter".

[0139] 3. Layered Optimization: Layered model training avoids the waste of computational resources caused by extensive use of uniform sampling, which results in minimal improvement due to a limited number of key data points. Simultaneously, it avoids the situation where, under conditions of low base accuracy, the low accuracy of the surrogate model negatively impacts the value judgment of sample points by the few-sample method.

[0140] 1) Low-precision model training phase: Low correlation sampling of Holden sequences to build a basic surrogate model.

[0141] 2) High-precision model training phase: Based on the active learning strategy of EI, by quantifying the model uncertainty and expected improvement, the marginal benefit of new samples to accuracy improvement is maximized, so that the model accuracy continues to improve (e.g., Figure 6 As shown), the improvement in the optimization objective is significant (e.g. Figure 8 (As shown).

[0142] II. Optimization Effects and Prospects

[0143] Experiments have verified that, based on the method provided by this invention, a high-precision surrogate model was trained with a very small sample size, and the characteristic parameters of fuel spray were significantly optimized in the target direction. The optimized spray characteristics can significantly improve the mixing quality of fuel and air, thereby increasing combustion efficiency. Spray optimization can also reduce the formation of locally excessively rich or lean mixtures, thus reducing pollutant emissions.

[0144] In the future, this method can be used for overall engine design, combining spray parameters with other engine parameters and operating conditions to achieve higher performance through multi-parameter collaborative optimization.

[0145] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for forward development and optimization design of an engine liquid fuel atomization system, characterized in that, The method includes the following steps: S101: Employ numerical simulation methods to simulate engine liquid fuel spray and calculate high-precision transient three-dimensional spray characteristics; S103: Use the scrambled Holton sampling method to perform low-correlation sampling on the high-dimensional parameter space to construct an initial sample library; S105: Construct a proxy model and train the proxy model using the training set to obtain a low-precision proxy model; S107: Based on the trained low-precision surrogate model, the NSGA-II genetic algorithm is used to optimize the optimization objective to obtain the Pareto front solution; S109: Use K-means clustering to classify the Pareto front solutions, select a finite set of solutions from the classified Pareto front solutions, carry out high-precision numerical simulation, calculate the error between the prediction results and simulation results of the low-precision surrogate model, and if the error meets the error threshold requirement, output the final high-precision surrogate model and execute step S113; otherwise, execute step S111. S111: Filter potential optimal sample points based on the expected improvement function, add the optimal sample points to the sample library, retrain the low-precision surrogate model using the sample library, and execute step S107. S113: Based on the finally trained high-precision proxy model, and according to the optimization objective, the structural parameters and injection parameters of the engine liquid fuel atomization system are inversely derived.

2. The method as described in claim 1, characterized in that, In step S101, when simulating the engine liquid fuel spray, the simulation includes the structural parameters and operating parameters of the engine liquid fuel atomization system. The structural parameters include the swirl chamber tilt angle, nozzle length, swirl channel rise angle, and number of swirl channels. The operating parameters include fuel temperature, injection pressure, and ambient pressure.

3. The method as described in claim 2, characterized in that, In step S101, computational fluid dynamics is used to perform numerical simulation on the sample, and the simulation data is post-processed to extract atomization characteristic parameters, including oil film thickness, atomization cone angle, and Sottle average diameter.

4. The method as described in claim 3, characterized in that, In step S105, if the model accuracy of the proxy model does not meet the requirements, Holton sampling expansion is performed, and the proxy model is retrained based on the new sample library to complete the training of the low-precision proxy model. The accuracy of the model is calculated using the coefficient of determination R. 2 To evaluate, the coefficient of determination R 2 The following method is used for calculation: Among them, SS res S is the sum of squared residuals, and SS is the sum of squared differences between observed and predicted values. tot The sum of squares is the total sum of squares, and the sum of squares of the differences between the observed values ​​and the mean of the observed values.

5. The method as described in claim 4, characterized in that, The Holden sampling expansion method continues to use the scrambled Holden sampling method to sample in the original parameter space, and the new sampling points are recalculated through simulation and added to the sample library.

6. The method as described in claim 5, characterized in that, In step S107, the optimization objective is multi-objective, including oil film thickness, atomization cone angle, and Sottle mean diameter. Based on the trained low-precision surrogate model, a large number of samples are taken in the sampling space and input into the low-precision surrogate model to calculate the objective function value. The objective function value is sorted by non-dominated order and crowding is calculated to obtain the parent of the elite solution. The parent samples are subjected to selection, crossover, and mutation operations. After multiple rounds of iteration, the Pareto front solution is finally obtained.

7. The method as described in claim 6, characterized in that, In step S109, when calculating the prediction result of the low-precision surrogate model, the prediction result is calculated using the following method: y0 = f(x; θ) Where x is the input vector, y0 is the prediction result, and θ is the model parameter, which makes y0 as close as possible to the true output y.

8. The method as described in claim 7, characterized in that, In step S111, the desired improvement function is: Where EI(x) is the expected improvement function, x is the new sample point, σ(x) is the standard deviation of the test point, μ(x) is the predicted mean of the test point, and y best For the optimal observation, Φ(·) is the cumulative distribution function of the standard normal distribution.

9. The method as described in claim 8, characterized in that, In step S113, the structural parameters and the injection parameters are obtained by inverse derivation based on the Pareto front solution, specifically including the following process: S1131: Set different weights for multiple optimization objectives to obtain the weight-objective curve; S1132: Solve for the intersection of the weight-target curve and the Pareto front to obtain the corresponding optimal solution; S1133: Based on the optimal solution, the structural parameters and the injection parameters are derived in reverse.

10. The method as described in claim 9, characterized in that, In step S113, the structural parameters and injection parameters obtained by reverse derivation are simulated and calculated to obtain actual calculation results. The actual calculation results are then compared with the expected results to calculate the actual error and complete the closed-loop verification.

Images