Multi-objective optimization design of linear oscillating actuator based on hybrid surrogate model Authors: Minghu Yu, Jiekun Lin, Yuqiu Zhang Published: 2026 ABSTRACT Thrust fluctuations, including electromagnetic and cogging forces, significantly limit the efficiency of linear oscillating actuators (LOAs) and hinder their ability to generate reciprocating driving forces and stroke ranges, presenting a complex multi-objective optimization challenge. This study introduces an innovative optimization strategy utilizing a hybrid surrogate model (HSM) in conjunction with an improved nondominated sorting genetic algorithm II (NSGA-II). The HSM integrates four complementary data-driven surrogate models through a heuristic weighting scheme based on the coefficient of determination, effectively capturing both nonlinear behaviors and local characteristics. This integration enhances the accuracy and robustness of motor performance predictions and facilitates the rapid generation of candidate solutions for optimal exploration within a broad design space. Furthermore, a modified NSGA-II improves the crowding degree model, thereby increasing the number of viable solutions in the target region. The proposed optimization strategy results in significant enhancements in average performance and a reduction in force fluctuations in LOAs, ultimately improving system controllability. Empirical validation through static and dynamic tests of the optimized prototype, coupled with finite element analysis, substantiates the efficacy of the HSM and the proposed optimization method. This research contributes a robust, data-driven framework for high-performance LOA design. The proposed strategy may offer insights for addressing similar complex multi-objective optimization problems in engineering. Introduction Traditional rotary motors and their intermediate conversion mechanisms often face issues such as low efficiency, large size, and poor reliability in high-frequency linear reciprocating drive industrial applications, such as linear compressors, active suspension systems, and vibration energy harvesters1,2. In this context, the linear oscillating actuator (LOA) emerges as a viable alternative, thanks to its compact design and absence of a crank drive. The moving-magnet LOA, in particular, offers notable benefits, such as reduced volume, high power density, and rapid dynamic response, attracting significant research interest2–7. However, the short lengths of both the stator and the mover in the LOA result in more pronounced end effects compared to other types of linear motors that feature only one shorter side. Thus thrust fluctuations restrict the mover’s stroke range and overall performance; therefore, greater emphasis on optimization is essential during the design process. Advancements in motor topology2, materials, and drive circuits3 hold the potential for more cost-effective solutions. Furthermore, optimizing key dimensions of the stator and rotor can markedly enhance LOA performance. Primary optimization goals include increasing electromagnetic (EM) force for greater power density, minimizing the nonlinear effects of lateral and cogging force fluctuations on motion, and refining the back electromotive force (EMF) waveform to boost system controllability. These optimization efforts generally rely on finite element analysis (FEA) which requires extensive scans of various design parameters4–7, but may become inefficient and computationally demanding when dealing with multiple objectives and a high-dimensional design space. To enhance optimization reliability and efficiency, researchers have turned to intelligent algorithms, such as Multi-Objective Particle Swarm Optimization8 and Multi-Objective Genetic Algorithms9–12, which have shown effectiveness when integrated with FEA. These algorithms optimize crucial performance indicators, such as output torque, cogging torque, vibration noise, and thermal characteristics, across various rotary motor types. However, in the previous research on multi-objective optimization design of motors, a notable limitation persists: The crowding mechanism of traditional intelligent algorithms takes the extremum form as the optimization objectives, which has limitations and low flexibility in the multi-objective optimization designs that take the interval form as the optimization objectives. In addition, considering that the reliance on FEA still impose significant computational costs, data-driven surrogate models have begun to replace traditional FEA approaches, categorized into statistical models (SM)13,14, machine learning (ML)15,16, and deep learning (DL)17,18. These surrogate models have been widely applied to various engineering applications for its high computational efficiency and generalization ability19–21. However, SM offer rapid approximations using explicit functions but struggle with high-dimensional data and complex nonlinear relationships. ML techniques improve prediction accuracy but can be prone to overfitting in low-dimensional contexts. DL methods, utilizing multi-layer neural networks, excel in handling complex high-dimensional mappings but require substantial data and computation, particularly at risk of overfitting with smaller datasets. Thus, achieving a balance between efficiency and accuracy in high-dimensional, nonlinear multi-objective optimization remains complex. And the substantial nonlinearity of the ferromagnetic materials, along with the numerous topological structure parameters and the strongly coupled nonlinear relationships among these parameters and the output functions, results in a high-dimensional and highly nonlinear multi-objective optimization problem. To address the aforementioned issues, hybrid surrogate models are introduced into the multi-objective optimization of passenger car rear seats. Multiple sub-surrogate models are employed to predict multi-objective functions, with the optimal-performing sub-surrogate model selected for each objective function based on prediction errors22–24. Furthermore, heuristic weighting schemes are proposed based on prediction errors. By weighted combination of multiple surrogate models according to prediction errors, more robust hybrid surrogate models are constructed25. In addition, genetic algorithms are utilized to perform multiple rounds of iterative optimization on weighting coefficient combinations, which effectively enhances the prediction accuracy of hybrid surrogate models26. Inspired by the above research, this paper proposes a novel multi-objective optimization strategy based on hybrid surrogate models (HSM) and an improved crowding distance mechanism, aimed at accurately predicting the EM performance of a dual E-core LOA (DELOA) during operation, promising faster runtime and enhanced robustness. Different from the aforementioned heuristic weighting schemes25,26, the proposed HSM comprehensively integrates four distinct data-driven sub-models from statistical models (SM), machine learning (ML), and deep learning (DL). It aims to provide fast approximations through explicit functions while being applicable to the high-dimensional, strongly coupled nonlinear characteristics of motor models. Secondly, the weights of the sub-models are allocated based on the coefficient of determination, avoiding complex evolutionary algorithms or uncertainty quantification processes and ensuring that the hybrid model achieves high prediction accuracy while minimizing computational costs. Moreover, an improved crowding degree model using in the nondominated sorting genetic algorithm II (NSGA-II) enhances the solution density within effective target regions, increasing optimization efficiency without additional iterations. This optimization strategy leads to significant improvements in average performance and mitigation of force fluctuations in LOAs, thus enhancing system controllability. The accuracy and effectiveness of the proposed model are validated through FEA and experimental results in this study, suggesting its potential applicability in other fields facing multi-objective optimization challenges. Principles and performance of DELOA Operating principles and modeling To enable the piston of the refrigeration linear compressor to achieve high-frequency reciprocating linear motion, DELOA, as illustrated in Fig. 1, is employed in conjunction with mechanical springs on either side of the mover to form the driving system. Fig. 1DELOA topology diagram. DELOA topology diagram. Two armature windings are arranged on E-type stator cores, which are stacked along the z-axis and placed symmetrically in x-axis opposite orientations. Two PMs, magnetized oppositely along the y-axis, are embedded in the non-magnetic mover bracket situated within the air gap. The interaction between the armature magnetic field and the magnetic field of the PMs generates a force that propels the piston in the x-direction, while the normal attraction in the y-direction cancels out due to symmetry, thereby preventing y-direction offsets and vibrations. The mover bracket connects to mechanical springs on both sides, facilitating potential energy conversion and establishing a mechanical resonance system for the entire DELOA. When sinusoidal excitation is applied, the mover produces linear reciprocating motion, with the system achieving maximum efficiency when the electrical frequency aligns with the mechanical resonance frequency. The circuit model for armature winding and the force conditions acting on mover are established as illustrated in Fig. 2. Then, the equivalent voltage equation and equivalent mechanical equation can be derived as27: Fig. 2Equivalent model of DELOA. (a) Equivalent circuit. (b) Mechanism. 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$u\left( t \right)=Ri\left( t \right)+{{Ldi\left( t \right)} \mathord{\left/ {\vphantom {{Ldi\left( t \right)} {dt}}} \right. \kern-0pt} {dt}}+e\left( t \right)$$\end{document} 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_i}\left( t \right)+{F_k}+{F_f}+{F_g}=ma$$\end{document} Equivalent model of DELOA. (a) Equivalent circuit. (b) Mechanism. where t, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$u\left( t \right)$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i\left( t \right)$$\end{document}, R, L represent time, terminal voltage, current, resistance, inductance, respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$e\left( t \right)$$\end{document} is the back EMF, which satisfies \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$e={K_i}v$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_i}$$\end{document} is the electromagnetic thrust coefficient and v is the speed of mover. m is the total mass of mover, a is the acceleration of mover, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_i}\left( t \right)$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_k}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_g}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_f}$$\end{document} represent the total thrust provided by the motor, spring force, nonlinear gas force, frictional force, respectively. The total thrust provided by motor consists of the EM force \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_e}$$\end{document} between the magnet and the armature field, along with the reluctance force \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_{cog}}$$\end{document} between the magnet and the stator core28, as expressed in:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_i}\left( t \right)={F_e}+{F_{cog}}={K_i}i\left( t \right)+{F_{cog}}$$\end{document} The main indicators of the DELOA, derived from practical linear compressor requirements, are detailed in Table 1 together with the overall dimensions of the mass production driving DELOA prototype of linear compressor for refrigerator. Table 1Main indicators and dimensions of mass production DELOA prototype.ParameterValueRated current (A)2Rated frequency (Hz)56EM force coefficient (N/A)32Mechanical spring stiffness (N/mm)22Maximum stroke of the mover (mm) ± 8 Overall x, y, z dimensions of the prototype (mm)15.60, 5.24, 20.7 Table 1Main indicators and dimensions of mass production DELOA prototype.ParameterValueRated current (A)2Rated frequency (Hz)56EM force coefficient (N/A)32Mechanical spring stiffness (N/mm)22Maximum stroke of the mover (mm) ± 8 Overall x, y, z dimensions of the prototype (mm)15.60, 5.24, 20.7 Main indicators and dimensions of mass production DELOA prototype. Thrust performance analysis by FEA The 2-dimensional FEA transient model of the mass production DELOA prototype was developed from the initial values of the design parameters lisited in Table 2 of Section Ⅱ to obtain accurate thrust performance using Ansys software. When a direct current of 0 A and 1 A is applied to the stator winding, the distributions of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_{cog}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_i}$$\end{document} at various positions throughout the stroke range can be obtained as illustrated in Fig. 3. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_{cog}}$$\end{document} exhibits approximate anti-symmetry around the equilibrium position(x=0mm) for the symmetrical structure design of DELOA. In addition, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_{cog}}$$\end{document} at the middle section of the oscillation stroke has relatively small fluctuations, while significant amplitude changes occur at the stroke ends, with a peak-to-peak value of 117.9 N for the more obvious cogging effect and end effect of DELOA without tooth crown. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_i}$$\end{document} shows similar trends as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_{cog}}$$\end{document}, with relatively flat middle section and increased amplitude and fluctuations at endpoints due to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_e}$$\end{document}. Fig. 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_{cog}}$$\end{document} distributions of initial prototype. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_{cog}}$$\end{document} distributions of initial prototype. Then, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_i}$$\end{document} is calculated according (3) and shown in Fig. 4. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_i}$$\end{document} has even symmetry with a mean value of 29.5 N/A within the stroke range, which is smaller than the requirement of linear compressor. And there have been special issues that greatly affect the linearization of the EM force provided by the motor. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_i}$$\end{document} drops drastically when the position exceeds ± 2 mm, with a peak-to-peak value of 20.9 N/A. The coefficient value on the right side of the equilibrium point decreases more rapidly than that on the left side, accompanied by a greater amplitude change. This phenomenon is attributed to the nonlinear increase in the saturation level of the iron core, which results from the superposition of the armature magnetic field and the magnet’s magnetic field when a positive current is applied. Fig. 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_i}$$\end{document} distribution of initial prototype. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_i}$$\end{document} distribution of initial prototype. These findings from Figs. 3 and 4 indicate that the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_e}$$\end{document} can enable the mover to achieve reciprocating motion, while they do not linearly correlate with current variations, complicating the stroke control of DELOA, amplifying nonlinearity, and increasing the difficulty of precisely controlling displacement and speed. Data-driven multi-objective optimization of DELOA based on HSM Significant thrust performance defects for high performance control appear in the original mass production prototype, necessitating improvements to the thrust of DELOA. The optimization design for DELOA is a high-dimensional and highly nonlinear multi-objective optimization problem. This section will detail the new data-driven multi-objective optimization strategy illustrated in Fig. 5 based on HSM, which can more accurately predict the EM performance of DELOA within its stroke, with higher optimization efficiency. Fig. 5Flowchart of multi-objective optimization of DELOA. Flowchart of multi-objective optimization of DELOA. Optimization objectives and parameter settings For a LOA, thrust is an important indicator28. Reducing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_{cog}}$$\end{document} and the fluctuation of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_i}$$\end{document} help to improve the linearity of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_i}$$\end{document}, which is beneficial for estimation of mover stroke and response speed. From the above analysis, there are significant defects of thrust in the mass production prototype, making it difficult to meet high performance operational requirements. Thus, the optimization goals are to minimize the peak-to-peak values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_{cog}}$$\end{document}, denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p2p.{K_i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p2p.{F_{cog}}$$\end{document}, while maintaining the mean value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_i}$$\end{document}, denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$avg.{K_i}$$\end{document}, at 32 N/A. The objective functions selected are as follows:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Function:\left\{ \begin{gathered} {f_1}\left( {avg.{K_i}} \right) \to 32{N \mathord{\left/ {\vphantom {N A}} \right. \kern-0pt} A}, \hfill \\ Min{f_2}\left( {p2p.{K_i}} \right), \hfill \\ Min{f_3}\left( {p2p.{F_{cog}}} \right) \hfill \\ \end{gathered} \right.$$\end{document} The parameter model of DELOA are shown in Fig. 5, with a total of 9 parameters, the initial value of which are given in Table 2. These parameters collectively affect the magnetic reluctance of the DELOA, thereby altering the thrust performance, and are therefore set as optimization parameters. After considering the manufacturing process, volume and cost comprehensively, the optimization range of these parameters are also listed in Table 2 to reduce invalid cases calculated in the optimization progress. Table 2Initial values and optimization range of parameters.Optimization variablesParametersInitial valuesRangesP1(mm) lslot 10.7010–18P2(mm) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l_1}$$\end{document} 00–6P3(mm) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h_1}$$\end{document} 00–6P4(mm) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h_2}$$\end{document} 00–6P5(mm) lpm 15.6015–23P6(mm) lpms 6.350.5–6.5P7(mm) lout 10.7512–18P8(mm) loutend 15.6012–17P9(mm) hsback 11.7510–14 Table 2Initial values and optimization range of parameters.Optimization variablesParametersInitial valuesRangesP1(mm) lslot 10.7010–18P2(mm) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l_1}$$\end{document} 00–6P3(mm) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h_1}$$\end{document} 00–6P4(mm) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h_2}$$\end{document} 00–6P5(mm) lpm 15.6015–23P6(mm) lpms 6.350.5–6.5P7(mm) lout 10.7512–18P8(mm) loutend 15.6012–17P9(mm) hsback 11.7510–14 Initial values and optimization range of parameters. Pearson sensitivity analysis and dimensionality reduction As previously indicated, the initial design parameters are considerable, totaling 9, which significantly increases the sample size required for multi-objective optimization. Although the electromagnetic characteristics of linear oscillating actuator exhibits high nonlinearity within the local parameter range, there is often a dominant global monotonic trend between key parameters and primary performance objectives when evaluated from the entire design space. Therefore, a design parameter that has a substantial impact on performance usually exhibits a certain directional correlation in Pearson correlation analysis. Therefore, to mitigate computational costs, Pearson sensitivity analysis is employed for dimensionality reduction within the design space. All correlation coefficients are shown in Fig. 6. Within the constraint range, the three objective functions are influenced to varying degrees by the 9 parameters, resulting in inconsistent variation patterns among the objectives. The correlation coefficients of P8 and P9 to the three objective functions are all very low, which indicates that the changes in these 2 parameters within the range have the lowest impact on the thrust. Therefore, P8 and P9 are disregarded, thereby reducing the original optimization problem from 9 design parameters to 7. Fig. 6Pearson sensitivity analysis results. Pearson sensitivity analysis results. Heuristic construction of HSM for DELOA The data-driven surrogate model offers significant cost advantages over the EM model developed using the FEA. However, a single surrogate model often proves inadequate for multi-input multi-output optimization problems. To address this, a HSM was constructed through a heuristic weighted combination of multiple surrogate models based on the determination coefficient. Four data-driven sub-models are selected to form HSM of DELOA, including the Response Surface Model (RSM), Kriging Model (KM), Back Propagation Neural Network (BPNN), and Support Vector Regression (SVR), all of which have been proven effective in predicting the EM characteristics of other kind of motors. These models encompass the previously mentioned SM, ML, and DL, leveraging their complementary strengths to improve the overall prediction accuracy of the HSM. RSM, based on polynomial regression, establishes a second-order functional relationship between observation points and the objective function8. It is computationally efficient but limited in fitting highly nonlinear data KM, relying on spatial interpolation, predicts the spatial mapping between observation points and the objective function, making it highly effective for spatially distributed data9. However, its performance depends on covariance function selection and parameter estimation, which may introduce interpolation errors in sparse data regions. SVR incorporates an insensitive loss function during regression, penalizing errors that exceed a predefined threshold, thereby enhancing generalization and reducing the risk of over fitting15. However, the threshold is often set empirically, lacking theoretical support, and it is difficult to find the optimal balance. BPNN, based on a multilayer feedforward neural network, adjusts weights using the back propagation algorithm to minimize prediction errors29. Its nonlinear activation functions enable accurate capture of complex relationships in the data but introduce a risk of over fitting. By combining the advantages of these four sub-models, the proposed HSM enhances the fitting of nonlinear data, reduces dependence on model parameters, improves generalization, and minimizes overfitting, enabling effective and accurate predictions for multiple DELOA optimization objectives. Its fundamental constructed form is defined as follows:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{gathered} f_{i}^{{HSM}}=\sum\limits_{{j=1}}^{N} {{w_j}f_{i}^{{SMj}}\left( p \right)} , \hfill \\ \sum\limits_{{j=1}}^{N} {{w_j}=1} \hfill \\ \end{gathered} \right.$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{i}^{{HSM}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{i}^{{SMj}}\left( p \right)$$\end{document} is the predicted value of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i{\mathrm{th}}$$\end{document} objective function as determined by the HSM and the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$j{\mathrm{th}}$$\end{document} sub-model, respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${w_j}$$\end{document}is the weight of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$j{\mathrm{th}}$$\end{document} sub-model, N is the total number of sub-models, and p is a vector encompassing the 7 design parameters. The predictive ability of model is quantified by the determination coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R^2}$$\end{document}, with values close to 1 indicating better fit with existing data. Recognizing that some sub-models may exhibit unusually high or low values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R^2}$$\end{document} due to random factors, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${w_j}$$\end{document} is composed of two components: the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R^2}$$\end{document} of the individual sub-model and the mean values of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R^2}$$\end{document} of all sub-models, as follows:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{gathered} {w_j} = {{w_j^*} \mathord{\left/ {\vphantom {{w_j^*} {\sum\limits_{k = 1}^N {w_k^*} }}} \right. \kern-\nulldelimiterspace} {\sum\limits_{k = 1}^N {w_k^*} }},w_j^* = {\left( {R_j^2 + \alpha \overline {{R^2}} } \right)^\beta },\overline {{R^2}} = {{\sum\limits_{j = 1}^N {R_j^2} } \mathord{\left/ {\vphantom {{\sum\limits_{j = 1}^N {R_j^2} } N}} \right. \kern-\nulldelimiterspace} N}, \hfill \\ R_j^2 = 1 - {{\sum\limits_{m = 1}^M {{{\left( {f_i^{FEA}\left( {{p_m}} \right) - f_i^{SM\hat j}\left( {{p_m}} \right)} \right)}^2}} } \mathord{\left/ {\vphantom {{\sum\limits_{m = 1}^M {{{\left( {f_i^{FEA}\left( {{p_m}} \right) - f_i^{SM\hat j}\left( {{p_m}} \right)} \right)}^2}} } {}}} \right. \kern-\nulldelimiterspace} {}} \hfill \\ \left( {\sum\limits_{m = 1}^M {{{\left( {f_i^{FEA}\left( {{p_m}} \right) - {{\sum\limits_{m = 1}^M {f_i^{FEA}\left( {{p_m}} \right)} } \mathord{\left/ {\vphantom {{\sum\limits_{m = 1}^M {f_i^{FEA}\left( {{p_m}} \right)} } M}} \right. \kern-\nulldelimiterspace} M}} \right)}^2}} } \right) \hfill \\ \end{gathered} \right.$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R_{j}^{2}$$\end{document} is the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R^2}$$\end{document} of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$j{\mathrm{th}}$$\end{document} sub-model, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${p_m}$$\end{document} is the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$m{\mathrm{th}}$$\end{document} sample point, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{i}^{{FEA}}\left( {{p_m}} \right)$$\end{document} is the computed value of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i{\mathrm{th}}$$\end{document} objective function based on FEA for the sample point \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${p_m}$$\end{document}, M is the total number of sample points, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_i^{SM\hat j}\left( {{p_m}} \right)$$\end{document} is the predicted value of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i{\mathrm{th}}$$\end{document} objective function obtained from the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$j{\mathrm{th}}$$\end{document} sub-model, constructed using all sample points except for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${p_m}$$\end{document}. α is used to adjust the balance between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R^2}$$\end{document} of the sub-model and the average \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R^2}$$\end{document}, which prevents the excessively amplified weight of an individual sub-model that exhibits an abnormally high \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R^2}$$\end{document} value due to local overfitting and thus enhances the overall robustness of the hybrid model. Therefore, an excessively high value of α will over-smooths the performance differences among sub-models and lead to weight convergence, whereas an excessively low value will weaken the regularization effect. Finally, among four discrete candidate values (0.5, 0.1, 0.05 and 0.01), the value of 0.05, which achieves a good balance between performance and robustness, is selected as the final value of α. And β is used to regulate the dispersion of the sub-model weight allocation. The value of β is fixed at 1, which ensures that the final weights directly reflect the performance ranking of each sub-model after adjustment by α, thus facilitating a clear comparison and interpretation of the contribution of each sub-model in subsequent analyses. This approach reduces the impact of extreme local characteristics on the HSM’s predictive accuracy. To verify the effectiveness and robustness of the four surrogate sub-models and the proposed HSM, Latin Hypercube Sampling (LHS) was employed to extract 1,000 sample points from the 7-parameter design space after dimensionality reduction. Objective values for each sample motor were computed using FEA and subsequently utilized to form training and testing datasets in a 4:1 ratio for parameter training and performance evaluation. The primary time cost associated with training a surrogate model arises from FEA computation, making the training time for the proposed HSM comparable to that of a single sub-model. The prediction errors of both the proposed HSM and the individual surrogate models for random samples are presented in Fig. 7, while the average prediction errors are summarized in Table 3. The collaborative integration of four distinct types of surrogate models within the HSM capitalizes on their strengths in addressing complex nonlinear issues related to motors, resulting in a significant enhancement in predictive accuracy. This improvement is depicted in Fig. 7, which contrasts the predicted data from multiple randomly sampled prototypes. Additionally, as shown in Table 3, this method exhibits a lower overall average prediction error. These findings imply that the proposed DELOA HSM combination method is not only justified but also demonstrates enhanced generalization capabilities, allowing for adaptation to diverse operating conditions and variables. Furthermore, it showcases improved robustness and interference resistance, effectively capturing intricate nonlinear relationships among objectives and facilitating the swift identification of optimal trade-off solutions in the multi-objective optimization of DELOA. Fig. 7The prediction errors of sub-models and HSM for different motors. (a) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$avg.{K_i}$$\end{document}.(b) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p2p.{K_i}$$\end{document}.(c) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p2p.{F_{cog}}$$\end{document}. The prediction errors of sub-models and HSM for different motors. (a) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$avg.{K_i}$$\end{document}.(b) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p2p.{K_i}$$\end{document}.(c) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p2p.{F_{cog}}$$\end{document}. Table 3Average Prediction Errors of Sub-models and Proposed HSM.Average prediction error \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_1}(p2p.{F_{cog}})$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_2}(avg.{K_i})$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_3}(p2p.{K_i})$$\end{document} RSM 9.62452.36792.8318 KM 5.17172.00791.7834 BPNN 7.21351.92432.8771 SVR 5.89832.00792.6751 HSM 1.84661.27881.4445 Table 3Average Prediction Errors of Sub-models and Proposed HSM.Average prediction error \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_1}(p2p.{F_{cog}})$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_2}(avg.{K_i})$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_3}(p2p.{K_i})$$\end{document} RSM 9.62452.36792.8318 KM 5.17172.00791.7834 BPNN 7.21351.92432.8771 SVR 5.89832.00792.6751 HSM 1.84661.27881.4445 Average Prediction Errors of Sub-models and Proposed HSM. Improved multi-objective optimization algorithm NSGA-II is an intelligent optimization algorithm that generates new populations using genetic algorithms and selects and retains them based on a crowding degree ranking mechanism. The new populations evolves into a higher-performing population through selection, crossover, and mutation. The improved NSGA-II for multi-objective optimization of DELOA consists of three components: genetic algorithms, non-dominated sorting, and crowding degree sorting as follows. Genetic algorithm for generating new individuals: Genetic algorithm for generating new individuals: Genetic algorithm for generating new individuals: The genetic algorithm generates new individuals while preserving characteristics of superior solutions through crossover and mutation. Crossover involves randomly selecting 60% of samples to form pairs that exchange parameter variables, creating new motor design solutions. Mutation simulates gene changes by randomly adjusting an individual’s features with a specific probability to avoid local optima. In each iteration, 5% of the samples are selected, and one variable from each is adjusted using a normally distributed random number, rounded and scaled by the mutation step size. 2)Non-dominated sorting of the sample population: 2)Non-dominated sorting of the sample population: Non-dominated sorting of the sample population: In the motor sample population, sample A dominates sample B, if they satisfy the following condition: 7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{gathered} {f_1}\left( A \right) \geqslant {f_1}\left( B \right),{f_2}\left( A \right) \leqslant {f_2}\left( B \right),{f_3}\left( A \right) \leqslant {f_3}\left( B \right), \hfill \\ {f_1}\left( A \right)>{f_1}\left( B \right) \cup {f_2}\left( A \right)<{f_2}\left( B \right) \cup {f_3}\left( A \right)<{f_3}\left( B \right) \hfill \\ \end{gathered} \right.$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_i}\left( X \right)$$\end{document} represents the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i{\mathrm{th}}$$\end{document} objective function of sample X. That means the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$avg.{K_i}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p2p.{K_i}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p2p.{F_{cog}}$$\end{document} for sample A are not inferior to those of sample B, with at least one objective being superior. Based on the dominance relationship, the samples are classified into layers, with the first layer referred to as the Pareto front. 3)Sorting of sample crowding degree: 3)Sorting of sample crowding degree: Sorting of sample crowding degree: To ensure the diversity of the Pareto front, the traditional NSGA-II method calculates the crowding degree of samples following non-dominated sorting. This calculation reflects the distribution of adjacent points around each sample; thus, a lower crowding degree indicates a sparser surrounding area. By prioritizing samples with lower crowding degrees, the Pareto front can effectively expand in three-dimensional space after a specified number of genetic iterations, which are defined by the number of optimization objectives as dimensions. As the iterations progress, the distance between the optimization parameters of the Pareto front samples in each dimension increases, leading to a gradual expansion of the surface formed. To optimize the objective function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_1}$$\end{document} of the samples within the final Pareto front around the design target of 32 N/A, the calculation of crowding degree has been updated based on the proportion of the distance between the output and target values as follows: 8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{gathered} {{D\left[ i \right]= - \left( {D\left[ i \right].{f_1} - 32} \right)} \mathord{\left/ {\vphantom {{D\left[ i \right]= - \left( {D\left[ i \right].{f_1} - 32} \right)} {\left( {f_{1}^{{\hbox{max} }} - f_{1}^{{\hbox{min} }}} \right)}}} \right. \kern-0pt} {\left( {f_{1}^{{\hbox{max} }} - f_{1}^{{\hbox{min} }}} \right)}}+ \hfill \\ {{\sum\limits_{{m=2}}^{{n - 1}} {\left( {D\left[ {i+1} \right].{f_m} - D\left[ {i - 1} \right].{f_m}} \right)} } \mathord{\left/ {\vphantom {{\sum\limits_{{m=2}}^{{n - 1}} {\left( {D\left[ {i+1} \right].{f_m} - D\left[ {i - 1} \right].{f_m}} \right)} } {\left( {f_{m}^{{\hbox{max} }} - f_{m}^{{\hbox{min} }}} \right)}}} \right. \kern-0pt} {\left( {f_{m}^{{\hbox{max} }} - f_{m}^{{\hbox{min} }}} \right)}} \hfill \\ \end{gathered}$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D\left[ i \right]$$\end{document} denotes the crowding degree of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i{\mathrm{th}}$$\end{document} sample, comprising 2 components: the first is the ratio of the distance from the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$avg.{K_i}$$\end{document} of the sample to 32 N/A, relative to the maximum distance; the second component represents the crowding degree of the remaining optimization objectives. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D\left[ i \right].{f_m}$$\end{document} indicates the value of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$m{\mathrm{th}}$$\end{document} objective function for the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i{\mathrm{th}}$$\end{document} sample, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{m}^{{\hbox{max} }}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{m}^{{\hbox{min} }}$$\end{document} denote the maximum and minimum values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_m}$$\end{document} within the sample population. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D\left[ i \right]$$\end{document} is subsequently sorted in descending order. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_1}$$\end{document} of the sample group is generated near 32 N/A, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_2}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_3}$$\end{document} being differentiated as much as possible to enhance the diversity of the motor design solutions, thereby avoiding premature convergence to a single point. In addition, crowding distance of each solution remains determined collectively by all three objective functions both before and after the modification, which means that the trade-off among \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_1}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_2}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_3}$$\end{document} is preserved in the selection process. And solutions that perform poorly in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_2}$$\end{document} or \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_3}$$\end{document} will still exhibit larger crowding distances due to their contributions from these objectives, thereby maintaining the diversity and convergence quality of the Pareto front across all objectives. To verify the effectiveness of the improved crowding distance mechanism, this study uses a 50-iteration scheme to evaluate the optimization performance of NSGA-II before and after improvements. In each iteration, the new population undergoes non-dominated sorting and crowding degree sorting sequentially. When the iteration count reaches 50, the algorithm terminates, with the top 120 samples forming the final Pareto front. Fig. 8 illustrates the Pareto fronts obtained before and after refinement of the crowding distance mechanism, together with a comparative overview. As shown in Fig. 8a and c, the conventional NSGA-II produces a progressively expanding Pareto front over the 50-generation run, distributing motor solutions across a broad \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$avg.{K_i}$$\end{document} range of 25–45 N/A. In contrast, Fig. 8b and c demonstrate that the improved NSGA-II yields a more convergent distribution, with samples clustering tightly within 31–35 N/A and progressively contracting toward the target \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$avg.{K_i}$$\end{document} value of 32 N/A. This distinct behavior indicates that the enhanced crowding distance mechanism effectively maintains a high density of selectable solutions within the desired target region. Consequently, this approach facilitates the identification of the optimal motor design and significantly improves optimization efficiency. Fig. 8The Pareto front of the 50-iteration scheme by traditional NSGA-II and improved NSGA-II. (a). Traditional NSGA-II. (b). Improved NSGA-II. (c). Comparison. The Pareto front of the 50-iteration scheme by traditional NSGA-II and improved NSGA-II. (a). Traditional NSGA-II. (b). Improved NSGA-II. (c). Comparison. Discussion of optimization results Since the surrogate model is trained prior to the NSGA-II process, and the genetic algorithm has a fixed population size and mutation rate, multiple optimization validations have shown that the resulting Pareto front exhibits high similarity and the distribution and range of the samples tend to stabilize after 50 iterations. Therefore, the proposed point is selected from the 50th generation, as shown in Fig. 8b. Taking into account a certain margin of prediction error for HSM, the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f_1}$$\end{document} value of the proposed point marginally exceeds the design standard. The proposed optimized point is indicated on the Pareto front in Fig. 8b, with their corresponding design parameters detailed in Table 4. Considering the machining precision of DELOA, the dimensions are standardized as the final optimized dimensions provided in Table 4. Table 4Parameter values of the proposed point and optimized scheme.ParameterProposed pointOptimized schemeP1(mm)14.086614.08P2(mm)4.55134.55P3(mm)4.44294.44P4(mm)3.19963.20P5(mm)22.448522.45P6(mm)5.30015.30P7(mm)15.333515.33 Table 4Parameter values of the proposed point and optimized scheme.ParameterProposed pointOptimized schemeP1(mm)14.086614.08P2(mm)4.55134.55P3(mm)4.44294.44P4(mm)3.19963.20P5(mm)22.448522.45P6(mm)5.30015.30P7(mm)15.333515.33 Parameter values of the proposed point and optimized scheme. To validate the accuracy of the HSM and the effectiveness of the optimization strategy, the final optimized scheme is incorporated into FEA model. The distributions of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_i}$$\end{document}and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_{cog}}$$\end{document} obtained from FEA before and after optimization within the stroke range are shown in Figs. 9 and 10, and the objective function values are listed in Table 5. As shown in Fig. 9, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_i}$$\end{document} of the initial design is slightly higher than the optimized scheme within the − 3 mm to 3 mm range. However, the average values across the entire stroke are similar, and the fluctuations of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_i}$$\end{document} significantly decreased after optimization. This finding is further corroborated by the comparison in Table 5. The optimized \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$avg.{K_i}$$\end{document}is 32.13 N/A, meeting the design requirements and showing a 8.84% increase compared to the initial design. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p2p.{K_i}$$\end{document} is 5.72 N/A, reflecting a reduction of approximately 72.59%. The optimization of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_{cog}}$$\end{document} is also apparent, as shown in Fig. 10; Table 5, where the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p2p.{F_{cog}}$$\end{document} decreases by 88.18%. The primary reason for the differences between the optimized prototype and the original prototype is the incorporation of enhanced E-type core teeth and crowns, along with improvements in the width of the end teeth and the length of the magnets. These modifications effectively mitigate the inherent end effect of linear motors, leading to reduced variations in no-load thrust and electromagnetic force during operation when the mover approaches the edge of the stator. Additionally, the optimized \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_i}$$\end{document} shown in Fig. 9 exhibits a significant improvement in symmetry on both sides of the equilibrium point, which suggests a reduction in the asymmetric nonlinear effects of ferromagnetic materials on the magnetic field distribution. As a result, within the stroke range, the relationship between current and thrust becomes more linear, which is a favorable condition for precise control. Overall, these modifications signify a substantial enhancement in the output characteristics of the DELOA, yielding smoother output thrust and a increase in the efficiency of energy conversion from electrical to mechanical. Fig. 9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_i}$$\end{document} distribution of initial prototype and optimized scheme. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_i}$$\end{document} distribution of initial prototype and optimized scheme. Fig. 10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_{cog}}$$\end{document} distribution of initial prototype and optimized scheme. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_{cog}}$$\end{document} distribution of initial prototype and optimized scheme. Table 5Objective function values before and after optimization.Objective functionInitial prototype (FEA)Optimized scheme (FEA)Optimized scheme (HSM)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$avg.{K_i}$$\end{document}(N/A)29.522632.131433.7614\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p2p.{K_i}$$\end{document}(N/A)20.86225.71906.3590\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p2p.{F_{cog}}$$\end{document}(N)117.932213.939415.5094 Table 5Objective function values before and after optimization.Objective functionInitial prototype (FEA)Optimized scheme (FEA)Optimized scheme (HSM)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$avg.{K_i}$$\end{document}(N/A)29.522632.131433.7614\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p2p.{K_i}$$\end{document}(N/A)20.86225.71906.3590\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p2p.{F_{cog}}$$\end{document}(N)117.932213.939415.5094 Objective function values before and after optimization. The predicted objective function values of the optimized DELOA from HSM is also presented in Table 5. The absolute error values of the three objectives between HSM and FEA are 1.57 N/A, 1.63 N/A, and 0.64 N, respectively, with the prediction accuracy consistently maintained at a high level. This fully demonstrates the stability and reliability of the proposed HSM for predicting the electromagnetic performance of linear oscillating actuators. Experimental verification Following the optimization process and verification via FEA, the optimized DELOA was fabricated, and an experimental platform was developed as shown in Fig. 11. This platform includes the optimized DELOA prototype functioning as the driving motor, a KTR11 linear position sensor for capturing the actual mover position signal, a 3.6 kW DC power supply, a driving circuit board with a designated switching frequency of 10 kHz, and a control circuit board with an STM32H743 microcontroller operating at a core frequency of 480 MHz. Comprehensive static and dynamic performance tests of the motor have been conducted, confirming both the model and the optimization effects. Fig. 11Experimental platform. Experimental platform. A series of static experiments were conducted by directly connecting a DC power supply to the actuator’s winding, causing the mover to shift a specific distance from its zero position. Since the mover is connected to two mechanical springs and is not linked to a compressor piston, the thrust acting on the DELOA can be calculated using (2), where only the spring force is considered, with both friction and nonlinear gas forces is zero. As illustrated in Figs. 3 and 4, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F_{cog}}$$\end{document} exhibit even and odd function characteristics, respectively, in relation to the zero position. By superimposing and averaging the recorded total thrust data from the negative and positive segments of the stroke, the actual \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_i}$$\end{document} across the entire stroke range were obtained. Its distribution during the forward stroke is presented in Fig. 12, together with the FEA results. The measured EM force coefficient has an average value of 31.55 N/A, maintaining a high level throughout the stroke, with a fluctuation of 6.70 N/A. This demonstrates the motor’s capability to continuously provide stable thrust under all operating conditions. In 0–6 mm, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_i}$$\end{document} remains relatively constant, which indicates a more distinct linear relationship between thrust and current. Furthermore, the FEA data closely aligns with the experimental results, demonstrating a high degree of correlation within 0–6 mm. In 6–8 mm, however, the measured data shows a slight decrease. This deviation is attributed to the presence of leakage flux at the mover’s end, which leads to minor discrepancies between the measured data and the FEA results. Thus, this overall behavior demonstrates the motor’ s capability to continuously provide stable thrust under all operating conditions. Across the entire stroke range, the error of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$avg.{K_i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p2p.{K_i}$$\end{document} are 0.58 and 0.98 N/A, respectively, indicating a high level of agreement. Thus, the static tests not only validate the accuracy of the HSM and FEA predictions in Table 5, but also confirm the linear drive characteristics of the optimized DELOA prototype. Fig. 12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_i}$$\end{document} distributions of the optimized prototype from static experiment and FEA. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_i}$$\end{document} distributions of the optimized prototype from static experiment and FEA. Then the circuit boards driving and implementing resonant frequency tracking control were used during dynamic experiments. The resonant frequency of the LOA and springs is designed to be 56 Hz under no-load conditions. Thus sinusoidal currents with frequencies of 30 Hz, 56 Hz, and 70 Hz were fed to the winding to verify the resonant characteristics of the LOA prototype, maintaining a constant peak value of 2 A. The current and position signals were collected synchronously and depicted in Fig. 13a, b, and c. When the current is sinusoidal, the displacement of the mover also exhibits a corresponding sinusoidal characteristic. This observation indicates that throughout the entire stroke under the three operational conditions, the EM driving force provided by the motor is essentially proportional to the current, with the EM force coefficient remaining relatively constant. At 56 Hz in (b), the phase differences between the two signals was measured at 90.72°, which is approximately 90° and occurs when the driving current frequency matches the mechanical system resonant frequency30. The actual resonant frequency aligns closely with the design specifications, demonstrating that the influence of cogging force on the equivalent spring coefficient is minimal, while fluctuations of the cogging force remain low following optimization. By comparing the three position waveforms in Fig. 13, it can be observed that under the same input current amplitude, the mover’s oscillation amplitude is below 0.5 mm in (a) and (c), corresponding to the non-resonant state. Whereas in the resonant state in Fig. 13b, the amplitude approaches 8 mm, significantly higher than the non-resonant state. This demonstrates that the system achieves the highest energy conversion efficiency in the resonant state. Fig. 13The position and current signals obtained from dynamic experiment. (a) 30 Hz current. (b) 56 Hz current. (c) 70 Hz current. The position and current signals obtained from dynamic experiment. (a) 30 Hz current. (b) 56 Hz current. (c) 70 Hz current. Thus, the static and dynamic experiments not only support the conclusions above but also thoroughly validate the outstanding EM performance, stability, and resonant characterics of the DELOA prototype developed in this study. Conclusion This paper investigates the thrust characteristics of a DELOA and presents an improved multi‑objective optimization design method that integrates a data‑driven hybrid surrogate model (HSM). The prediction capability of the HSM and the effectiveness of the optimization approach are evaluated through FEA and experimental measurements. The main findings are summarized as follows: Due to the short lengths of both the stator and the mover in the LOA, end effects are pronounced; therefore, special attention must be given to thrust performance during the design process. By optimizing key dimensions, such as the stator’s tooth crown and end tooth width, as well as the length of the mover’s magnets, end effects can be effectively suppressed. This minimizes the significant fluctuations in thrust, both under no-load and load conditions, when the mover approaches the stator’s edge. This improvement enhances the linearity of the EM force coefficient, thereby supporting more precise control of linear reciprocating motion.By integrating various data-driven sub-models with a coefficient-of-determination-based weighting mechanism, a high-accuracy HSM for DELOA is developed. The HSM captures nonlinear and local features of the DELOA, demonstrating high accuracy and robustness in predicting linear oscillating actuator performance.Improved crowding degree model to refine the ranking mechanism of the NSGA-II algorithm significantly increases the density of solutions in the objective space without additional computational cost, thus boosting optimization efficiency.The hybrid surrogate model proposed in this paper adopts four sub-models, namely RSM, KM, BPNN and SVR, each of which has its own advantages in the field of electromagnetic devices. It exhibits distinct advantages such as high fitting accuracy and strong robustness in data-driven modeling, performance prediction and multi-objective optimization. However, this also limits its specific application scope to motors and other electromagnetic devices. Therefore, in future work, when the optimization method is further applied to other high-dimensional, highly nonlinear and multi-output systems, more advantageous sub-models in the corresponding fields can be selected to construct the hybrid surrogate model. Due to the short lengths of both the stator and the mover in the LOA, end effects are pronounced; therefore, special attention must be given to thrust performance during the design process. By optimizing key dimensions, such as the stator’s tooth crown and end tooth width, as well as the length of the mover’s magnets, end effects can be effectively suppressed. This minimizes the significant fluctuations in thrust, both under no-load and load conditions, when the mover approaches the stator’s edge. This improvement enhances the linearity of the EM force coefficient, thereby supporting more precise control of linear reciprocating motion. Due to the short lengths of both the stator and the mover in the LOA, end effects are pronounced; therefore, special attention must be given to thrust performance during the design process. By optimizing key dimensions, such as the stator’s tooth crown and end tooth width, as well as the length of the mover’s magnets, end effects can be effectively suppressed. This minimizes the significant fluctuations in thrust, both under no-load and load conditions, when the mover approaches the stator’s edge. This improvement enhances the linearity of the EM force coefficient, thereby supporting more precise control of linear reciprocating motion. By integrating various data-driven sub-models with a coefficient-of-determination-based weighting mechanism, a high-accuracy HSM for DELOA is developed. The HSM captures nonlinear and local features of the DELOA, demonstrating high accuracy and robustness in predicting linear oscillating actuator performance. By integrating various data-driven sub-models with a coefficient-of-determination-based weighting mechanism, a high-accuracy HSM for DELOA is developed. The HSM captures nonlinear and local features of the DELOA, demonstrating high accuracy and robustness in predicting linear oscillating actuator performance. Improved crowding degree model to refine the ranking mechanism of the NSGA-II algorithm significantly increases the density of solutions in the objective space without additional computational cost, thus boosting optimization efficiency. Improved crowding degree model to refine the ranking mechanism of the NSGA-II algorithm significantly increases the density of solutions in the objective space without additional computational cost, thus boosting optimization efficiency. The hybrid surrogate model proposed in this paper adopts four sub-models, namely RSM, KM, BPNN and SVR, each of which has its own advantages in the field of electromagnetic devices. It exhibits distinct advantages such as high fitting accuracy and strong robustness in data-driven modeling, performance prediction and multi-objective optimization. However, this also limits its specific application scope to motors and other electromagnetic devices. Therefore, in future work, when the optimization method is further applied to other high-dimensional, highly nonlinear and multi-output systems, more advantageous sub-models in the corresponding fields can be selected to construct the hybrid surrogate model. The hybrid surrogate model proposed in this paper adopts four sub-models, namely RSM, KM, BPNN and SVR, each of which has its own advantages in the field of electromagnetic devices. It exhibits distinct advantages such as high fitting accuracy and strong robustness in data-driven modeling, performance prediction and multi-objective optimization. However, this also limits its specific application scope to motors and other electromagnetic devices. Therefore, in future work, when the optimization method is further applied to other high-dimensional, highly nonlinear and multi-output systems, more advantageous sub-models in the corresponding fields can be selected to construct the hybrid surrogate model.