KFKai Fu

Kai Fu

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Kai Fu has a total of 22 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Surrogate-Assisted Level Set Topology Optimization for Pixelated Antenna Design

    by Haohua Chen, Kai Fu, Xiwen Cai, Lei Yue, Liang Gao

    This letter proposes a novel antenna topology optimization framework consisting of two components. First, a Kriging-based level set method (KLSM) is employed to formulate the antenna optimization problem. The KLSM defines level set values at a small number of knot points and uses Kriging interpolation to generate the geometric representation of the antenna topology. The KLSM reduces the design space dimensionality and transforms the complex discrete pixelated antenna optimization problem into a low-dimensional continuous optimization problem via variable substitution. Second, a dual surrogate-assisted genetic algorithm (DSGA) is proposed to solve the KLSM-based antenna optimization problem. The DSGA employs two surrogate-assisted strategies to improve the traditional genetic algorithm. A local search is performed around the current best individual by using the radial basis function (RBF). A global search is performed by using expected improvement of high-dimensional Kriging to prescreen potential candidate individuals. The proposed surrogate-assisted topology optimization method is validated through the design of a pixelated monopole antenna. Results show that the proposed method outperforms the state-of-the-art methods in both convergence speed and optimization accuracy.


  • NeuroStorm: An Adaptive Neural-Network-Driven Algorithm for Mixed-Variable Antenna Design

    by Kai Fu, Kwok Wa Leung

    Many practical antennas contain continuous, binary, and categorical geometrical variables. Antenna designs involving such mixed-type variables fall under the category of mixed-variable antenna optimization problems (MVAOPs). Most existing studies have focused on antenna designs with continuous variables only and therefore the MVAOP is a relatively unexplored area with several challenges to be addressed. This paper presents an adaptive neural-network-driven algorithm, named NeuroStorm, as a solution for MVAOPs. NeuroStorm integrates three innovative modules: (1) an adaptive neural network with dynamically adjustable hidden layer nodes, (2) a complementary objective estimation method for reducing simulation costs, and (3) an enhanced brainstorm optimization capable of simultaneously handling continuous, binary, and categorical variables. A stacked dense dielectric patch antenna is used to illustrate the step-by-step workflow of NeuroStorm. Results show that NeuroStorm achieves faster convergence and yields a better dielectric patch antenna design compared with three existing algorithms. The generality of NeuroStorm is validated through four benchmark functions and two additional antenna designs, demonstrating its effectiveness and adaptability for solving diverse MVAOPs.


  • Machine-Learning-Assisted Swarm Intelligence Algorithm for Antenna Optimization With Mixed Continuous and Binary Variables

    by Kai Fu, Kwok Wa Leung

    Many existing surrogate-assisted optimization algorithms are limited to designing antennas with continuous variables only. However, numerous challenges emerge when tackling antenna optimization problems that involve both continuous and binary design variables. This paper proposes an efficient surrogate-assisted mixed continuous/binary particle swarm optimization (SAMPSO) algorithm to address these mixed-variable antenna optimization problems. The SAMPSO tightly integrates machine learning (ML) models with PSO in two key aspects: an ML-guided swarm updating method and an ML-assisted prescreening strategy. In addition, a novel local ML model training method is developed to reduce the algorithm time complexity. To verify its effectiveness, the SAMPSO is compared with two existing algorithms in solving benchmark functions and designing antennas. Results demonstrate that SAMPSO can achieve design objectives with a faster convergence speed.


  • Surrogate-Assisted PSO for Patch Antenna Designs

    by Kai Fu, Kwok Wa Leung

    In this study, we employed a surrogate-assisted particle swarm optimization (ESPSO) method to efficiently design wideband patch antennas. The ESPSO algorithm significantly reduced the computational burden typically associated with antenna design, requiring approximately 2000 EM simulations to meet the design specifications. The optimized antenna achieved a 42.95% bandwidth, covering the frequency range from 1.81 GHz to 2.8 GHz, with broadside radiation patterns and a maximum realized gain of 9.74 dBi. The effectiveness of ESPSO was demostrated in achieving high-performance wideband antenna designs within a limited computational budget.

  • An Efficient Surrogate Assisted Particle Swarm Optimization for Antenna Synthesis

    by Kai Fu, Xiwen Cai, Bo Yuan, Yang Yang, Xin Yao

    By virtue of the prediction abilities of machine learning (ML) methods, the ML-assisted evolutionary algorithm has been treated as an efficient solution for antenna design automation. This article presents an efficient ML-based surrogate-assisted particle swarm optimization (SAPSO). The proposed algorithm closely combines the particle swarm optimization (PSO) with two ML-based approximation models. Then, a novel mixed prescreening ( mix P) strategy is proposed to pick out promising individuals for full-wave electromagnetic (EM) simulations. As the optimization procedure progresses, the ML models are dynamically updated once new training data are obtained. Finally, the proposed algorithm is verified by three real-world antenna examples. The results show that the proposed SAPSO- mix P can find favorable results with a much smaller number of EM simulations than other methods.


  • Modeling and Characterization of Differential Multibit Carbon-Nanotube Through-Silicon Vias

    by Qing-Hao Hu, Wen-Sheng Zhao, Kai Fu, Gaofeng Wang

    By virtue of multi-bit through-silicon via (TSV), which consists of carbon-nanotube (CNT) filled TSV with separate metal pads, a design of compact differential multi-bit CNT TSVs (DMC-TSVs) is proposed and studied. For modeling such DMCTSVs, an equivalent circuit model is introduced, in which the partial-element equivalent circuit (PEEC) method is adopted to extract the impedance. Moreover, the circuit model is simplified as an odd-mode half-circuit and is used to design passive RL equalizer.


  • A Passive Equalizer Design for Shielded Differential Through-Silicon Vias in 3-D IC

    by Kai Fu, Wen-Sheng Zhao, Gaofeng Wang, Madhavan Swaminathan

    This letter develops a simplified circuit model for the shielded differential through-silicon vias (SD-TSVs). The model is consisted of four frequency-independent elements and can accurately predict the insertion loss of the SD-TSVs up to 20 GHz. Based on the model, a parallel resistor–capacitor equalizer is proposed. The proposed passive equalizer achieves wide-band equalization and perfect compensation for high-speed differential signaling.


  • Changes in femoral trochlear morphology following surgical correction of recurrent patellar dislocation associated with trochlear dysplasia in children

    by K. Fu, G. Duan, C. Liu, J. Niu, F. Wang

    Aims: The aim of this study was to investigate the changes in femoral trochlear morphology following surgical correction of recurrent patellar dislocation associated with trochlear dysplasia in children.Patients and Methods: A total of 23 patients with a mean age of 9.6 years (7 to 11) were included All had bilateral recurrent patellar dislocation associated with femoral trochlear dysplasia. The knee with traumatic dislocation at the time of presentation or that had dislocated most frequently was treated with medial patellar retinacular plasty (Group S). The contralateral knee served as a control and was treated conservatively (Group C). All patients were treated between October 2008 and August 2013. The mean follow-up was 48.7 months (43 to 56). Axial CT scans were undertaken in all patients to assess the trochlear morphological characteristics on a particular axial image which was established at the point with the greatest epicondylar width based on measurements preoperatively and at the final follow-up.Results: Preoperatively, there were no statistically significant differences between the trochlear morphology in the two groups (sulcus angle, p 0.852; trochlear groove depth, p 0.885; lateral trochlear inclination, p 0.676; lateral-to-medial facet ratio, p 0.468; lateral condylar height, p 0.899; medial condylar height, p 0.816). Many radiological parameters of trochlear morphology were significantly different between the two groups at the final follow-up, including well-known parameters, such as the mean sulcus angle (Group S, 146.27° (sd 7.18); Group C, 160.61° (sd 9.29); p < 0.001), the mean trochlear groove depth (Group S, 6.25 mm (sd 0.41); Group C, 3.48 mm (sd 0.65); p < 0.001) and the mean lateral trochlear inclination (Group S, 20.99° (sd 3.87); Group C, 12.18° (sd 1.85); p < 0.001). Lesser known parameters such as the ratio of the lateral to medial trochlear length (Group S, 1.46 (sd 0.19); Group C, 2.14 (sd 0.42); p < 0.001), which is a measurement of facet asymmetry, and the lateral and medial condylar height were also significantly different between the two groups (p < 0.001).Conclusion: The femoral trochlear morphology can be improved by early (before epiphyseal closure) surgical correction in children with recurrent patellar dislocation associated with femoral trochlear dysplasia. Cite this article: Bone Joint J 2018;100-B:811–21.


  • Tests of new type of precast sandwich panel with FRP connectors

    by Weichen Xue, Ya Li, Jialin Yang, Kai Fu, Zhijie Li