Liu Wenhang, Zhang Youwei, Lyu Yongtao, Bosiakov Sergei, Liu Yadong
Department of Engineering Mechanics, Dalian University of Technology, Dalian, China.
DUT-BSU Joint Institute, Dalian University of Technology, Dalian, China.
Front Bioeng Biotechnol. 2023 Sep 7;11:1241151. doi: 10.3389/fbioe.2023.1241151. eCollection 2023.
Triply periodic minimal surface (TPMS) is widely used in the design of bone scaffolds due to its structural advantages. However, the current approach to designing bone scaffolds using TPMS structures is limited to a forward process from microstructure to mechanical properties. Developing an inverse bone scaffold design method based on the mechanical properties of bone structures is crucial. Using the machine learning and genetic algorithm, a new inverse design model was proposed in this research. The anisotropy of bone was matched by changing the number of cells in different directions. The finite element (FE) method was used to calculate the TPMS configuration and generate a back propagation neural network (BPNN) data set. Neural networks were used to establish the relationship between microstructural parameters and the elastic matrix of bone. This relationship was then used with regenerative genetic algorithm (RGA) in inverse design. The accuracy of the BPNN-RGA model was confirmed by comparing the elasticity matrix of the inverse-designed structure with that of the actual bone. The results indicated that the average error was below 3.00% for three mechanical performance parameters as design targets, and approximately 5.00% for six design targets. The present study demonstrated the potential of combining machine learning with traditional optimization method to inversely design anisotropic TPMS bone scaffolds with target mechanical properties. The BPNN-RGA model achieves higher design efficiency, compared to traditional optimization methods. The entire design process is easily controlled.
三重周期极小曲面（TPMS）因其结构优势而被广泛应用于骨支架设计。然而，目前利用TPMS结构设计骨支架的方法仅限于从微观结构到力学性能的正向过程。开发一种基于骨结构力学性能的反向骨支架设计方法至关重要。本研究利用机器学习和遗传算法，提出了一种新的反向设计模型。通过改变不同方向的单元数量来匹配骨的各向异性。采用有限元（FE）方法计算TPMS构型并生成反向传播神经网络（BPNN）数据集。利用神经网络建立微观结构参数与骨弹性矩阵之间的关系。然后在反向设计中，将这种关系与再生遗传算法（RGA）结合使用。通过比较反向设计结构的弹性矩阵与实际骨的弹性矩阵，证实了BPNN-RGA模型的准确性。结果表明，以三个力学性能参数为设计目标时，平均误差低于3.00%，以六个设计目标时，平均误差约为5.00%。本研究证明了将机器学习与传统优化方法相结合来反向设计具有目标力学性能的各向异性TPMS骨支架的潜力。与传统优化方法相比，BPNN-RGA模型具有更高的设计效率。整个设计过程易于控制。
