以基於Transformer的對抗生成網路生成三維仿生微結構

Abstract

生物材料因其複雜的材料微結構而具有獨特的特性，這些特性造就其特別適用於創造具有多樣化功能的高性能材料。本研究聚焦於透過具有自注意機制的深度生成網絡，即AE-transformer GAN，生成具有預定特性的仿生微結構。模型的初始訓練是基於兩種數據類型，分別為等向性多孔狀微結構和異向性生物微觀結構。並通過基於圖像的分析和利用3D列印試體的物理實驗驗證並且評估了模型的有效性。下一步，我們將模型拓展至有條件式的生成模型，以便更精確地控制生成過程。此外，透過將局部熵作為結構指標加入條件式模型的生成過程，建立具有訂製特性的混合微結構之生成策略。本研究提出了一種高效的生成方式，用於學習與生成仿生微結構，並利用條件式生成模型，奠定新的生成方式能夠探索從未見過並具有目標特性的混種微結構。

Biomaterials are distinguished by their unique attributes, which are largely attributable to their intricate microstructures. These properties make them particularly attractive for the creation of high-performance materials with diverse functionalities. This research focuses on the generation of bioinspired microstructures with predetermined characteristics. This is achieved through a deep generative network equipped with a self-attention mechanism, specifically the AE-transformer GAN. The initial training of the model was based on two types of data: isotropic porous structures and anisotropic biological microstructures. The efficacy of the model was then evaluated using image-based analysis and physical examination of 3D-printed specimens. Subsequently, the model was augmented to become a conditional generative model, providing more precise control over the generation process. Moreover, the incorporation of local entropy as a structural index facilitated the strategic generation of hybrid microstructures with bespoke properties. In summary, this thesis introduces a highly efficient generative methodology for learning and generating bioinspired microstructures. Notably, it paves the way for the design of previously unexplored hybrid microstructures with targeted properties, thanks to the conditional generative models.