- 學位論文
以深度學習與4D列印反向設計3D複雜曲面──以人臉面具為例
Inverse design of Complex 3D Gridshell by Deep Learning and 4D Printing: A Case Study of Face Mask Design
摘要
4D列印奠基在3D列印技術之上，利用形狀記憶效應使物件經過外界如熱或光等刺激後能再次變形，其優勢是在印製空心或懸空的網格結構時能省下大量的支撐材，並進而大幅加速製造的速度。過去雖有研究使用形狀記憶聚合物組成的平面網格透過4D列印來進行立體網格的製作，但由於變形機制的高度非線性與鄰近網格的相互牽連，反向設計的過程非常困難。因此，本研究探討了形狀記憶聚合物作為平面網格材料的設計空間，希望能以深度學習自動化反向設計的過程。其中本研究利用熔融堆疊式的3D列印機列印SMP55時儲存的預應力作為4D列印的機制，結合PLA產生的遇熱會彎曲的雙層結構，產生共四種的單元網格配置的平面網格設計空間。本研究先以人為反向設計藉由嘗試錯誤的方法，搭配有限元素法與繪圖軟體反向設計三個日本能面，驗證了此設計空間的多樣性。接著深度學習反向設計的部分，本研究將人臉面具的平面網格設計以多項式的參數來生成大量的隨機人臉面具，並搭配有限元素模擬產生對應的變形形狀作為深度學習模型訓練的資料集。模型架構上本研究選擇通常用於影像分割(Image segmentation)任務的全卷積網路(Fully convolutional network)模型進行反向設計，模型會根據目標形狀的深度照片來生成平面網格設計。在測試資料集中全卷積網路生成的人臉面具能夠有超過0.95的相素準確度與0.9的平均並交比，代表網格設計變形形狀的深度照片也有約0.9的結構相似性與7.5的均方誤差。雖然模型在資料集外如日本能面的反向設計結果不是很理想，卻已足夠證明此一方法的可行性。本研究也以日本能面為例以泡熱水實驗與石膏鋪膜的方法改善了人臉面具的製程，其結果不僅能夠驗證有限元素的模擬，也能製造出與能面相似的面具。
並列摘要
4D printing is a technology that is built upon 3D printing. By utilizing shape memory effect, 3D-printed objects have the ability to deform again due to certain external stimuli such as heat and light. It can greatly reduce the waste of supporting material and printing time for certain structures such as hanging or hollow structure. However, it is very difficult to inverse design this process because of the nonlinearity of the morphing mechanism and the entanglement between different parts. In this study, we use deep learning techniques to overcome this difficulty and automate the inverse design process. Specifically, this research studies a 2D grid design space that deforms to 3D gridshell through 4D printing process. The 2D grid is composed of rectangular arranged double-layered segments. Each layer is made of shape memory polymer (SMP55) or PLA, resulting in four combinations for each segment. The size and material combination of each 2D grid is specified to control both global and local curvatures of the deformed gridshell which can achieve a variety of complex structures. Three traditional Japanese “Noh masks” are chosen as the target shapes because Noh masks are an ideal model system as each mask has unique aesthetic features. We use parametric polynomial functions describe face features and generate random mask designs. Combining with the deformed shapes simulated by FEM software, we produced a dataset which contained 60k data. We use fully convolutional network (FCN) which typically used in image segmentation task to inverse design 2D grids based on the depth images of the desire shapes. The trained FCN model can predict 2D grid designs with over 0.95 pixel accuracy and 0.9 mean IOU. Also, by the calculation of structural similarity, the average similarity of 3D gridshells deformed from FCN-output designs and target 3D gridshells is 0.9. Although the model is limited to the distribution of the training dataset and performs poorly on Noh masks, it is still a successful proof of concept that deep learning can be utilized in the inverse design problem of 4D printing 2D grid design.