生物支架材料大略可分為天然材料跟人工合成兩大類，目前最為廣泛使用的天然材料有膠原蛋白、明膠及殼聚醣等。由於上述材料係從動、植物取得，每批量配置出的生物墨水總有些許差距。生物列印之製造參數大多彼此相依，因此很難藉由理論推導出有效的預測公式。往往只能使用試誤法來調整製造參數來達到預設的尺寸目標。多層感知器為深度學習的一支，其可透過大量過往實驗數據訓練出一個有效預測模型，在新的數據輸入此一模型後進而預測結果，幫助改進生物支架的外形品質。 本研究之目的是以深度學習的方式改進生物支架的外形品質。本研究運用多層感知器進行列印參數分析，研究不同的材料性質與製造參數對於支架品質的關聯性。製造參數包含了動力黏度、擠出壓力、噴頭溫度、噴頭移動速度、沉積平台溫度等等，根據這些模型可以看出動力黏度為列印過程中最相關的參數。實際製作75個支架做為訓練集，8個支架做為測試集，將動力黏度、噴頭溫度、噴頭移動速度、沉積平台溫度、環境壓力及室內露點溫度六種參數做為模型預測之重要參數，模型使用兩層隱藏層，所預測之評分與平均線徑的均方誤差最小為0.0246。 ;The materials of bio-scaffolds can be roughly divided into two categories: natural and synthetic materials. The most widely used natural materials are collagen, gelatin and chitosan. As these natural materials are obtained from flora and fauna, there is always a slight difference between the bio-inks in each batch. Most of the fabrication parameters have dependence with each other, it is difficult to derive effective predictive formulas for manufacturing scaffolds. Often can only use the trial and error method to adjust the fabrication parameters for achieving the preset target. Multilayer perceptron is one kind of Deep learning, and it can train a model through a large amount of historical data, and predictable results obtain after inputting new data into the model. The capability can help to improve the appearance quality of the bio-scaffold. The purpose of this study is to improve the appearance quality of the bio-scaffold by utilizing deep learning. The multilayer perceptron is used to investigate the material properties and the fabrication parameters in order to achieve optimal quality of the scaffold. The fabrication parameters are based on kinematic viscosity, print pressure, nozzle temperature, nozzle speed, platform temperature, and so on. According to these models, it can be seen that the kinematic viscosity is the most relevant parameter in the printing process. 75 scaffolds were actually used as the training dataset and 8 scaffolds were used as the test set. Six parameters including dynamic viscosity, nozzle temperature, nozzle speed, platform temperature, print pressure and dew point temperature are used as important parameters for the model prediction. The model uses two hidden layers, and The minimum mean square error of predicted score and average line width is 0.0246.
