機器學習驅動強度與韌性最佳化仿生珍珠母微結構設計

Machine Learning Enabled Strength and Toughness Optimization for Nacre-inspired Microstructural Design

蔡松霖(Sung-Lin Tsai)

指導教授 ： 陳俊杉

國立臺灣大學/工學院/土木工程學研究所/碩士(2020年)

https://doi.org/10.6342/NTU202001570

摘要

本研究針對珍珠母啟發微結構複合材料建立了一套以機器學習驅動的材料性質最佳化設計系統，希望能在強度及韌性之間的衝突以及廣大設計空間中找到符合目標材料性質的微結構設計。材料的設計與發展從古至今一直以來都與我們日常生活息息相關。近年來，隨著電腦模擬、3D列印以及機器學習等先進技術不斷蓬勃發展，設計材料所能拓展的層面也越來越廣。因此，如何在廣泛的可能性中找尋新穎、符合需求的材料設計，是本研究主要的研究動機。

本研究首先分別訓練了卷積神經網絡（CNN）預測模型，預測珍珠母啟發設計空間中微結構的最大強度以及韌性(吸收能量)，藉此建立微結構以及材料性質之間的關聯性。強度預測模型準確度高達99％，而韌性預測模型亦有80％。因此，將預測模型作為快速篩選的工具，結合基因演算法（GA），探索以珍珠母啟發，可能組合高達10^10的設計空間。為了找尋既堅固又堅韌的微結構，本篇研究中針對不同的材料性質進行單目標及多目標的最佳化。在額外條件硬材體積分率0.95的限制下，透過結合機器學習與多目標最佳化演算法，設計出的微結構強度從129.7增加到150.0 MPa，韌性從0.095增加到0.123 mJ/mm^3。最後，本研究亦訓練了變分自動編碼器（VAE）模型，以建立珍珠質啟發設計空間和潛在空間之間的轉換關係，進而擴展設計更多可能性。

關鍵字

機器學習 ； 多目標最佳化 ； 微結構設計 ； 拓樸最佳化 ； 仿生珍珠母結構 ； 基因演算法 ； 卷積神經網路

並列摘要

In this research, a machine learning-enabled optimized design system for nacre-inspired microstructural composites is developed to tackle the conflict between strength and toughness and the challenge of design space exploration. The development of materials is always an essential issue that deeply affects our daily life. Recently, owing to the continuous advances in computational simulation, 3-D printing experiments, and machine learning techniques, the process of material design has been elevated to a new level.

This study first creates linkages between microstructure representations and material properties by respectively training convolutional neural networks (CNN) predictive models for ultimate strength and toughness (absorbed energy) with a nacre-inspired dataset of combinations of soft and stiff materials up to 10^10 order derived from computational simulations. An impressive 99% accuracy for ultimate strength prediction and 80% accuracy for toughness prediction are achieved. Accordingly, served as high-throughput mapping tools, predictive models were implemented in conjunction with genetic algorithm (GA) to explore nacre-inspired design space for obtaining microstructural designs with target properties in the image size of 512 by 512. Single objective and multi-objective optimization tasks were conducted to find microstructures that are both strong and tough. For volume fraction ratio of hard material 0.95, strength increases from 129.7 to 150.0 MPa, and toughness increases from 0.095 to 0.123 mJ/mm^3, through machine-learning enabled multi-objective microstructural design. Last but not least, a variational autoencoder (VAE) model is trained for bridging between nacre-inspired design space and a latent space, which broadens the design.

並列關鍵字

Machine Learning ； Multi-objective Optimization ； Microstructural Design ； Topology Optimization ； Nacre-inspired Composite ； Genetic Algorithm ； Convolutional Neural Networks

參考文獻

Abueidda, D. W., Almasri, M., Ammourah, R., Ravaioli, U., Jasiuk, I. M., Sobh, N. A. (2019). Prediction and optimization of mechanical properties of composites using convolutional neural networks. Composite Structures, 227.

Barthelat, F., Mirkhalaf, M. (2013). The quest for stiff, strong and tough hybrid materials: An exhaustive exploration. Journal of the Royal Society Interface, 10(89).

Bekah, S., Rabiei, R., Barthelat, F. (2012). The Micromechanics of Biological and Biomimetic Staggered Composites. Journal of Bionic Engineering, 9(4), 446–456.

Benyus, J. M. (2002). Biomimicry: innovation inspired by nature. New York: Perennial. Carpinteri, A., Accornero, F. (2018). Multiple snap-back instabilities in progressive

microcracking coalescence. Engineering Fracture Mechanics, 187, 272–281.