{"title":"基于集成深度学习和强化学习的人工智能驱动编织复合材料优化","authors":"Mao-Ken Hsu, Bo-Yu Huang, Chi-Hua Yu","doi":"10.1016/j.matdes.2025.114798","DOIUrl":null,"url":null,"abstract":"<div><div>Woven carbon fiber composites are increasingly adopted in advanced structural applications due to their exceptional strength-to-weight ratio and tunable design features. However, high-fidelity simulations of their complex woven architecture are computationally intensive. This study presents a hybrid deep learning framework that combines a dual-input Convolutional Neural Network (CNN) for mechanical property prediction with a Deep Q-Network (DQN) for reinforcement learning-based optimization. The CNN achieves R<sup>2</sup> values above 0.96 for elastic deformation, plastic deformation, and strain energy density prediction. Using the DQN, the optimized design achieves a 2.37-fold improvement in strain energy density, increasing from 3590.78 J/m<sup>3</sup> to 8527.85 J/m<sup>3</sup>. Furthermore, by replacing the original woven geometry with a reduced model using stress–strain behavior, simulation time is reduced from 534 min to 2 min, a 267-fold speedup. This approach significantly enhances efficiency in composite design and optimization workflows, enabling rapid exploration of high-performance configurations.</div></div>","PeriodicalId":383,"journal":{"name":"Materials & Design","volume":"259 ","pages":"Article 114798"},"PeriodicalIF":7.9000,"publicationDate":"2025-09-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Ai-driven optimization of woven composite via integrated deep and reinforcement learning\",\"authors\":\"Mao-Ken Hsu, Bo-Yu Huang, Chi-Hua Yu\",\"doi\":\"10.1016/j.matdes.2025.114798\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<div><div>Woven carbon fiber composites are increasingly adopted in advanced structural applications due to their exceptional strength-to-weight ratio and tunable design features. However, high-fidelity simulations of their complex woven architecture are computationally intensive. This study presents a hybrid deep learning framework that combines a dual-input Convolutional Neural Network (CNN) for mechanical property prediction with a Deep Q-Network (DQN) for reinforcement learning-based optimization. The CNN achieves R<sup>2</sup> values above 0.96 for elastic deformation, plastic deformation, and strain energy density prediction. Using the DQN, the optimized design achieves a 2.37-fold improvement in strain energy density, increasing from 3590.78 J/m<sup>3</sup> to 8527.85 J/m<sup>3</sup>. Furthermore, by replacing the original woven geometry with a reduced model using stress–strain behavior, simulation time is reduced from 534 min to 2 min, a 267-fold speedup. This approach significantly enhances efficiency in composite design and optimization workflows, enabling rapid exploration of high-performance configurations.</div></div>\",\"PeriodicalId\":383,\"journal\":{\"name\":\"Materials & Design\",\"volume\":\"259 \",\"pages\":\"Article 114798\"},\"PeriodicalIF\":7.9000,\"publicationDate\":\"2025-09-19\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Materials & Design\",\"FirstCategoryId\":\"88\",\"ListUrlMain\":\"https://www.sciencedirect.com/science/article/pii/S0264127525012183\",\"RegionNum\":2,\"RegionCategory\":\"材料科学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q1\",\"JCRName\":\"MATERIALS SCIENCE, MULTIDISCIPLINARY\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Materials & Design","FirstCategoryId":"88","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S0264127525012183","RegionNum":2,"RegionCategory":"材料科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"MATERIALS SCIENCE, MULTIDISCIPLINARY","Score":null,"Total":0}
Ai-driven optimization of woven composite via integrated deep and reinforcement learning
Woven carbon fiber composites are increasingly adopted in advanced structural applications due to their exceptional strength-to-weight ratio and tunable design features. However, high-fidelity simulations of their complex woven architecture are computationally intensive. This study presents a hybrid deep learning framework that combines a dual-input Convolutional Neural Network (CNN) for mechanical property prediction with a Deep Q-Network (DQN) for reinforcement learning-based optimization. The CNN achieves R2 values above 0.96 for elastic deformation, plastic deformation, and strain energy density prediction. Using the DQN, the optimized design achieves a 2.37-fold improvement in strain energy density, increasing from 3590.78 J/m3 to 8527.85 J/m3. Furthermore, by replacing the original woven geometry with a reduced model using stress–strain behavior, simulation time is reduced from 534 min to 2 min, a 267-fold speedup. This approach significantly enhances efficiency in composite design and optimization workflows, enabling rapid exploration of high-performance configurations.
期刊介绍:
Materials and Design is a multi-disciplinary journal that publishes original research reports, review articles, and express communications. The journal focuses on studying the structure and properties of inorganic and organic materials, advancements in synthesis, processing, characterization, and testing, the design of materials and engineering systems, and their applications in technology. It aims to bring together various aspects of materials science, engineering, physics, and chemistry.
The journal explores themes ranging from materials to design and aims to reveal the connections between natural and artificial materials, as well as experiment and modeling. Manuscripts submitted to Materials and Design should contain elements of discovery and surprise, as they often contribute new insights into the architecture and function of matter.