物理化学学报

上一篇    下一篇

人工智能驱动的多功能聚酰亚胺设计:数据、生成与合成范式革命

张泽鹏, 吴奇坤, 汪昕隆, 廖彦淞, 林钰洲, 李紫芹, 吴绍平, 张权聪, 肖宗源, 高铭滨, 苏文湫, 洪文晶   

  1. 固体表面物理化学国家重点实验室, 化学化工学院&人工智能研究院&福建省能源材料科学与技术创新实验室(IKKEM), 厦门大学, 福建 厦门 361005
  • 收稿日期:2025-11-12 修回日期:2026-01-14
  • 通讯作者: 高铭滨, 苏文湫, 洪文晶 E-mail:mbgao@xmu.edu.cn;suwq@xmu.edu.cn;whong@xmu.edu.cn
  • 基金资助:
    洪文晶感谢中国国家自然科学基金(22250003和22325303)的资助支持。高铭滨感谢中国厦门市自然科学基金(20241134)、中国国家自然科学基金(22208337)以及中央高校基本科研业务费专项资金(20720240060)的资助支持。苏文湫感谢新一代人工智能国家科技重大专项(2025ZD0121903)的资助支持。李紫芹感谢新一代人工智能国家科技重大专项(2025ZD0121800)的资助支持。吴绍平感谢中国国家自然科学基金(22503038)的资助支持。人工智能驱动的实验、模拟和模型训练均在中国科学院机器人AI-Scientist平台上完成。

AI-Driven design of multifunctional polyimides: revolutionizing the paradigm from data to synthesis

Zepeng Zhang, Qikun Wu, Xinlong Wang, Yansong Liao, Yuzhou Lin, Ziqin Li, Shaoping Wu, Quancong Zhang, Zongyuan Xiao, Mingbin Gao, Wenqiu Su, Wenjing Hong   

  1. State Key Laboratory of Physical Chemistry of Solid Surfaces, College of Chemistry and Chemical Engineering & Institute of Artificial Intelligence & Innovation Laboratory for Sciences and Technologies of Energy Materials of Fujian Province (IKKEM), Xiamen University, Xiamen 361005, Fujian Province, China
  • Received:2025-11-12 Revised:2026-01-14
  • Contact: Mingbin Gao, Wenqiu Su, Wenjing Hong E-mail:mbgao@xmu.edu.cn;suwq@xmu.edu.cn;whong@xmu.edu.cn

摘要: 人工智能(Artificial intelligence, AI)技术正在重塑高分子材料研发范式,尤其在聚酰亚胺(Polyimide, PI)等高性能聚合物领域展现出巨大潜力。本文系统梳理了AI在高分子科学中的三大核心应用:数据驱动的分子生成、面向性能的PI材料设计、自动化实验平台构建。通过融合实验数据库、文献挖掘与高通量计算,机器学习模型(如图神经网络与Transformer)已实现PI热学、电学、力学、光学及气体分离等多性质的高精度预测。生成式AI (如VAE与GPT)进一步突破传统试错法,逆向设计满足极端环境需求的虚拟PI结构,极大拓展了材料探索边界。而机器人平台(如Polybot与PANDA)结合贝叶斯优化与主动学习,构建起“设计-合成-表征-反馈”的闭环系统,将研发周期从数年缩短至数月。尽管面临数据标准化、模型可解释性、实验与模拟鸿沟等挑战,AI仍推动PI研发进入多目标协同优化与跨尺度整合的新阶段,未来有望在5G/6G、航空航天等领域实现“订制化”高分子材料创新。

关键词: AI驱动设计, 数据驱动分子生成, 生成式设计, 多性质优化, 自驱动实验室(SDL)

Abstract: Artificial intelligence (AI) technologies are reshaping the paradigm of polymer material research and development, especially showing great potential in the field of high-performance polymers such as polyimide (PI). This article systematically reviews the three core applications of AI in polymer science: data-driven molecular generation, performance-oriented PI material design, and the construction of automated experimental platforms. By integrating experimental databases, literature mining, and high-throughput computing, machine learning models (such as graph neural networks and Transformers) have achieved high-precision predictions of multiple properties of PI, including thermal, electrical, mechanical, optical, and gas separation properties. Generative AI (such as VAE and GPT) has further broken through the traditional trial-and-error method, inversely designing virtual PI structures that meet extreme environmental requirements, significantly expanding the boundaries of material exploration. Meanwhile, robotic platforms (such as Polybot and PANDA), combined with Bayesian optimization and active learning, have built a closed-loop system of “design-synthesis-characterization-feedback,” reducing the R&D cycle from years to months. Despite challenges such as data standardization, model interpretability, and the gap between experiments and simulations, AI is still driving PI research and development into a new stage of multi-objective collaborative optimization and cross-scale integration. In the future, it is expected to achieve “customized” polymer material innovation in fields such as 5G/6G and aerospace.

Key words: AI-Driven design, Data-driven molecular generation, Generative design, Multi-property optimization, Self-Driving Laboratory (SDL)