人工智能辅助空天新材料设计研究进展

基金项目

国家重点研发计划(2022YFB3707800);上海市教育委员会“人工智能促进科研范式改革、赋能学科跃升计划”专项;国家自然科学基金(12572125,12421002)。

中图分类号:

TP18TB3

文献标识码:

A

编辑

责编 :逸飞

引文格式

孙升, 陈祎远, 尚卿. 人工智能辅助空天新材料设计研究进展[J]. 航空制造技术, 2025, 68(18): 26–44.

AI-Assisted Design of Aerospace Advanced Materials: Recent Advances and Perspectives

Citations

SUN Sheng, CHEN Yiyuan, SHANG Qing. AI-assisted design of aerospace advanced materials: recent advances and perspectives[J]. Aeronautical Manufacturing Technology, 2025, 68(18): 26–44.

航空制造技术    第68卷    第18期    26-44
Aeronautical Manufacturing Techinology    Vol.68    No.18 : 26-44
DOI: 10.16080/j.issn1671-833x.2025.18.026
封面文章(COVER STORY)

人工智能辅助空天新材料设计研究进展

  • 孙升 1,2
  • 陈祎远 1
  • 尚卿 1
1.上海大学材料基因组工程研究院上海 200444
2.上海市力学信息学前沿科学研究基地上海 200072

基金项目

国家重点研发计划(2022YFB3707800);上海市教育委员会“人工智能促进科研范式改革、赋能学科跃升计划”专项;国家自然科学基金(12572125,12421002)。

中图分类号:

TP18TB3

文献标识码:

A

引文格式

孙升, 陈祎远, 尚卿. 人工智能辅助空天新材料设计研究进展[J]. 航空制造技术, 2025, 68(18): 26–44.

摘要

极端的工作服役环境,是新一代航空航天材料面临的巨大挑战。传统的材料设计方法面临效率低、成本高、研发周期长等挑战,已严重制约航空航天材料的发展。空天新材料的研发亟需创新且高效精准的材料研发范式。人工智能(Artificial intelligence,AI)技术,尤其是机器学习和深度学习的迅猛进步,为航空航天材料研发提供了强有力的工具,可显著提升新材料设计效率和性能预测的准确性。本文系统综述了AI在航空航天材料领域的研究进展,首先介绍了AI辅助的多尺度计算模拟与智能化试验,接着系统性地介绍了代理模型加速的材料优化设计方法和以大模型为核心的新型材料设计流程,并详细探讨了AI技术在合金材料、复合材料及超材料研发中的具体应用案例。最后,总结了AI辅助航空航天材料设计的优势与挑战,并对未来研究方向进行展望。

关键词

航空航天材料;人工智能(AI);材料基因组;代理模型;优化设计;机器学习(ML);

AI-Assisted Design of Aerospace Advanced Materials: Recent Advances and Perspectives

  • SUN Sheng 1,2
  • CHEN Yiyuan 1
  • SHANG Qing 1
1.Materials Genome Institute, Shanghai University, Shanghai 200444, China
2.Shanghai Frontier Science Center of Mechanoinformatics, Shanghai 200072, China

Citations

SUN Sheng, CHEN Yiyuan, SHANG Qing. AI-assisted design of aerospace advanced materials: recent advances and perspectives[J]. Aeronautical Manufacturing Technology, 2025, 68(18): 26–44.

Abstract

Extreme operating environments pose significant challenges for the next generation of aerospace materials. Traditional materials design methods, characterized by low efficiency, high costs, and long development cycles, have severely hindered the advancement of aerospace materials development. The development of new aerospace materials calls for innovative, highly efficient, and precise research and development paradigms. Artificial intelligence (AI) technologies, particularly the rapid advances in machine learning and deep learning, have emerged as powerful tools for aerospace materials research, markedly enhancing the efficiency of new material designs and the accuracy of performance predictions. This paper provides a systematic review of the research progress of AI in the aerospace materials field. It begins with an introduction to AI-assisted multiscale computational simulation and intelligent experimentation, then comprehensively presents surrogate model-accelerated materials optimization methods and a new materials design process centered on large-scale models. Detailed case studies are also presented on AI applications in the research and development of alloy materials, composite materials, and metamaterials. Finally, the paper summarizes the advantages and challenges of AI-assisted aerospace materials design and offers insights into future research directions.

Keywords

Aerospace materials; Artificial intelligence (AI); Materials genome; Surrogate models; Optimization design; Machine learning (ML);



随着航空航天任务复杂性的不断增加,以及极端服役条件对材料性能要求的不断提高,先进航空航天材料的研发需求日益迫切。例如,为满足长距离飞行和载人飞行任务对长航程、高载荷能力的要求,开发轻质高强材料已成为关键研究方向,这类先进材料可有效实现飞行器结构的轻量化,显著降低燃料消耗,从而提高飞行器的燃油经济性,促进远程航天任务和星际探索的实现[   ZHU L, LI N, CHILDS P R N. Light-weighting in aerospace component and system design[J]. Propulsion and Power Research, 2018, 7(2): 103–119.
1
]
。同时,高超声速飞行器在实际飞行过程中须应对极端严酷的工作环境,包括比常规太阳辐射高3~7个数量级的热流密度、从–170 ℃到3000 ℃的剧烈温度梯度、105~107 Pa的高停滞压力,以及等离子体侵蚀导致的严重材料氧化和结构退化问题。因此,开发具备卓越高温稳定性、抗氧化性能和高机械性能的先进热防护材料,对于保证飞行器在极端条件下的结构完整性和可靠性具有重要意义[   PETERS A B, ZHANG D J, CHEN S, et al. Materials design for hypersonics[J]. Nature Communications, 2024, 15: 3328.
2
]
。此外,激光武器(Laser weapons)和定向能武器(Directed energy weapons)的广泛应用也为飞行器材料的开发带来了新的挑战。飞行器材料须承受这类武器产生的高能电磁辐射或粒子束所带来的剧烈热冲击、机械损伤及等离子体侵蚀。因此,开发具备高效电磁屏蔽和热防护性能等在极端工作环境下服役的新型材料,已成为提升飞行器综合防护能力的重要研究方向[   SONI R, VERMA R, KUMAR GARG R, et al. Progress in aerospace materials and ablation resistant Coatings: A focused review[J]. Optics & Laser Technology, 2024, 177: 111160.
3
]

近年来,航空航天领域涌现出诸多新型材料,如高熵合金、3D打印制备的碳纤维复合材料等。这些先进材料的开发涉及众多设计工艺参数、材料变量和结构参数等,例如高熵合金须优化合金成分及微观结构控制[   YE Y F, WANG Q, LU J, et al. High-entropy alloy: Challenges and prospects[J]. Materials Today, 2016, 19(6): 349–362.
4
]
,3D打印碳纤维复合材料需考虑打印路径、工艺参数及层间结合性能等[   LIU G, XIONG Y, ZHOU L M. Additive manufacturing of continuous fiber reinforced polymer composites: Design opportunities and novel applications[J]. Composites Communications, 2021, 27: 100907.
5
]
。随着材料设计空间的显著扩大,传统依赖试错法的研发模式成本高昂且周期过长,已难以满足现代航空航天对新材料快速、高效开发的需求。近年来,借助人工智能(Artificial intelligence,AI)技术,材料研发正在向数据智能驱动转型。目前,AI技术已经全面渗透至数据驱动的材料研发全过程,为新材料的设计与优化提供了全新的方法论。

一方面,在AI辅助下,通过模拟计算和试验获取高质量数据的效率大幅提升[   LI J C, YE H L, DONG Y J, et al. An efficient deep learning-based topology optimization method for continuous fiber composite structure[J]. Acta Mechanica Sinica, 2024, 41(4): 424207.
6
]
。例如,在模拟计算领域,将分子动力学与机器学习力场相结合,不仅提升了计算精度,还显著提高了计算效率[   WANG H, ZHANG L F, HAN J Q, et al. DeePMD–kit: A deep learning package for many-body potential energy representation and molecular dynamics[J]. Computer Physics Communications, 2018, 228: 178–184.
7
]
。与此同时,AI与自动化技术的深度融合使智能化试验成为可能[   RACCUGLIA P, ELBERT K C, ADLER P D F, et al. Machine-learning-assisted materials discovery using failed experiments[J]. Nature, 2016, 533: 73–76.
8
]
。在智能实验室中,机器人能够自主执行试验,并实时调整和优化试验参数,从而大幅提高试验效率和数据质量[   SZYMANSKI N J, RENDY B, FEI Y, et al. An autonomous laboratory for the accelerated synthesis of novel materials[J]. Nature, 2023, 624: 86–91.
9
]
。另一方面,机器学习(Machine learning,ML)[   CHEN Q B, HE Z J, ZHAO Y, et al. Stacking ensemble learning assisted design of Al–Nb–Ti–V–Zr lightweight high-entropy alloys with high hardness[J]. Materials & Design, 2024, 246: 113363.
10
]
和深度学习(Deep learning,DL)[   YAN Y G, LIAO Y L, WANG K. Accelerated discovery of oxidation-resistant ultra-high temperature ceramics via data driven methodology[J]. Corrosion Science, 2023, 223: 111457.
11
]
算法凭借其强大的数据挖掘和模式识别能力,被用来训练代理模型,显著加速了材料设计和优化过程[   ON H I, JEONG L, SEO T M, et al. Novel method of performance-optimized metastructure design for electromagnetic wave absorption in specific band using deep learning[J]. Engineering Applications of Artificial Intelligence, 2024, 137: 109274.
12
]
。此外,近年来基于大语言模型(Large language model,LLM)的材料开发方法也开始崭露头角[   SRIRAM A, MILLER B K, CHEN R T Q, et al. FlowLLM: Flow Matching for material generation with large language models as base distributions[C]//Proceedings of NeurIPS 2024. San Diego: NeurIPS Foundation, 2024.
13
]
。借助其在自然语言处理和知识整合方面的优势,这些方法有助于突破传统思维局限,为材料科学领域带来更多创新性解决方案和前瞻性思路[   LIU X, SUN P, ZHANG L, et al. Perovskite–LLM: Knowledge-enhanced large language models for perovskite solar cell research[J/OL]. (2025–02–18)[2025–03–15]. https://doi.org/10.48550/arXiv.2502.12669.
14
]

本文将从数据驱动角度系统地介绍AI在航空航天新材料研发方面的方法和应用。首先,介绍AI技术在多尺度计算模拟与智能化试验中的应用情况;其次,深入探讨以AI算法为核心的材料优化设计方法,涵盖经典的机器学习代理模型及基于代理模型的优化设计方法介绍,并重点介绍近年来快速兴起的LLM在材料设计中的创新应用;接着,通过具体案例展示AI在航空航天用合金材料、复合材料及超材料开发中的实际应用;最后,本文对AI技术在航空航天材料研发中的应用进行总结,并展望了未来研究亟待解决的关键问题与挑战。本文有望为航空航天材料领域的研究人员提供理论参考和技术支持,推动AI与材料科学的深度融合与协同创新。

1     结合AI的材料计算模拟与智能化试验

计算模拟和智能化试验作为材料数据获取的两种主要途径,在数据驱动的材料研发中发挥着关键作用,高质量数据的获取是建立准确材料结构与性能关系的基础。当前,AI与计算模拟技术相结合,大幅提升了模拟计算的效率与准确性;同时,AI驱动的智能化试验方法通过自动化和并行化的试验流程,显著提高了数据采集效率与质量。

1.1     AI辅助多尺度模拟计算

1.1.1     微观尺度

密度泛函理论(Density functional theory,DFT)作为量子力学尺度的重要工具,其核心思想是将能量看作电子密度的泛函,而电子密度又是空间坐标的函数。从而将多电子问题简化为电子密度的泛函优化问题。该理论使用交换–关联泛函(如局域密度近似、广义梯度近似)近似处理电子间的相互作用。美国的Materials Project[   JAIN A, ONG S P, HAUTIER G, et al. Commentary: The Materials Project: A materials genome approach to accelerating materials innovation[J]. APL Materials, 2013, 1(1): 011002.
15
]
存储了数十万种材料的DFT计算数据,年计算量超过1×108 CPU·h,能够为航空材料的筛选提供重要支持;近年来发展起来的NOMAD[   SCHEIDGEN M, HIMANEN L, LADINES A N, et al. NOMAD: A distributed web-based platform for managingmaterials science research data[J]. Journal of Open Source Software, 2023, 8(90): 5388.
16
]
数据库更通过标准化数据格式,结合python库(如ASE[   HJORTH LARSEN A, JØRGEN MORTENSEN J, BLOMQVIST J, et al. The atomic simulation environment—A Python library for working with atoms[J]. Journal of Physics: Condensed Matter, 2017, 29(27): 273002.
17
]
、pymatgen[   ONG S P, RICHARDS W D, JAIN A, et al. Python materials genomics (PYMATGEN): A robust, open-source Python library for materials analysis[J]. Computational Materials Science, 2013, 68: 314–319.
18
]
)实现了跨平台DFT数据的智能检索与对比分析。作为第三范式的技术支柱,AI已深度重构DFT计算范式:Xu等[   TANG Z C, LI H, LIN P Z, et al. A deep equivariant neural network approach for efficient hybrid density functional calculations[J]. Nature Communications, 2024, 15: 8815.
  WANG Y X, LI Y, TANG Z C, et al. Universal materials model of deep-learning density functional theory Hamiltonian[J]. Science Bulletin, 2024, 69(16): 2514–2521.
19-20
]
提出了DeepH模型,利用神经网络高效预测了DFT哈密顿量,在大幅降低计算成本的同时实现了亚meV级别的精度;Wang等[   WANG J Q, LIU J P, WANG H S, et al. A comprehensive transformer-based approach for high-accuracy gas adsorption predictions in metal-organic frameworks[J]. Nature Communications, 2024, 15: 1904.
21
]
则开发了深度学习框架Uni–MOF,通过自监督预训练学习三维MOF/COF材料的结构表征,实现了在不同温度、压力和气体条件下对MOF吸附性能的高精度预测;Meta发布了基于开源大规模无机材料DFT数据集OMat24的预训练模型EquiformerV2[   BARROSO-LUQUE L, SHUABI M, FU X, et al. Open materials 2024 (OMat24) inorganic materials dataset and models[J/OL]. (2024–10–16)[2025–03–15]. https://doi.org/10.48550/arXiv.2410.12771.
22
]
,能量误差不超过 10 meV/atom;微软开发的MatterSim[   YANG H, HU C, ZHOU Y, et al. MatterSim: A deep learning atomistic model across elements, temperatures and pressures[J/OL]. (2024–05–08)[2025–03–15]. https://doi.org/10.48550/arXiv.2405.04967.
23
]
模型在0~5000 K和0~1000 GPa条件下高精度计算材料的能量、力和应力,可准确获得晶格动力学、力学及热力学性质,通过高精度的自由能预测支撑相图构建。

1.1.2     原子尺度

在量子尺度的DFT计算基础上,原子尺度的分子动力学模拟(Molecular dynamics simulations,MD)基于经典牛顿力学模拟原子/分子体系的动态演化。通过数值积分牛顿运动方程来描述分子系统中每个原子的运动。每个原子在每个时间步上的位置和速度由作用在其上的力决定,这些力来自于分子间的相互作用和外部条件。MD模拟优势在于可模拟纳米尺度下材料的动态响应过程,如位错运动、裂纹扩展等。其核心运动方程可表示为mid2ridt2=Fi(其中,mi是粒子质量,ri是位置,Fi是受力)。传统MD依赖预设势函数(如Lennard–Jones势[   SCHWERDTFEGER P, WALES D J. 100 years of the lennard-Jones potential[J]. Journal of Chemical Theory and Computation, 2024, 20(9): 3379–3405.
24
]
和反应力场ReaxFF[   VAN DUIN A C T, DASGUPTA S, LORANT F, et al. ReaxFF: A reactive force field for hydrocarbons[J]. Journal of Physical Chemistry A, 2001, 105(41): 9396–9409.
25
]
)描述粒子间相互作用。NOMAD[   SCHEIDGEN M, HIMANEN L, LADINES A N, et al. NOMAD: A distributed web-based platform for managingmaterials science research data[J]. Journal of Open Source Software, 2023, 8(90): 5388.
16
]
数据库包含原子结构、电子结构、力场参数等多种类型的数据,美国国家标准与技术研究院设置了专门的力场数据库[   HALE L. NIST interatomic potentials repository[EB/OL]. (2016–02–25)[2025–03–15]. https://doi.org/10.18434/M37.
26
]
。机器学习力场(Machine learning force fields,MLFF)通过数据驱动方式重构原子间势能函数,突破了传统经验势函数精度与泛化能力的限制,其核心原理是将势能面建模为原子局部环境特征的函数,例如,DeepMD[   WANG H, ZHANG L F, HAN J Q, et al. DeePMD–kit: A deep learning package for many-body potential energy representation and molecular dynamics[J]. Computer Physics Communications, 2018, 228: 178–184.
7
]
利用深度神经网络构建机器学习立场,通过原子类型、位置及其近邻原子的几何信息(如距离、角度)构建局部环境描述符,实现势能和原子受力的端到端预测;MACE[   BATATIA I, KOVÁCS D P, SIMM G N C, et al. MACE: Higher order equivariant message passing neural networks for fast and accurate force fields[C]//Proceedings of NeurIPS 2022. San Diego: NeurIPS Foundation, 2022.
27
]
基于等变神经网络架构,引入多体展开理论,确保势能预测满足物理对称性;GemNet[   GASTEIGER J, BECKER F, GÜNNEMANN S. GemNet: Universal directional graph neural networks for molecules[J]. Advances in Neural Information Processing Systems, 2021, 34: 6790–6802.
28
]
利用图神经网络显式建模原子间几何关系,通过边–节点交互机制捕捉长程相互作用与多体效应;而PANNA[   PELLEGRINI F, LOT R, SHAIDU Y, et al. PANNA 2.0: Efficient neural network interatomic potentials and new architectures[J]. Journal of Chemical Physics, 2023, 159(8): 084117.
29
]
则通过人工神经网络直接从第一性原理计算数据中学习高维势能面,显著降低了计算成本;GPTFF[   XIE F K, LU T L, MENG S, et al. GPTFF: A high-accuracy out-of-the-box universal AI force field for arbitrary inorganic materials[J]. Science Bulletin, 2024, 69(22): 3525–3532.
30
]
利用transformer和注意力机制,在3780万单点能量数据基础上训练出可以精准计算结构优化、相变模拟的模型,预测能量、原子力和应力的平均绝对误差分别低至32 meV/atom、71 meV/Å、0.365 GPa。

1.1.3     宏观尺度

在宏观连续介质力学问题中,有限元方法(Finite element method,FEM)通过将连续域离散化为有限个简单单元,在单元上近似求解偏微分方程。其基本思想是在每个单元内采用函数展开,使用近似函数来描述场变量(如温度、压力等),再通过组装各单元方程构成全局方程组进行求解。Li等[   LI L, SHAO Q, YANG Y C, et al. A database construction method for data-driven computational mechanics of composites[J]. International Journal of Mechanical Sciences, 2023, 249: 108232.
31
]
采用有限元技术对代表体积元(Representative volume element,RVE)进行数值计算,获得了复合材料在弹塑性行为下的高保真应力—应变数据,并构建了一个小规模但高精度的样本数据库;随后,利用这些样本数据训练神经网络,对数据进行扩充,生成了一个高密度数据库,从而显著降低了大规模RVE计算所需的时间和成本。Liu等[   LIU Z L, FLEMING M, LIU W K. Microstructural material database for self-consistent clustering analysis of elastoplastic strain softening materials[J]. Computer Methods in Applied Mechanics and Engineering, 2018, 330: 547–577.
32
]
则采用有限元方法采集了因孔洞和夹杂物引起的高保真应力—应变数据,并通过k–means聚类对这些数据进行降维和压缩,构建出反映材料微观结构异质性的数据库。与此同时,Kohar等[   KOHAR C P, GREVE L, ELLER T K, et al. A machine learning framework for accelerating the design process using CAE simulations: An application to finite element analysis in structural crashworthiness[J]. Computer Methods in Applied Mechanics and Engineering, 2021, 385: 114008.
33
]
提出了一种基于3D–CNN自动编码器和LSTM网络的机器学习框架,通过对有限元网格数据进行自动特征提取,实现了对结构碰撞响应的快速预测。该方法大幅加速了设计流程,其预测速度在保持较高精度的前提下,比传统有限元方法快了数百~数百万倍。此外,Kim等[   KIM S, SHIN H. Deep learning framework for multiscale finite element analysis based on data-driven mechanics and data augmentation[J]. Computer Methods in Applied Mechanics and Engineering, 2023, 414: 116131.
34
]
开发了一种基于深度学习的数据驱动多尺度有限元分析框架,通过自适应数据增强和距离最小化方法构建宏观材料基因数据库,既降低了离线RVE计算成本,又提高了仿真精度。

1.2     AI赋能智能化试验

AI实验室在材料研发领域展现出了强大的能力,尤其是在智能化试验数据获取方面发挥着关键作用,为材料研发带来了前所未有的变革。AI实验室通过构建智能化试验平台,结合自动化技术和机器人操作,能够快速进行大量材料样品的制备和性能测试。Angelopoulos等[   ANGELOPOULOS A, CAHOON J F, ALTEROVITZ R. Transforming science labs into automated factories of discovery[J]. Science Robotics, 2024, 9(95): eadm6991.
35
]
将实验室自动化级别分为五类,如图1所示,其范围从“辅助(Assistive)”到“全面(Full)”;“辅助”阶段指自动化仅协助完成单个步骤,而“全面”阶段则实现实验室自主完成整个过程,包括管理不确定性和自我维护,无需人为干预;在完全自动化阶段,机器人和AI能够实现自主运行和安全管理。

图1     实验室自动化水平[   ANGELOPOULOS A, CAHOON J F, ALTEROVITZ R. Transforming science labs into automated factories of discovery[J]. Science Robotics, 2024, 9(95): eadm6991.
35
]
Fig.1     Laboratory automation level[   ANGELOPOULOS A, CAHOON J F, ALTEROVITZ R. Transforming science labs into automated factories of discovery[J]. Science Robotics, 2024, 9(95): eadm6991.
35
]

近年来,AI实验室的突破性进展正在重构材料研发范式。由多伦多大学Alán Aspuru-Guzik教授领衔的国际团队展示了一种分布式、异步、闭环发现流程[   STRIETH-KALTHOFF F, HAO H, RATHORE V, et al. Delocalized, asynchronous, closed-loop discovery of organic laser emitters[J]. Science, 2024, 384(6697): eadk9227.
36
]
,通过整合全球多个实验室的自动化合成和表征模块,以及基于云的AI优化器,成功实现了有机激光发射器增益材料的高效发现;该研究不仅克服了传统材料发现中地域限制和资源分散的难题,还通过模块化合成策略和自动化工作流程,大幅提高了材料发现的效率和成功率;研究团队共发现了21种具有改进发射增益截面的小分子发射器,其中部分材料在薄膜器件中表现出同类最佳的放大自发发射阈值。

Google DeepMind团队开发的A–Lab自主试验平台[   SZYMANSKI N J, RENDY B, FEI Y, et al. An autonomous laboratory for the accelerated synthesis of novel materials[J]. Nature, 2023, 624: 86–91.
9
]
,通过整合计算模拟、历史数据、机器学习与自动化试验技术,在17天内成功合成了74%的目标无机粉末材料,其中包括多种氧化物和磷酸盐,合成配方由自然语言模型提出并基于热力学主动学习进一步优化,体现了AI驱动的自主材料研发平台的高效性。

美国北卡罗来纳州立大学开发的Fast–Cat催化实验室[   BENNETT J A, OROUJI N, KHAN M, et al. Autonomous reaction Pareto-front mapping with a self-driving catalysis laboratory[J]. Nature Chemical Engineering, 2024, 1(3): 240–250.
37
]
能够自主进行高温、高压、气液反应的试验规划、执行和优化,仅用5天就完成了45个反应的自主试验,远超传统方法在6个月周期内所能获得的数据量。Dai等[   DAI T W, VIJAYAKRISHNAN S, SZCZYPIŃSKI F T, et al. Autonomous mobile robots for exploratory synthetic chemistry[J]. Nature, 2024, 635: 890–897.
38
]
介绍的一种基于自主移动机器人的探索性合成化学平台,整合分布式合成和分析设备,实现了从反应筛选到产物验证的全流程自动化,其模块化和可扩展的设计适用于多种不同的试验领域。这些智能化试验平台和自动化技术能够缩短航空航天材料研发的周期,提高研发效率。

在这一过程中,AI算法发挥着关键作用,它能够根据试验数据实时调整试验参数和方案,实现试验过程的智能优化。AI实验室利用先进的计算模拟技术,结合AI算法,能够对航空航天材料在复杂工况下的性能进行多物理场耦合模拟,为材料的设计和应用提供重要的理论支持。中国科大设计的化学平台“ChemAgents”[   SONG T, LUO M, ZHANG X L, et al. A multiagent-driven robotic AI chemist enabling autonomous chemical research on demand[J]. Journal of the American Chemical Society, 2025, 147(15): 12534–12545.
39
]
是一种基于多智能体系统的机器人AI化学家,能够实现按需自主化学研究;该平台通过整合LLM和智能化实验室技术,展示了AI在复杂化学试验中的能力。

Ada平台[   MACLEOD B P, PARLANE F L, MORRISSEY T D, et al. Self-driving laboratory for accelerated discovery of thin-film materials[J]. Science Advances, 2020, 6(20): eaaz8867.
40
]
能够自主合成、处理和表征有机薄膜,自训练在没有任何先验知识的情况下找到目标参数,实现迭代试验设计,从而最大限度地提高每个样本的信息增益。这种有机薄膜的高空穴迁移率和良好成膜性,使之在航空航天电子设备和传感器中具有潜在应用价值。

AI实验室在试验效率、可重复性与自适应能力等方面展现出显著优势,然而当前仍难以实现真正意义上的自主高通量试验,其方案设计与执行过程仍高度依赖人工干预。为此,需要构建集成高通量试验设计与在线表征的闭环体系,将试验过程嵌入自适应决策与反馈优化算法中,从而实现“无人工干预—自驱闭环—持续优化”的全流程智能化高通量研发模式。

2     以AI算法为核心的材料设计

2.1     机器学习代理模型

在航空航天新材料设计中,精确模拟[   CURRAN R, RAGHUNATHAN S, PRICE M. Review of aerospace engineering cost modelling: The genetic causal approach[J]. Progress in Aerospace Sciences, 2004, 40(8): 487–534.
41
]
与实际试验[   AYKAN M, ÇELIK M. Vibration fatigue analysis and multi-axial effect in testing of aerospace structures[J]. Mechanical Systems and Signal Processing, 2009, 23(3): 897–907.
42
]
通常需要耗费巨额成本和大量时间,例如高保真的飞机飞行模拟[   KROLL N, ABU-ZURAYK M, DIMITROV D, et al. DLR project digital–X: Towards virtual aircraft design and flight testing based on high-fidelity methods[J]. CEAS Aeronautical Journal, 2016, 7(1): 3–27.
43
]
、飞机设计中的风洞测试[   CATTAFESTA L, BAHR C, MATHEW J. Fundamentals of wind-tunnel design[M]//Encyclopedia of Aerospace Engineering. Hoboken: John Wiley & Sons Inc., 2010.
44
]
等。为了在确保设计性能的前提下降低研发成本并缩短设计周期,代理模型作为一种高效的计算替代工具得到了广泛应用[   AZARHOOSH Z, ILCHI GHAZAAN M. A review of recent advances in surrogate models for uncertainty quantification of high-dimensional engineering applications[J]. Computer Methods in Applied Mechanics and Engineering, 2025, 433: 117508.
  FORRESTER A I J, SÓBESTER A, KEANE A J. Engineering design via surrogate modelling: A practical guide[M]. Hoboken: John Wiley & Sons Inc., 2008.
45-46
]
。代理模型f^x)通过近似高计算成本的原始模型f:ℝd→ℝ的输入–输出关系,替代昂贵且耗时的试验和仿真过程,其中d表示输入变量的维度。

传统代理模型主要依托物理理论、数学近似和统计方法近似描述复杂系统的输入–输出关系,例如通过多阶段贝叶斯代理模型、协方差方法等构建代理模型。尽管这些方法在一定程度上可替代模拟和试验,但在处理高维、非线性及多目标优化问题时存在局限,并需大量人工调试和参数设定。

代理模型构建中,机器学习和深度学习算法的引入为传统方法带来了全新的改进。基于AI的代理模型利用数据进行训练,能够自动捕捉复杂的非线性关系和高维特征,经过训练的代理模型通常能够比全精度仿真快几个数量级,能快速完成从输入到输出的计算,并在预测准确性方面为后续精细设计提供可靠支持[   SAMADIAN D, MUHIT I B, DAWOOD N. Application of data-driven surrogate models in structural engineering: A literature review[J]. Archives of Computational Methods in Engineering, 2025, 32(2): 735–784.
47
]
。基于这些代理模型,各种优化算法能够高效进行设计优化,进一步缩短设计周期、降低资源消耗,加速材料设计过程[   KOCHENDERFER M J, WHEELER T A. Algorithms for optimization[M]. Cambridge: The MIT press, 2019.
48
]
。下面介绍响应面模型(Response surface model,RSM)、克里金(Kriging)代理模型、径向基函数(Radial basis function,RBF)模型、支持向量机回归(Support vector regression,SVR)模型和深度学习代理模型这5种经典的机器学习代理模型。

2.1.1     RSM

RSM是一种经典且广泛应用于工程领域的代理模型[   KHURI A I, MUKHOPADHYAY S. Response surface methodology[J]. WIREs Computational Statistics, 2010, 2(2): 128–149.
49
]
。其基本思想是将代理模型f^x)近似为输入变量x的多项式函数(主要为一阶和二阶)的线性组合,以逼近原始模型fx)的输出。RSM的一阶形式为

f^(x,a)=a0+i=1maixi=a0+a1x1++amxm
(1)

其中,xim维输入变量中的第i个分量,向量a=(a0a1,…,amT可通过最小化真实值与预测值之间的误差平方和进行求解,即最小二乘优化问题为

min(yrealypredicted)2
(2)

其中,yreal是原始模型的真实值,ypredicted是代理模型的预测值。

对于非线性问题,通常采用更高阶的多项式模型。例如,RSM的二阶模型可表示为

f^(x,a,α)=a0+i=1maixi+i=1mαiixi2+i=1m1j=i+1mαijxixj
(3)

其中,线性项aixi描述变量的直接影响,二次项αiixi2刻画变量的非线性效应(如曲率),交互项αijxixj反映变量间的相互作用。

RSM适用于低维、单峰、平滑的优化问题,例如,Ferdosian等[   FERDOSIAN I, CAMÕES A. Eco-efficient ultra-high performance concrete development by means of response surface methodology[J]. Cement and Concrete Composites, 2017, 84: 146–156.
50
]
采用RSM对超高性能混凝土的硅灰、超细粉煤灰和砂的配比进行了多目标优化,成功提升了超高性能混凝土的生态效益和力学性能。然而,RSM在应对高维、多峰和强非线性设计空间时表现受限,模型精度和构建效率显著下降,难以胜任复杂材料设计任务。

2.1.2     Kriging

Kriging代理模型是一种基于高斯过程建模的代理方法[   RASMUSSEN C E, WILLIAMS C K I. Gaussian processes for machine learning[M]. Cambridge: The MIT press, 2006.
51
]
,其基本形式为

f^(x)=j=1majgj(x)+ϵ(x)
(4)

其中,gjx)表示m个相互独立且已知的基函数,用以描述点x处均值预测的趋势,aj为待估计参数,而ϵx)表示在点x处的随机误差,该误差服从均值为零的正态分布。

用于预测的Kriging代理模型可进一步表述为

f^(x)=g(x)Ta*+r(x)Tα*
(5)

其中,gx)=[g1x),g2x),…,gmx)]T为基函数在点x处的取值向量;rx)为一个n×1的相关性向量,描述了ϵx)与各样本点上误差ϵxi)之间的相关性。而a*α*的计算公式为

a*=(GTR1G)1GTR1yα*=R1(yGa*)
(6)

其中,Rn×n的协方差矩阵,其第(ij)个元素描述了ϵxi)与ϵxj)之间的相关性;G=[gx(1)),gx(2)),…,gxn)]Tn×m的矩阵,而y表示在可用试验点处的观测值。此外,过程方差σ2可计算为

σ2=1n(yGa*)TR1(yGa*)
(7)

在kriging代理模型中,假设随机变量ϵx)的相关性遵循由超参数θ参数化的相关模型R(·,·),这些超参数通常采用最大似然估计方法进行求解[   KAYMAZ I. Application of Kriging method to structural reliability problems[J]. Structural Safety, 2005, 27(2): 133–151.
52
]

另一个重要特性是选择合适的协方差函数。通常,Kriging模型采用的协方差函数是平稳的,其表达式为

R(x,x)=jψj(θ,mj),mj=xjxj
(8)

这种形式具有两大优点。首先,可以通过多个一维相关函数的乘积来表达多变量函数之间的相关性;其次,相关性是平稳的,仅依赖于两点xx′在第j维上的距离。表1列出了常用的核函数模型,用于刻画输入变量之间的相关性,其中m表示输入变量在某一维度上的距离,θ为尺度参数(scale parameter),控制相关性随距离增大的衰减速率。Exponential模型适用于“非平滑”的目标函数;Gaussian模型适合于平滑且可微的函数建模;Matérn模型则引入平滑度参数ν,可灵活调节函数光滑性,Γ(ν)为归一化所用的伽马函数,Kν(·)为第ν阶修正贝塞尔函数,用于描述远距离下的相关性衰减。当ν较大时,Matérn模型的形式逐渐趋近于Gaussian模型。

表1     常用的核函数及其数学形式
Table 1     Common kernel functions and their mathematical forms
名称 数学形式
Exponential ψ(θ,m)=exp(θ|m|)
Gaussian ψ(θ,m)=exp(θm2)
Matérn ψ(θ,m)=1Γ(ν)2ν1(2νθ|m|)νKν(2νθ|m|)

Kriging模型不仅能够进行响应预测,还具备量化不确定性的能力,适用于样本有限、计算密集型的材料设计问题,特别适合处理中等维度、局部变化显著的复杂响应函数。它在多尺度建模、可靠性分析和灵敏度评估中表现优越。在材料设计中,Zhou等[   ZHOU C C, LI C, ZHANG H L, et al. Reliability and sensitivity analysis of composite structures by an adaptive Kriging based approach[J]. Composite Structures, 2021, 278: 114682.
53
]
采用自适应Kriging方法结合蒙特卡洛模拟,对复合材料结构的可靠性与灵敏度进行了定量分析;首先利用有限元构建复杂结构(如复合结构)的响应模型,再构建高效的Kriging代理模型来描述材料和几何参数的不确定性;通过自适应采样不断更新近似模型,该方法显著减少了真实模型的调用次数,并在高维与非线性问题中准确估计了失效概率。

2.1.3     RBF

在处理复杂的原始模型时,通常会采用便于分析的基函数来逼近系统的输入–输出关系,例如2.1.1节中介绍的多项式模型。另一种在工程领域中广泛应用的方法是RBF模型[   GUTMANN H M. A radial basis function method for global optimization[J]. Journal of Global Optimization, 2001, 19(3): 201–227.
54
]

首先,通过试验设计在设计空间中选取若干样本点。设第i个样本点为

xi=(xi1,xi2,xi3,,xim)(i=1,2,,n)
(9)

其中每个样本点均为m维。RBF模型的基本思想是利用待测点x与每个样本点xi之间的欧几里得距离ri=‖xxi‖作为自变量,通过对应的径向函数ϕri)构造基函数,然后将这些径向函数进行线性组合以获得待测点x的响应值。模型表达式可写为

f^(x)=i=1nwiϕ(ri)
(10)

其中,wi为通过最小二乘法计算得到的权重系数。

RBF模型适用于中低维、函数平滑的插值问题,具有较强的局部拟合能力。它在样本充足且均匀分布时能实现高精度建模,常用于工程结构与材料参数空间的快速响应面构建。需要注意的是,RBF模型在插值任务中的表现取决于所选用的基函数类型。表2列出了常用的RBF模型,其中r表示待测点x与任一样本点之间的欧几里得距离,c是大于零的常数参数,也称为形状参数。不同的径向函数会赋予模型不同的性质和逼近效果。例如,Sun等[   SUN G Y, LI G Y, GONG Z H, et al. Radial basis functional model for multi-objective sheet metal forming optimization[J]. Engineering Optimization, 2011, 43(12): 1351–1366.
55
]
在钣金成形优化研究中,将RBF代理模型与多目标粒子群优化算法相结合,实现了对汽车内侧面板断裂和起皱风险的同步优化;构建的RBF模型高效近似有限元分析结果,有效提升了预测精度并缩短了设计周期。

表2     常见的RBF模型
Table 2     Common RBF models
名称 ϕr 条件
Gaussian exp(r2c2) c >0
Multi-quadratic (MQ) r2+c2 c >0
Inverse multi-quadratic 1r2+c2 c >0
Thin-plate splines (rc2)ln(rc) c >0

2.1.4     SVR

SVR是将支持向量机(Support vector machine,SVM)扩展到回归问题的一种方法[   SMOLA A J, SCHÖLKOPF B. A tutorial on support vector regression[J]. Statistics and Computing, 2004, 14(3): 199–222.
56
]
,其核心目标是构建一个回归函数f^x),使得大部分训练样本的预测误差不超过预设的阈值δ,同时保持模型平滑,从而提高泛化能力。

SVR的预测模型通常表示为一系列预先设定的基函数ψ(在SVR中也称为核函数)的加权和,再加上一个常数项μ,即

f^(x)=μ+i=1nwiψ(x,x(i))
(11)

对于线性映射,利用内积运算可以简化为:

f^(x)=μ+w,x
(12)

在构建线性SVR模型时,其目标是寻找一组权重w,使得大部分样本点的预测值与真实值之间的偏差不超过δ。这一目标可以通过求解如式(13)的凸优化问题实现。

min12w2+γi=1n(ξi++ξi)
(13)

约束条件为

{y(i)w,x(i)μδ+ξi+,i=1,,nw,x(i)+μy(i)δ+ξi,i=1,,n ξi+,ξi0,i=1,,n
(14)

其中,松弛变量ξi+ξi用于容许部分样本的预测误差超出δ的范围,而正则化参数γ则平衡模型复杂度与容忍误差之间的权衡。利用拉格朗日对偶方法可以求解上述优化问题,从而得到最优权重w

拉格朗日对偶方法不仅适用于传统的线性回归模型,还能结合核函数ψxx′)有效处理非线性问题。常见的核函数包括线性核函数(ψxx′)=〈xx′〉)和高斯核函数(ψxx′)=exp(xx2σ2))。通过这些核函数将输入数据映射到高维特征空间,SVR能捕捉更复杂的输入–输出关系;而映射后的优化问题仍为凸二次规划问题,通过最大化拉格朗日对偶函数,实现高效的参数求解。

SVR基于结构风险最小化原则,具备良好的泛化能力,适用于中小样本、特征维度适中、数据噪声较低的材料预测任务。它对异常值鲁棒且模型结构稳定,适合处理复杂性受控的非线性问题。在材料设计中,Zhai等[   ZHAI X Y, CHEN M T. Machine learning aided design of Bi2WO6/MIL–53(Al) nanocomposites[J]. Computational Materials Science, 2024, 233: 112737.
57
]
采用SVR构建了纳米复合材料的光催化活性预测模型:首先建立了基准数据集,并通过特征选择和遗传算法筛选出关键变量;随后优化了SVR–RBF模型的超参数,实现了对光催化活性的高精度预测。

2.1.5     深度学习代理模型

人工神经网络(Artificial neural network,ANN)是一种非线性数学模型,通过改变“神经元”层数和每层的个数,可以逼近任何函数[   WU Y C, FENG J W. Development and application of artificial neural network[J]. Wireless Personal Communications, 2018, 102(2): 1645–1656.
58
]
。每个神经元先将前一层输出按权重加权、加上偏置,然后通过非线性激活函数(如Sigmoid或ReLU)生成输出。多层感知器(Multilayer perceptron,MLP)是最常见的ANN结构,通常包括输入层、一个或多个隐藏层及输出层,训练时采用前馈传播结合反向传播算法。第l层第j个神经元的输出可表示为

aj(l)=σ(i=1nlwij(l)ai(l1)+bj(l))
(15)

其中,wijl为连接权重,bjl为偏置,σ(·)为激活函数。网络可通过最小化均方误差损失函数来更新权重和偏置,损失函数可表示为

L=12i=1n(yiy^i)2
(16)

一般把多于两层神经元的ANN称为深度神经网络,对深度神经网络的训练称为深度学习。根据自然语言处理、图像识别和时间序列数据处理等不同应用,形成了不同的深度学习网络和方法。在此基础上,先进的深度学习代理模型进一步融合了多种技术,并在材料开发领域取得了显著成果。例如,卷积神经网络(Convolutional neural network,CNN)通过堆叠更多层(例如卷积层、池化层等)来提取更高层次特征,从而进一步提升模型的学习与预测能力[   O’SHEA K, NASH R. An Introduction to Convolutional Neural Networks[J/OL]. (2015–11–26)[2025–03–15]. https://doi.org/10.48550/arXiv.1511.08458 (2015).
59
]
。Xie等[   XIE T, GROSSMAN J C. Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties[J]. Physical Review Letters, 2018, 120(14): 145301.
60
]
提出了基于晶体图卷积神经网络的方法,通过将晶体结构转化为包含多重边的拓扑图表示,并构建多层卷积与池化网络,逐层整合局部原子环境信息,实现了对形成能、带隙、模量等多种材料性质的高精度预测。同时,对抗神经网络(Generative adversarial network,GAN)[   CRESWELL A, WHITE T, DUMOULIN V, et al. Generative adversarial networks: An overview[J]. IEEE Signal Processing Magazine, 2018, 35(1): 53–65.
61
]
等生成式模型在材料开发中也展现出独特优势。Qian等[   QIAN C, TAN R K, YE W J. Design of architectured composite materials with an efficient, adaptive artificial neural network-based generative design method[J]. Acta Materialia, 2022, 225: 117548.
62
]
提出了一种基于GAN与CNN相结合的逆向生成设计方法,通过自适应学习与数据扩展策略,在显著减少标记数据的前提下,实现了对高韧性、高刚度及具有指定体积与剪切模量的复合材料的高效精准设计。此外,Transformer模型及其注意力机制通过捕捉全局依赖关系,为复杂系统多尺度特征的建模提供了全新思路[   POPEL M, BOJAR O. Training tips for the transformer model[J]. The Prague Bulletin of Mathematical Linguistics, 2018, 110(1): 43–70.
63
]
。Chen等[   CHEN W, GAO Y, LI Y Y, et al. Broadband solar metamaterial absorbers empowered by transformer-based deep learning[J]. Advanced Science, 2023, 10(13): 2206718.
64
]
提出的基于Transformer的深度学习方法,通过将宽带光谱分块处理并构建逆向与正向设计网络,实现了快速精确地设计出高吸收率(94%)且低热辐射的超材料太阳能吸收器。

相较于传统代理模型,深度学习代理模型在处理大样本、高维非线性和多目标优化等复杂问题中展现出更高的建模精度与扩展性,尤其适用于复杂材料系统。但该类方法对数据和计算资源依赖较强,可解释性较弱,存在“黑箱”风险,且其泛化性能易受模型结构和训练策略的影响。

2.2     基于代理模型的优化设计

在复杂工程或科学问题中,直接采用原始模型进行优化往往面临高昂的计算代价。利用训练好的代理模型进行优化设计,不仅能够保留原问题的关键特性,还能大幅降低计算开销。下面将从数据点的采样策略、代理模型的训练和优化算法3个方面介绍基于代理模型的优化设计。

2.2.1     采样策略

在构建数据驱动的代理模型时,模型的表现与用于训练的样本点在设计空间中的分布密切相关[   FUHG J N, FAU A, NACKENHORST U. State-of-the-art and comparative review of adaptive sampling methods for Kriging[J]. Archives of Computational Methods in Engineering, 2021, 28(4): 2689–2747.
65
]
。因此,合理的采样策略至关重要,其主要目标是均匀且有效地填充整个设计空间,以确保模型能够捕捉到变量之间的复杂关系与交互效应,尽量避免“外推拟合”现象发生。常见的采样方法包括均匀采样[   FANG K-T, LIN D K J, WINKER P, et al. Uniform design: Theory and application[J]. Technometrics, 2000, 42(3): 237–248.
66
]
、蒙特卡洛采样[   MORRIS M D, MITCHELL T J. Exploratory designs for computational experiments[J]. Journal of Statistical Planning and Inference, 1995, 43(3): 381–402.
67
]
和拉丁超立方体采样[   MCKAY M D, BECKMAN R J, CONOVER W J. A comparison of three methods for selecting values of input variables in the analysis of output from a computer code[J]. Technometrics, 2000, 42(1): 55–61.
68
]
等。

均匀采样旨在使所有样本点在设计空间中分布均匀,通过在每个维度上以相等的间隔选取样本点,从而最大限度地减少局部区域数据过密或过稀的现象,确保设计空间的全面覆盖,使训练得到的代理模型在不同区域内都能获得较为稳定的预测效果。蒙特卡洛采样依赖于随机生成样本点,其优势在于简单易行且适用于高维问题,尽管这种方法可能会导致部分区域样本较为稀疏且计算开销较大,但其随机性有助于捕捉设计空间中潜在的非线性和复杂性,尤其在处理黑箱函数时表现出较高的适用性。而拉丁超立方体采样则通过将每个设计变量划分为若干等长区间,并在每个区间中随机选取样本点来确保每个区间至少有一个样本,从而在保证均匀性的同时兼顾了随机采样的灵活性,在高维设计空间中实现更优的空间填充效果,特别适用于在有限样本量下最大限度地覆盖设计变量变化范围的场景。

2.2.2     代理模型训练

代理模型的训练过程主要包括模型选择、模型训练与超参数优化及模型验证3个关键步骤。首先,在模型选择阶段,需要确保输入和输出数据的格式与所选模型匹配,并对数据进行归一化或标准化处理,以保证各变量在相同尺度下进行计算;同时,应将采样点合理划分为训练集、验证集和测试集,其中训练集(约占70%)用于模型参数的学习和训练,验证集(约占10%)用于模型选择和超参数调优,测试集(约占20%)则用于最终评估,确保评估结果不受模型训练和选择过程的干扰。通过验证集对候选模型在新数据上的表现进行初步筛选,可以选出最合适的模型。接下来,在模型训练与超参数优化阶段,通过已有数据训练模型并调整各项超参数,以达到最优配置。以ANN为例,在训练过程中常采用反向传播算法更新网络的权重参数W,并通过误差反向传递不断降低预测误差E;同时为了提升模型的泛化能力,需要对诸多超参数进行系统调优,包括学习率(Learning rate,α)、正则化系数(Regularization coefficient,λ)、隐藏层数(Number of hidden layers,L),以及激活函数(Activation function,σ)等。在这一阶段,常用的调优方法包括网格搜索和贝叶斯优化,它们能够在较大的参数空间内寻找最优配置。此外,还可采用交叉验证、早停和正则化等技术进一步提升模型的鲁棒性和稳定性。最后,在确定模型结构和超参数后,利用训练集和验证集的全部数据对模型进行重新训练,模型通过测试数据集进行评估,常用的评估指标包括平均绝对误差(Mean absolute error,MAE)、决定系数(Coefficient of determination,R2)、相对平均绝对误差(Relative average absolute error,RAAE)以及相对最大绝对误差(Relative maximum absolute error,RMAE)。这些指标能够全面反映模型在未见数据上的预测能力,确保代理模型不仅在训练数据上具有高精度,同时在实际应用中也具备良好的泛化能力,有效避免过拟合。

2.2.3     优化算法

在航空航天新材料的设计中,优化算法发挥着关键作用,其核心目标是在庞大而复杂的设计空间中高效筛选出兼具性能与经济性的设计方案。依托训练良好的代理模型,优化算法能够充分利用其预测能力,在显著降低计算成本的同时,有效探索高维非线性空间,实现材料结构的逆向设计与性能优化。

种群优化算法[   BABALOLA A E, OJOKOH B A, ODILI J B. A review of population-based optimization algorithms[C]//Proceedings of 2020 International Conference in Mathematics, Computer Engineering and Computer Science (ICMCECS). New York: IEEE, 2020.
69
]
是一类受到自然界启发、模拟生物进化或群体协同行为的全局优化方法。该类算法具备搜索能力强、自适应性高、对目标函数形式要求低等优势,广泛应用于非线性、多峰、多目标及高维度等复杂优化问题中,包括遗传算法(Genetic algorithm,GA)[   LAMBORA A, GUPTA K, CHOPRA K. Genetic algorithm- a literature review[C]//Proceedings of 2019 International Conference on Machine Learning, Big Data, Cloud and Parallel Computing (COMITCon). New York: IEEE, 2019.
70
]
、差分进化(Differential evolution,DE)[   DAS S, SUGANTHAN P N. Differential evolution: A survey of the state-of-the-art[J]. IEEE Transactions on Evolutionary Computation, 2011, 15(1): 4–31.
71
]
、粒子群优化(Particle swarm optimization,PSO)[   WANG D S, TAN D P, LIU L. Particle swarm optimization algorithm: An overview[J]. Soft Computing, 2018, 22(2): 387–408.
72
]
、蚁群算法(Ant colony optimization,ACO)[   DORIGO M, BIRATTARI M, STUTZLE T. Ant colony optimization[J]. IEEE Computational Intelligence Magazine, 2006, 1(4): 28–39.
73
]
等。例如,GA通过交叉、变异和选择等操作实现迭代优化,特别适用于非线性和多模态问题。Kneidin等[   KNEIDING H, BALCELLS D. Augmenting genetic algorithms with machine learning for inverse molecular design[J]. Chemical Science, 2024, 15(38): 15522–15539.
74
]
将ANN与GA相结合,构建了用于自旋交叉配合物性能预测的优化框架;该研究基于DFT数据训练了用于预测高低自旋态能量差的ANN模型,并将其作为GA的代理适应度函数,实现了结构性能的高效评估。同时,PSO算法通过模拟粒子间信息交互,结合个体历史最优与全局最优信息动态更新粒子位置与速度,实现快速收敛。Jin等[   JIN P, YANG S, XU L J, et al. Particle swarm optimization for realizing bilayer thermal sensors with bulk isotropic materials[J]. International Journal of Heat and Mass Transfer, 2021, 172: 121177.
75
]
将PSO与有限元模拟结合,提出了一种面向双层热传感器的智能逆向设计框架;该方法以温度分布误差为核心构建多目标适应度函数,并将结构几何尺寸作为优化变量,在高维空间中实现高效寻优;数值模拟与试验结果表明,该方法在实现热隐身效果的同时保持了高精度的温度探测能力,展现出良好的工程适用性与通用性。

另外一类强大的优化算法是贝叶斯优化[   SHAHRIARI B, SWERSKY K, WANG Z Y, et al. Taking the human out of the loop: A review of Bayesian optimization[J]. Proceedings of the IEEE, 2016, 104(1): 148–175.
76
]
(其流程见图2),它通常与代理模型(如高斯过程)相结合,对目标函数进行概率预测建模。该方法首先通过采样少量初始样本点(可采用2.2.1节的采样策略)并获取目标函数值(例如通过模拟计算或试验测量)来构建初始训练数据集,为后续模型构建提供数据基础;然后利用高斯过程等模型对目标函数进行概率建模,预测每个设计点的函数均值,同时给出相应的不确定性(置信区间),这一过程在图2的右侧得到了直观展示;接下来,基于训练好的代理模型,通过构建如期望改进和置信上限等采集函数,在设计空间中评估每个候选点的潜在改进效果,从而平衡“探索”未知区域与“利用”已有高值区域的需求,如图2左侧所示,这些采集函数会给出一个或多个最优候选点xcand;随后,对采集函数选出的候选点进行实际目标函数的评估,并将新获得的样本数据加入训练数据集中以更新模型信息;此后,通过不断利用更新后的数据重新训练代理模型并重新计算采集函数,整个过程在有限的函数评估次数下不断迭代,逐步逼近全局最优解。基于高斯过程的不确定性量化与智能采样策略,使贝叶斯优化能够在较少的迭代次数内找到高质量解,从而极大降低了试验和计算成本,因而在航空航天新材料设计等高成本、高复杂度的优化问题中展现出显著优势。Serles等[   SERLES P, YEO J, HACHÉ M, et al. Ultrahigh specific strength by Bayesian optimization of carbon nanolattices[J]. Advanced Materials, 2025, 37(14): 2410651.
77
]
结合多目标贝叶斯优化与纳米制造,提出了一种用于开发高比强度碳纳米晶格材料的智能设计方法,通过高斯过程建模材料性能并采用期望超体积改进引导结构优化,实现了压缩模量、剪切模量与密度的多目标平衡,最终在保持超轻特性的同时,结构强度提升达118%,杨氏模量提升68%。Alvi等[   ALVI S M A A, JANSSEN J, KHATAMSAZ D, et al. Hierarchical Gaussian process-based Bayesian optimization for materials discovery in high entropy alloy spaces[J]. Acta Materialia, 2025, 289: 120908.
78
]
提出了一种异构深度高斯过程贝叶斯优化方法,针对FeCrNiCoCu高熵合金的多目标优化问题,实现了热膨胀系数的最小化与体积模量的最大化;该方法利用深度高斯过程捕捉热膨胀系数、体积模量与设计参数复杂的非线性相关性,并结合期望超体积改进与置信上限采集策略,实现异构查询,在仅50次迭代内超体积覆盖率便超过95%,显著提升了低热膨胀系数和高体积模量目标的优化效率。

图2     贝叶斯优化流程图
Fig.2     Bayesian optimization flowchart

此外,在航空航天新材料设计中,许多问题具有显著的多目标特性,往往需要在多个相互冲突的性能指标间进行权衡,例如强度与质量、耐腐蚀性与制造成本、热防护能力与结构刚度等。多目标优化方法旨在找到一组非支配解,也被称为帕累托前沿(Pareto front)。在这些解中,每个目标的改善必须以牺牲另一个目标为代价。常见的多目标优化技术包括加权和法、Pareto排序和分解型方法,其中NSGA–Ⅱ[   DEB K, PRATAP A, AGARWAL S, et al. A fast and elitist multiobjective genetic algorithm: NSGA–II[J]. IEEE Transactions on Evolutionary Computation, 2002, 6(2): 182–197.
79
]
(非支配排序遗传算法Ⅱ)和SPEA2[   ZITZLER E, LAUMANNS M, THIELE L. SPEA2: Improving the strength pareto evolutionary algorithm[R/OL]. Computer Engineering and Networks Laboratory of ETH Zurich, 2001. http://hdl.handle.net/20.500.11850/145755.
80
]
(强度Pareto进化算法2)因其良好的全局搜索能力和高效性被广泛应用。Wang等[   WANG J W, ZHOU L L, FAN C Z. A machine learning-based method for co-design and optimization of microwave-absorbing/load-bearing multifunctional structures[J]. Smart Materials and Structures, 2024, 33(4): 045023.
81
]
将NSGA–Ⅱ应用于多层频率选择表面(Frequency-selective surface)材料吸收器的设计中,通过构建非支配等级划分,协同优化其电磁吸收性能与结构承载能力;优化结果显示,该结构在2.5~18.0 GHz范围内实现了超过90%的吸收率,同时具备优异的力学性能(杨氏模量达334.8 MPa,抗压强度达4.95 MPa)。

2.3     LLM驱动的材料开发

近年来,依托自注意力机制和Transformer架构,LLM通过大规模预训练捕捉语言规律和知识表示,在多个领域取得了突破性进展,并迅速渗透至材料研发领域。如图3所示[   YANG J F, JIN H Y, TANG R X, et al. Harnessing the power of LLMs in practice: A survey on ChatGPT and beyond[J]. ACM Transactions on Knowledge Discovery from Data, 2024, 18(6): 1–32.
82
]
,基于Transformer的模型以非灰色显示,其中蓝色分支代表仅解码器模型,粉色分支代表仅编码器模型,绿色分支则表示编码器–解码器模型。图3中模型在时间轴上的垂直位置标示了其发布时间;同时,实心方块代表开源模型,空心方块代表闭源模型;右下角的堆叠条形图进一步展示了来自不同公司和机构的模型数量。这些模型为跨领域信息提取和知识融合提供了强大工具,但其训练通常需要巨大的计算资源。因此,在实际研发中,需要采用低资源适配策略,在提升特定任务性能的同时降低训练和推理成本。

图3     现代LLM的进化树[   YANG J F, JIN H Y, TANG R X, et al. Harnessing the power of LLMs in practice: A survey on ChatGPT and beyond[J]. ACM Transactions on Knowledge Discovery from Data, 2024, 18(6): 1–32.
82
]
Fig.3     Evolutionary tree of modern LLM[   YANG J F, JIN H Y, TANG R X, et al. Harnessing the power of LLMs in practice: A survey on ChatGPT and beyond[J]. ACM Transactions on Knowledge Discovery from Data, 2024, 18(6): 1–32.
82
]

2.3.1     LLM低资源适配策略

提示词是用户输入LLM的文本指令或问题,用于引导模型生成特定类型输出,可以由文本、图像、声音或其他媒体形式组成。提示词工程由3个重复步骤组成[   SCHULHOFF S, LLIE M, BALEPUR N, et al. The prompt report: a systematic survey of prompt engineering techniques[J/OL]. (2024–06–06)[2025–03–15]. https://doi.org/10.48550/arXiv.2406.06608.
83
]
:(1)对数据集进行推理;(2)评估模型性能;(3)修改提示词模板。在提示词工程的基础上,谷歌开发了思维链(Chain of thought)技术[   WEI J, WANG X, SCHUURMANS D, et al. Chain-of-thought prompting elicits reasoning in large language models[J/OL]. (2023–01–10)[2025–03–15]. https://doi.org/10.48550/arXiv.2201.11903.
84
]
,将LLM中隐含的推理步骤转化为明确的、可指导的序列,从而增强了模型产生基于逻辑推理输出的能力(尤其是在复杂问题解决的过程中)。思维树(Tree of thought)[   YAO S, YU D, ZHAO J, et al. Tree of thoughts: Deliberate problem solving with large language models[J/OL]. (2023–12–03) [2025–03–15]. https://doi.org/10.48550/arXiv.2305.10601.
85
]
是另一种提示技术,在每一个分支上都体现了一种替代的推理轨迹,使得LLM能够遍历不同的假设,在LLM展开不同的思路的同时,评估每一条思路的逻辑一致性和任务的相关性。思维树的每个示例都按照问题分解、思维生成、状态评估、搜索算法4个步骤执行。此外用于领域任务的专家提示(Expert prompting)、用于模型自我迭代的反射(Reflection)方法、用于简化复杂多步骤任务的链式(Chains)方法、自动提示工程(Automatic prompt engineering)等提示词技术[   AMATRIAIN X. Prompt design and engineering: Introduction and advanced methods[J/OL]. (2024–05–05)[2025–03–15]. https://doi.org/10.48550/ARXIV.2401.14423.
86
]
也被广泛应用。

参数高效微调[   LIALIN V, DESHPANDE V, YAO X, et al. Scaling down to scale up: A guide to parameter-efficient fine-tuning[J/OL]. (2024–11–22)[2025–03–15]. https://doi.org/10.48550/arXiv.2303.15647.
87
]
(Parameter-efficient fine-tuning,PEFT)是一种旨在高效、低成本微调LLM的技术,主要分为基于添加(Addition-based)、基于选择(Selection-based)和基于重参数化(Reparameterization-based)3类策略,具体的分类见图4。以低秩自适应方法(Low-rank adaptation,LoRA)[   HU E J, SHEN Y, WALLIS P, et al. LoRA: Low-rank adaptation of large language models[J/OL]. (2021–10–16) [2025–03–15]. https://doi.org/10.48550/ARXIV.2106.09685.
88
]
为例,LoRA通过在预训练模型的权重矩阵中引入低秩分解,将权重矩阵W分解为两个低秩矩阵AB,即W=W0W,其中ΔW=A·BW0是预训练模型的初始权重。AB是可训练的低秩矩阵,秩r远小于原始权重矩阵的维度。仅调整少量参数来实现模型微调,而不是更新整个模型的参数。QLoRA[   DETTMERS T, PAGNONI A, HOLTZMAN A, et al. QLoRA: Efficient finetuning of quantized LLMs[J/OL]. (2023–05–23)[2025–03–15]. https://doi.org/10.48550/arXiv.2305.14314.
89
]
在LoRA的基础上进一步引入了量化[   DETTMERS T, LEWIS M, BELKADA Y, et al. LLM.int8(): 8-bit matrix multiplication for transformers at scale[J/OL]. (2022–11–10)[2025–03–15]. https://doi.org/10.48550/ARXIV.2208.07339.
90
]
技术,通过降低模型权重的精度来减少内存占用和计算开销。结合分布式训练[   RAJBHANDARI S, RASLEY J, RUWASE O, et al. ZeRO: Memory optimizations toward training trillion parameter models[C]//SC20: International Conference for High Performance Computing, Networking, Storage and Analysis. New York: IEEE, 2020.
91
]
和混合精度计算[   DETTMERS T, LEWIS M, BELKADA Y, et al. LLM.int8(): 8-bit matrix multiplication for transformers at scale[J/OL]. (2022–11–10)[2025–03–15]. https://doi.org/10.48550/ARXIV.2208.07339.
90
]
,可大幅加速模型并行并减少计算消耗。DeepSeek–R1[   DEEPSEEK-AI, GUO D, YANG D, et al. DeepSeek–R1: Incentivizing reasoning capability in LLMs via reinforcement learning[J/OL]. (2025–01–22)[2025–03–15]. https://doi.org/10.48550/ARXIV.2501.12948.
92
]
采用了混合专家架构,通过动态路由机制仅激活部分模型参数,进一步降低了大模型的开发成本。

图4     参数高效微调方法分类[   LIALIN V, DESHPANDE V, YAO X, et al. Scaling down to scale up: A guide to parameter-efficient fine-tuning[J/OL]. (2024–11–22)[2025–03–15]. https://doi.org/10.48550/arXiv.2303.15647.
87
]
Fig.4     Classification of parameter-efficient fine-tuning methods[   LIALIN V, DESHPANDE V, YAO X, et al. Scaling down to scale up: A guide to parameter-efficient fine-tuning[J/OL]. (2024–11–22)[2025–03–15]. https://doi.org/10.48550/arXiv.2303.15647.
87
]

在面对复杂的科学文献与技术文档时,LLM能够迅速提取其中的关键信息,能为材料研发工作提供及时有力的信息支持。Kang等[   KANG Y, LEE W, BAE T, et al. Harnessing large language models to collect and analyze metal-organic framework property data set[J]. Journal of the American Chemical Society, 2025, 147(5): 3943–3958.
93
]
和Bai等[   BAI X F, HE S, LI Y, et al. Construction of a knowledge graph for framework material enabled by large language models and its application[J]. NPJ Computational Materials, 2025, 11: 51.
94
]
结合LLM的自然语言处理能力,在数以万计的期刊文章的基础上,建立了包含上百万节点与关系的知识图谱,用于LLM的检索增强生成(Retrieval-augmented generation,RAG)[   GAO Y, XIONG Y, GAO X, et al. Retrieval-augmented generation for large language models: A survey[J/OL]. (2024–03–27)[2025–03–15]. https://doi.org/10.48550/arXiv.2312.10997.
95
]
。在知识图谱与LLM集成后,LLM在问答方面准确率达到了91.67%,并且提供了精准的信息来源。此外,在材料研发的跨学科研究进程中,LLM能够有效整合物理学、化学及工程学等不同领域的知识资源,为航空航天材料的设计与创新提供全新的思路与方法,将有力推动航空航天材料的技术进步与发展。

2.3.2     面向材料创新的语言模型智能体

Boiko等[   BOIKO D A, MACKNIGHT R, KLINE B, et al. Autonomous chemical research with large language models[J]. Nature, 2023, 624(7992): 570–578.
96
]
基于GPT–4开发的名为Coscientist的人工智能系统,能够自主设计、规划和执行复杂的科学试验。通过整合互联网搜索、文档查询、代码执行及试验智能化工具,Coscientist成功优化了钯催化的交叉偶联反应,并在多种试验任务中表现出卓越的自主试验设计和执行能力。此外,Meta FAIR实验室联合阿姆斯特丹大学发布了材料生成模型FlowLLM[   SRIRAM A, MILLER B K, CHEN R T Q, et al. FlowLLM: Flow Matching for material generation with large language models as base distributions[C]//Proceedings of NeurIPS 2024. San Diego: NeurIPS Foundation, 2024.
13
]
,该模型使生成稳定材料的效率提升超300%,生成稳定、独特、新颖材料的效率提高约50%。FlowLLM通过LLM生成初始材料表示,然后使用黎曼流匹配模型进行迭代转换更新原子位置和晶格参数,从而实现高效稳定的材料生成。张统一院士团队[   LIU X, SUN P, ZHANG L, et al. Perovskite–LLM: Knowledge-enhanced large language models for perovskite solar cell research[J/OL]. (2025–02–18)[2025–03–15]. https://doi.org/10.48550/arXiv.2502.12669.
14
]
为钙钛矿太阳能电池构建了知识增强型系统,分别是用于领域特定知识辅助的Perovskite Chat LLM和用于科学推理任务的Perovskite Reasoning LLM,为研究人员提供了文献综述、试验设计和钙钛矿太阳能电池研究中复杂问题解决的有效工具。上海人工智能实验室发布的化学语言模型ChemLLM[   ZHANG D, LIU W, TAN Q, et al. ChemLLM: A chemical large language model[J/OL]. (2024–04–25)[2025–03–15]. https://doi.org/10.48550/arXiv.2402.06852.
97
]
,可通过对话交互执行化学学科的各种任务,在核心任务上的性能与GPT–4相当;研究人员将结构化化学知识融入对话系统,为开发各科学领域的LLM树立了新标准。麻省理工学院(MIT)的研究团队开发出AtomAgents[   GHAFAROLLAHI A, BUEHLER M J. AtomAgents: Alloy design and discovery through physics-aware multi-modal multi-agent artificial intelligence[J/OL]. (2024–07–13)[2025–03–15]. https://doi.org/10.48550/ARXIV.2407.10022.
98
]
,这是一个考虑物理法则的生成式AI框架,结合LLM智能及在不同领域中精通的AI代理的合作能力,能够设计出特性优于纯金属的合金,为能源等领域提供了新的材料设计思路。

3     AI算法开发航空航天材料具体示例

3.1     AI助力合金材料开发设计

合金材料凭借优异的力学性能、高温稳定性及耐腐蚀性能,长期以来在航空航天领域发挥着重要作用。然而,传统的合金开发方法存在试错成本高、设计周期长、效率低等问题,AI技术的引入有效地提升了合金材料开发的精准性和效率[   弭光宝, 孙圆治, 吴明宇, 等. 机器学习在航空发动机钛合金研究中的应用进展[J]. 航空制造技术, 2024, 67(1/2): 66–78.MI Guangbao, SUN Yuanzhi, WU Mingyu, et al. Applications of machine learning on aero-engine titanium alloys[J]. Aeronautical Manufacturing Technology, 2024, 67(1/2): 66–78.
99
]

铝合金由于低成本、易于制造和轻质高强特性,是重要的机身结构材料,在航空航天领域应用广泛[   JIANG L, FU H D, ZHANG Z H, et al. Synchronously enhancing the strength, toughness, and stress corrosion resistance of high-end aluminum alloys via interpretable machine learning[J]. Acta Materialia, 2024, 270: 119873.
100
]
。He等[   HE M, LI Y, GHAFFARI B, et al. Machine learning-augmented modeling on the formation of Si–dominated Non–β″early-stage precipitates in Al–Si–Mg alloys with Si supersaturation induced by non-equilibrium solidification[J]. Acta Materialia, 2025, 282: 120454.
101
]
通过ANN研究高冷却速率下形成的Al–Si–Mg合金中,Si富集的早期析出相的形成机制,以及与Si的快速扩散和高初始浓度的关系,从理论层面解释了高压铸造工艺中Si富集相产生的原因[   LIU T, PEI Z R, BARTON D, et al. Characterization of nanostructures in a high pressure die cast Al–Si–Cu alloy[J]. Acta Materialia, 2022, 224: 117500.
102
]
。Juan等[   JUAN Y F, NIU G S, YANG Y, et al. Knowledge-aware design of high-strength aviation aluminum alloys via machine learning[J]. Journal of Materials Research and Technology, 2023, 24: 346–361.
103
]
提出了一种结合机器学习与材料科学基础理论的知识感知设计系统,建立了包含5113个样本的航空铝合金数据库,并筛选出关键特征,利用XGBoost算法与PSO算法成功开发出抗拉强度达到812 MPa、屈服强度792 MPa的新型铝合金。Jiang等[   JIANG L, ZHANG Z H, HU H, et al. A rapid and effective method for alloy materials design via sample data transfer machine learning[J]. NPJ Computational Materials, 2023, 9: 26.
104
]
提出了一种基于样本数据迁移学习的新型合金设计方法,利用现有AA7xxx系列铝合金的工艺数据,通过迁移学习和NSGA–Ⅱ多目标优化,成功优化了E2合金的3阶段时效工艺,实现了合金抗拉强度和延展性的同步大幅提升。Hariharan等[   HARIHARAN A, ACKERMANN M, KOSS S, et al. High-speed 3D printing coupled with machine learning to accelerate alloy development for additive manufacturing[J]. Advanced Science, 2025, 12(17): 2414880.
105
]
通过可解释多项式回归模型[   ACKERMANN M, HAASE C. Machine learning-based identification of interpretable process-structure linkages in metal additive manufacturing[J]. Additive Manufacturing, 2023, 71: 103585.
106
]
,研究了增材制造过程中不同铝浓度下激光功率、扫描速度和路径、粉末混合比例对合金性能的的贡献,在较高铝含量、较低激光功率、层间0°扫描旋转下得到了最高屈服强度。

钛合金因具有高比强度、高温稳定性和耐腐蚀性,在航空航天结构及发动机部件(如涡轮叶片)[   TAKAHASHI R J, DE ASSIS J M K, FAZAN L H, et al. TBC development on Ti–6Al–4V for aerospace application[J]. Coatings, 2025, 15(1): 47.
107
]
中广泛使用。Lee等[   LEE J A, PARK J, SAGONG M J, et al. Active learning framework to optimize process parameters for additive-manufactured Ti-6Al-4V with high strength and ductility[J]. Nature Communications, 2025, 16: 931.
108
]
提出了一种基于主动学习与高斯过程回归的优化框架,通过高效筛选激光粉末床熔融的最佳工艺参数组合,实现了Ti–6Al–4V钛合金性能的显著提升;该方法首先利用高斯过程模型建立工艺参数与机械性能之间的映射关系,再采用期望超体积改进作为采集函数,在每一轮主动学习中推荐两个候选样本进行试验验证;图5(a)、(b)、(c)分别展示了第1次、第3次和第5次迭代的结果,其中黄色菱形表示模型预测的样本点性能,红色三角形表示试验测得的样本点性能;从图5(d)所示的Ti–6Al–4V合金抗拉强度(Ultimate tensile strength,UTS)与延展性(Total elongation,TE)的Ashby图中可以看出,相较于其他加工方法的数据(蓝色三角形代表向能量沉积(Directed energy deposition,DED)工艺样品,绿色五边形则代表锻造工艺样品),该主动学习框架得到的样本点(红色圆点)性能具有更为优异的表现;最终,该方法使Ti–6Al–4V合金的抗拉强度达到了1190 MPa,延伸率达16.5%,突破了传统方法中UTS与TE之间的矛盾。Montalbano等[   MONTALBANO T, NIMER S, DAFFRON M, et al. Machine learning enabled discovery of new L–PBF processing domains for Ti–6Al–4V[J]. Additive Manufacturing, 2025, 98: 104632.
109
]
采用高斯过程回归模型预测增材制造过程中,高激光功率、高扫描速度与小扫描间距等关键工艺参数下材料的密度与力学性能,结合贝叶斯优化策略高效引导试验设计,成功发现了一系列孔隙率体积分数低于1%的高致密材料制备工艺参数组合。

图5     基于主动学习的Ti–6Al–4V合金性能优化过程[   LEE J A, PARK J, SAGONG M J, et al. Active learning framework to optimize process parameters for additive-manufactured Ti-6Al-4V with high strength and ductility[J]. Nature Communications, 2025, 16: 931.
108
]
Fig.5     Performance optimization process of Ti–6Al–4V alloy based on active learning[   LEE J A, PARK J, SAGONG M J, et al. Active learning framework to optimize process parameters for additive-manufactured Ti-6Al-4V with high strength and ductility[J]. Nature Communications, 2025, 16: 931.
108
]

镁合金因密度低、比强度高,成为航空航天轻量化的重要候选材料。然而,其机械性能和耐腐蚀性相对不足,限制了镁合金在航空领域的应用范围[   BAI J Y, YANG Y, WEN C, et al. Applications of magnesium alloys for aerospace: A review[J]. Journal of Magnesium and Alloys, 2023, 11(10): 3609–3619.
110
]
。Wang等[   WANG Q H, QIN X, XIA S, et al. Interpretable machine learning excavates a low-alloyed magnesium alloy with strength-ductility synergy based on data augmentation and reconstruction[J]. Journal of Magnesium and Alloys, 2025, 13(6): 2866–2883.
111
]
利用数据增强和可解释机器学习方法,通过梯度提升模型与相关性分析,成功开发出抗拉强度超过380 MPa、延伸率达18%的低成本镁合金体系(Mg–Al–Ca–Mn–(Zn))。Liu等[   LIU Z Y, WANG T Y, JIN L, et al. Towards high stiffness and ductility—The Mg–Al–Y alloy design through machine learning[J]. Journal of Materials Science & Technology, 2025, 221: 194–203.
112
]
利用XGBoost与图像识别技术,定量提取了镁铝钇合金中增强相的体积分数、尺寸、形状和分散状态等6个关键参数,从而精准预测并优化合金性能;设计出的AW917合金实现了高杨氏模量(51.5 GPa)和优良的延伸率(7%),还展现了均匀的增强相分布和良好的界面匹配,有效平衡了刚性与延展性。Xue等[   XUE D, WEI W, SHI W, et al. Optimization of stabilized annealing of Al–Mg alloys utilizing machine learning algorithms[J]. Materials Today Communications, 2023, 35: 106177.
113
]
使用ANN优化Al–Mg合金的退火过程,揭示了镁含量、变形程度、退火温度和时间这4个因素对合金耐腐蚀性能的决定性影响;所构建的模型精准预测了合金的腐蚀行为,其平均绝对误差为0.195。

3.2     AI助力复合材料开发设计

复合材料具有高比强度、高比模量及优于传统合金材料的抗腐蚀和抗疲劳性能,在航空航天领域得到广泛应用,如空客A380和波音787飞机中复合材料的使用比例已分别超过25%和50%。近年来,AI算法的应用显著提高了复合材料性能优化和设计效率[   WANG Y F, WANG K, ZHANG C. Applications of artificial intelligence/machine learning to high-performance composites[J]. Composites Part B: Engineering, 2024, 285: 111740.
114
]

陶瓷基复合材料具备卓越的高温稳定性与耐烧蚀性能,广泛应用于航空航天热防护结构中,但传统试错方法效率低下且成本高昂[   KAUFMANN K, MARYANOVSKY D, MELLOR W M, et al. Discovery of high-entropy ceramics via machine learning[J]. NPJ Computational Materials, 2020, 6: 42.
115
]
。为此,Xiao等[   XIAO J, GUO W J, YANG J G, et al. Analysis and regularity of ablation resistance performance of ultra-high temperature ceramic matrix composites using data-driven strategy[J]. Ceramics International, 2024, 50(18): 31937–31945.
116
]
采用Pearson相关系数(图6(a))等评估方法对初始特征进行降维处理,最终筛选出5个对线性烧蚀率(Linear ablation rate,LAR)影响最大的关键特征,分别为陶瓷平均熔点、材料热导率、氧化物热膨胀系数、材料制备温度及烧蚀时间;基于这些特征构建了随机森林(Random forest,RF)模型,实现了对超高温陶瓷基复合材料LAR的高效精准预测,并结合符号回归方法(图6(b)),进一步得出了主要特征与LAR之间的显式数学表达式。Borkowski等[   BORKOWSKI L, SKINNER T, CHATTOPADHYAY A. Woven ceramic matrix composite surrogate model based on physics-informed recurrent neural network[J]. Composite Structures, 2023, 305: 116455.
117
]
利用物理信息约束的循环神经网络作为代理模型,通过整合多尺度广义单元法与基质损伤模型生成的训练数据,精确预测了编织陶瓷基复合材料在非单调加载下的应力–应变响应及切线模量,大幅提升计算效率,为大规模结构分析提供了高效的替代方案。Zhou等[   ZHOU J, LI L, LU L, et al. Machine learning-based quality optimisation of ceramic extrusion 3D printing deposition lines[J]. Materials Today Communications, 2024, 41: 110841.
118
]
提出了一种融合机器学习与视觉反馈的陶瓷挤出型3D打印质量优化方法;通过图像处理提取沉积线宽度,并建立它与4个关键工艺参数(螺杆转速、打印速度、喷嘴直径与打印高度)之间的映射关系;通过集成学习模型实现对沉积线宽度的预测与质量分类,获得了优异的预测性能(R²>0.99),为实现打印过程的稳定控制与缺陷抑制提供了有效手段。

图6     线性烧蚀率(LAR)与关键特征间的关系分析[   XIAO J, GUO W J, YANG J G, et al. Analysis and regularity of ablation resistance performance of ultra-high temperature ceramic matrix composites using data-driven strategy[J]. Ceramics International, 2024, 50(18): 31937–31945.
116
]
Fig.6     Correlation analysis between the linear ablation rate (LAR) and key features[   XIAO J, GUO W J, YANG J G, et al. Analysis and regularity of ablation resistance performance of ultra-high temperature ceramic matrix composites using data-driven strategy[J]. Ceramics International, 2024, 50(18): 31937–31945.
116
]

金属基复合材料具有高屈服强度、良好的断裂韧性、低热膨胀性和优异的耐磨性,在航空领域被广泛应用于机体结构(如起落架)和推进系统(如发动机导向叶片、液压歧管等)[   KACZMAR J W, PIETRZAK K, WŁOSIŃSKI W. The production and application of metal matrix composite materials[J]. Journal of Materials Processing Technology, 2000, 106(1–3): 58–67.
119
]
。但增强颗粒的加入通常会引起材料韧性下降,导致强度与韧性难以兼顾。为此,Zhong等[   ZHONG Z Y, AN J, WU D, et al. A machine learning strategy for enhancing the strength and toughness in metal matrix composites[J]. International Journal of Mechanical Sciences, 2024, 281: 109550.
120
]
提出一种多结构、多目标协同优化策略,利用RF模型与非支配排序遗传算法,构建覆盖叠层、网络、微纳混合和聚集结构的综合数据库,结合机器学习进行正向预测及遗传算法实现反向优化,有效实现了金属基复合材料强韧性的协同提升。Hasan等[   HASAN M S, WONG T, ROHATGI P K, et al. Analysis of the friction and wear of graphene reinforced aluminum metal matrix composites using machine learning models[J]. Tribology International, 2022, 170: 107527.
121
]
利用RF等多种机器学习模型,基于15个关键变量精准预测了石墨烯增强铝基复合材料在不同工况下的摩擦系数与磨损率。

聚合物基复合材料因高比强度、高比模量而广泛应用于航空航天结构,特别是碳纤维增强聚合物(Carbon fiber reinforced polymer,CFRP)在飞机副翼、襟翼和起落架舱门等结构中应用广泛,波音777和787飞机中CFRP用量已高达50%[   PIMENTA S, PINHO S T. Recycling carbon fibre reinforced polymers for structural applications: Technology review and market outlook[J]. Waste Management, 2011, 31(2): 378–392.
122
]
。Marr等[   MARR J, ZARTMANN L, REINEL-BITZER D, et al. Data-based prediction of the viscoelastic behavior of short fiber reinforced composites[J]. PAMM, 2023, 22(1): e202200085.
123
]
基于传统的注塑成型工艺,提出一种基于机器学习与自适应采样策略的数据驱动方法,以纤维体积分数、取向分布和弹性参数等微观结构特征为基础,构建了黏弹性性能的高精度代理模型;通过Kriging插值与全局优化技术实现了短纤维增强复合材料黏弹性行为的精确预测与结构优化,有效降低了计算成本。此外,AI与3D打印技术的深度融合,为轻质高性能复合材料开发带来了新思路[   LI M Z, ZHANG H W, LI S R, et al. Machine learning and materials informatics approaches for predicting transverse mechanical properties of unidirectional CFRP composites with microvoids[J]. Materials & Design, 2022, 224: 111340.
124
]
。Thomas等[   THOMAS A J, BAROCIO E, PIPES R B. A machine learning approach to determine the elastic properties of printed fiber-reinforced polymers[J]. Composites Science and Technology, 2022, 220: 109293.
125
]
提出了一种基于SVR的逆向推断方法,通过有限的拉伸试验数据,高效准确地预测了3D打印短纤维增强聚合物的弹性性能;该研究结合微观力学模型,利用网格搜索和交叉验证优化模型超参数,有效推断出纤维取向状态与聚合物基体性能,大幅降低了试验成本和复杂性。连续纤维增强聚合物复合材料的增材制造过程中,参数设置不当将引发纤维错位和磨损等缺陷的问题,Lu等[   LU L, HOU J, YUAN S Q, et al. Deep learning-assisted real-time defect detection and closed-loop adjustment for additive manufacturing of continuous fiber-reinforced polymer composites[J]. Robotics and Computer-Integrated Manufacturing, 2023, 79: 102431.
126
]
针对此提出了一种基于深度学习的实时缺陷检测与闭环参数优化方法,该方法采用YOLOv4算法进行实时缺陷检测,实现了90.42%的高检测精度,通过缺陷识别与实时反馈,动态优化挤出温度、打印速度、层厚和纤维供给速率等关键工艺参数,有效减少甚至消除了制造过程中的缺陷。Liu等[   LIU T, ZHANG T, CHEN Y, et al. Neural co-optimization of structural topology, manufacturable layers, and path orientations for fiber-reinforced composites[J/OL]. (2025–04–30)[2025–03–15]. https://doi.org/10.48550/arXiv.2505.03779.
127
]
提出了一种基于神经网络的并行拓扑优化框架,能够同步优化纤维增强复合材料的结构拓扑、可制造的打印层及纤维路径方向,从而显著提升结构强度并保障制造可行性;该方法引入3个隐式神经场,分别表示材料分布、材料沉积顺序和纤维排布方向,并将力学性能目标与多种增材制造中涉及的制造工艺约束(包括层曲率、路径曲率、层厚控制及打印方向限制)统一建模为可导的损失函数,构建了适用于多轴3D打印系统的可微分联合优化流程,相较于串行的拓扑优化方式,其最大失效载荷提升可达33.1%。

3.3     AI助力超材料开发设计

超材料是一类通过精细设计并重复排列特殊“单元结构”而获得的材料,其电磁、光学和机械等性能超越原有材料本身的性能[   SONG J, LEE J, KIM N, et al. Artificial intelligence in the design of innovative metamaterials: A comprehensive review[J]. International Journal of Precision Engineering and Manufacturing, 2024, 25(1): 225–244.
128
]
。近年来,得益于AI与增材制造技术的发展,超材料的设计与制备效率大幅提升。在航空航天领域,超材料被广泛应用于雷达隐身、振动与噪声控制及结构轻量化,通过调控飞行器表面或内部的微结构,优化气动性能和能量吸收能力,从而提高整体安全性与效率。

力学超材料[   NIAN Y Z, WAN S, AVCAR M, et al. Nature-inspired 3D printing-based double-graded aerospace negative Poisson’s ratio metastructure: Design, fabrication, investigation, optimization[J]. Composite Structures, 2024, 348: 118482.
129
]
是一类通过精细设计微结构单元而展现出传统材料所不具备的特殊静态与动态力学性能的材料[   WANG H F, ZHANG C, WANG C, et al. The application of negative Poisson’s ratio metamaterials in the optimization of a variable area wing[J]. Aerospace, 2025, 12(2): 125.
130
]
。例如,负泊松比材料在受拉时横向膨胀,表现出独特的变形机制。Wang等[   WANG M H, SUN S, ZHANG T Y. Machine learning accelerated design of auxetic structures[J]. Materials & Design, 2023, 234: 112334.
131
]
利用AI技术提出了一种基于机器学习的负泊松比超材料设计方法;首先采用二进制编码将拉胀结构中的re-entrant单元转换为由0和1构成的基因序列,精准捕捉微结构几何特征;随后通过有限元仿真获得初始数据,并利用进化计算进行全局搜索,不断优化结构以达到最大负泊松比;为进一步提升设计效率,研究人员引入机器学习代理模型来将结构基因编码直接映射为负泊松比预测值(图7(a))以加速结构性能预测;通过对比SVR、RF、ANN等多种机器学习算法在测试集上的R²表现(图7(b)),最终选定预测精度最高的ANN模型;相较传统FEM方法,其计算速度提升超过5个数量级,显著提高了超材料结构的设计效率。此外,Zhao等[   ZHAO M, LI X W, YAN X, et al. Machine learning accelerated design of lattice metamaterials for customizable energy absorption[J]. Thin-Walled Structures, 2025, 208: 112845.
132
]
利用CNN加速设计方法,通过伪随机生成条杆几何图案和有限元模拟数据训练CNN,实现了逆向设计具有定制能量吸收性能的晶格超材料,得到的结构中Model–A的比能吸收性能比Model–B提高超267.2%,且应力分布更均匀,承载能力更强。

图7     机器学习作为代理模型设计负泊松比结构[   WANG M H, SUN S, ZHANG T Y. Machine learning accelerated design of auxetic structures[J]. Materials & Design, 2023, 234: 112334.
131
]
Fig.7     Machine learning as a surrogate model for designing auxetic structures[   WANG M H, SUN S, ZHANG T Y. Machine learning accelerated design of auxetic structures[J]. Materials & Design, 2023, 234: 112334.
131
]

声学超材料通过设计亚声波长尺度的微结构实现卓越的声波调控,主要用于航空航天领域的噪声与振动控制及轻量化设计,有效抑制发动机的气动噪音及结构振动,从而提升飞行器安全性和舒适性[   FAN J X, SONG B, ZHANG L, et al. Structural design and additive manufacturing of multifunctional metamaterials with low-frequency sound absorption and load-bearing performances[J]. International Journal of Mechanical Sciences, 2023, 238: 107848.
133
]
。Zhang等[   ZHANG H J, LIU J W, MA W T, et al. Learning to inversely design acoustic metamaterials for enhanced performance[J]. Acta Mechanica Sinica, 2023, 39(7): 722426.
134
]
提出的基于机器学习的逆向设计框架,通过正向网络捕捉超材料结构与声吸收性能之间的非线性关系,并利用反向网络根据目标吸声谱反求最佳结构设计;该方法采用“超限探索”技术,突破训练数据参数范围,使逆向设计精度从9.2%提高至99.6%,显著提升了设计效率和性能。

电磁超材料通过精确控制微结构单元的参数,实现对电磁波传播、反射和吸收的高效调控,从而具备宽带吸收、低反射乃至电磁隐身等功能,满足航空航天领域对雷达隐身和电磁干扰控制的需求[   FENG M F, ZHANG K F, XIAO J J, et al. Material-structure collaborative design for broadband microwave absorption metamaterial with low density and thin thickness[J]. Composites Part B: Engineering, 2023, 263: 110862.
135
]
。Lim等[   LIM D D, IBARRA A, LEE J, et al. A tunable metamaterial microwave absorber inspired by chameleon’s color-changing mechanism[J]. Science Advances, 2025, 11(3): eads3499.
136
]
借助AI构建了一套数据驱动的设计流程,将有限元分析与机器学习代理模型及遗传算法相结合,实现了超材料结构的快速优化;该流程首先利用FEA模拟大量设计参数下的电磁响应数据,再通过神经网络训练出高精度代理模型,实现ms级性能预测,显著降低计算成本;随后,遗传算法在代理模型的辅助下全局搜索设计空间,筛选出满足目标性能(如在4~18 GHz频段内吸收率超90%)的最佳结构参数。

4     结论与展望

本文系统阐述了人工智能在航空航天新材料设计中的多维应用,涵盖了从AI辅助的计算模拟与智能化试验,到基于机器学习和深度学习的代理模型优化设计,再到LLM驱动的智能研发方法,并介绍了AI在航空航天新材料开发中的具体应用。通过多尺度AI模拟(涵盖量子、原子和宏观尺度),实现更高精度计算和高效的数据积累;AI辅助的智能化试验依托自动化机器人技术,有望实现自主材料研发;而代理模型与优化算法的结合则在降低研发成本、缩短设计周期方面展现出显著优势。此外,LLM为跨学科知识整合提供了新思路,有助于突破传统思维局限,成为航空航天材料创新设计的有力工具。

尽管当前人工智能在航空航天材料设计中已取得显著进展,但仍面临一些亟待解决的问题和挑战。(1)由于不同研究团队在数据采集、处理方法及标准上的差异,现有数据在一致性和质量上存在较大差异,急需建立统一的数据标准和规范;(2)当前的AI模型虽然在预测精度方面表现优异,但在处理未见过的设计空间或不同材料体系时,其泛化能力和可解释性仍不足,这在一定程度上限制了它在实际工程中的广泛应用;(3)计算模拟与试验数据之间的深度融合不足,导致模型预测结果与实际应用之间存在一定偏差;(4)高通量实验室的智能化水平尚需提升,以达到更快速地获取高质量试验数据的目标。

针对上述挑战,未来的研究可从以下4个方向深入探索。(1)建立标准化、多尺度的综合材料数据库,促进数据资源的深度融合和广泛共享,为高效精准的AI应用提供坚实的数据基础;(2)开发具有更高泛化能力和可解释性的AI算法,通过算法创新提升模型在不同材料体系中的适用性和工程应用性能;(3)大力发展自动化与智能化试验技术,实现其过程的自主优化和实时反馈,从而进一步提升试验效率;(4)探索机器学习代理模型与新兴大模型相结合的混合优化策略,充分发挥各类模型的优势,拓展航空航天材料设计方法的适用范围。

通过上述举措,有望实现航空航天材料设计领域的数字化、自动化和智能化转型,极大推动航空航天技术的快速发展,为未来航空航天产业提供坚实的技术支撑。

作者介绍



孙升 研究员,博士生导师,研究方向为力学信息学和材料信息学。

参考文献

[1]

ZHU L, LI N, CHILDS P R N. Light-weighting in aerospace component and system design[J]. Propulsion and Power Research, 2018, 7(2): 103119.

[2]

PETERS A B, ZHANG D J, CHEN S, et al. Materials design for hypersonics[J]. Nature Communications, 2024, 15: 3328.

[3]

SONI R, VERMA R, KUMAR GARG R, et al. Progress in aerospace materials and ablation resistant Coatings: A focused review[J]. Optics & Laser Technology, 2024, 177: 111160.

[4]

YE Y F, WANG Q, LU J, et al. High-entropy alloy: Challenges and prospects[J]. Materials Today, 2016, 19(6): 349362.

[5]

LIU G, XIONG Y, ZHOU L M. Additive manufacturing of continuous fiber reinforced polymer composites: Design opportunities and novel applications[J]. Composites Communications, 2021, 27: 100907.

[6]

LI J C, YE H L, DONG Y J, et al. An efficient deep learning-based topology optimization method for continuous fiber composite structure[J]. Acta Mechanica Sinica, 2024, 41(4): 424207.

[7]

WANG H, ZHANG L F, HAN J Q, et al. DeePMD–kit: A deep learning package for many-body potential energy representation and molecular dynamics[J]. Computer Physics Communications, 2018, 228: 178184.

[8]

RACCUGLIA P, ELBERT K C, ADLER P D F, et al. Machine-learning-assisted materials discovery using failed experiments[J]. Nature, 2016, 533: 7376.

[9]

SZYMANSKI N J, RENDY B, FEI Y, et al. An autonomous laboratory for the accelerated synthesis of novel materials[J]. Nature, 2023, 624: 8691.

[10]

CHEN Q B, HE Z J, ZHAO Y, et al. Stacking ensemble learning assisted design of Al–Nb–Ti–V–Zr lightweight high-entropy alloys with high hardness[J]. Materials & Design, 2024, 246: 113363.

[11]

YAN Y G, LIAO Y L, WANG K. Accelerated discovery of oxidation-resistant ultra-high temperature ceramics via data driven methodology[J]. Corrosion Science, 2023, 223: 111457.

[12]

ON H I, JEONG L, SEO T M, et al. Novel method of performance-optimized metastructure design for electromagnetic wave absorption in specific band using deep learning[J]. Engineering Applications of Artificial Intelligence, 2024, 137: 109274.

[13]

SRIRAM A, MILLER B K, CHEN R T Q, et al. FlowLLM: Flow Matching for material generation with large language models as base distributions[C]//Proceedings of NeurIPS 2024. San Diego: NeurIPS Foundation, 2024.

[14]

LIU X, SUN P, ZHANG L, et al. Perovskite–LLM: Knowledge-enhanced large language models for perovskite solar cell research[J/OL]. (2025–02–18)[2025–03–15]. https://doi.org/10.48550/arXiv.2502.12669.

[15]

JAIN A, ONG S P, HAUTIER G, et al. Commentary: The Materials Project: A materials genome approach to accelerating materials innovation[J]. APL Materials, 2013, 1(1): 011002.

[16]

SCHEIDGEN M, HIMANEN L, LADINES A N, et al. NOMAD: A distributed web-based platform for managingmaterials science research data[J]. Journal of Open Source Software, 2023, 8(90): 5388.

[17]

HJORTH LARSEN A, JØRGEN MORTENSEN J, BLOMQVIST J, et al. The atomic simulation environment—A Python library for working with atoms[J]. Journal of Physics: Condensed Matter, 2017, 29(27): 273002.

[18]

ONG S P, RICHARDS W D, JAIN A, et al. Python materials genomics (PYMATGEN): A robust, open-source Python library for materials analysis[J]. Computational Materials Science, 2013, 68: 314319.

[19]

TANG Z C, LI H, LIN P Z, et al. A deep equivariant neural network approach for efficient hybrid density functional calculations[J]. Nature Communications, 2024, 15: 8815.

[20]

WANG Y X, LI Y, TANG Z C, et al. Universal materials model of deep-learning density functional theory Hamiltonian[J]. Science Bulletin, 2024, 69(16): 25142521.

[21]

WANG J Q, LIU J P, WANG H S, et al. A comprehensive transformer-based approach for high-accuracy gas adsorption predictions in metal-organic frameworks[J]. Nature Communications, 2024, 15: 1904.

[22]

BARROSO-LUQUE L, SHUABI M, FU X, et al. Open materials 2024 (OMat24) inorganic materials dataset and models[J/OL]. (2024–10–16)[2025–03–15]. https://doi.org/10.48550/arXiv.2410.12771.

[23]

YANG H, HU C, ZHOU Y, et al. MatterSim: A deep learning atomistic model across elements, temperatures and pressures[J/OL]. (2024–05–08)[2025–03–15]. https://doi.org/10.48550/arXiv.2405.04967.

[24]

SCHWERDTFEGER P, WALES D J. 100 years of the lennard-Jones potential[J]. Journal of Chemical Theory and Computation, 2024, 20(9): 33793405.

[25]

VAN DUIN A C T, DASGUPTA S, LORANT F, et al. ReaxFF: A reactive force field for hydrocarbons[J]. Journal of Physical Chemistry A, 2001, 105(41): 93969409.

[26]

HALE L. NIST interatomic potentials repository[EB/OL]. (2016–02–25)[2025–03–15]. https://doi.org/10.18434/M37.

[27]

BATATIA I, KOVÁCS D P, SIMM G N C, et al. MACE: Higher order equivariant message passing neural networks for fast and accurate force fields[C]//Proceedings of NeurIPS 2022. San Diego: NeurIPS Foundation, 2022.

[28]

GASTEIGER J, BECKER F, GÜNNEMANN S. GemNet: Universal directional graph neural networks for molecules[J]. Advances in Neural Information Processing Systems, 2021, 34: 67906802.

[29]

PELLEGRINI F, LOT R, SHAIDU Y, et al. PANNA 2.0: Efficient neural network interatomic potentials and new architectures[J]. Journal of Chemical Physics, 2023, 159(8): 084117.

[30]

XIE F K, LU T L, MENG S, et al. GPTFF: A high-accuracy out-of-the-box universal AI force field for arbitrary inorganic materials[J]. Science Bulletin, 2024, 69(22): 35253532.

[31]

LI L, SHAO Q, YANG Y C, et al. A database construction method for data-driven computational mechanics of composites[J]. International Journal of Mechanical Sciences, 2023, 249: 108232.

[32]

LIU Z L, FLEMING M, LIU W K. Microstructural material database for self-consistent clustering analysis of elastoplastic strain softening materials[J]. Computer Methods in Applied Mechanics and Engineering, 2018, 330: 547577.

[33]

KOHAR C P, GREVE L, ELLER T K, et al. A machine learning framework for accelerating the design process using CAE simulations: An application to finite element analysis in structural crashworthiness[J]. Computer Methods in Applied Mechanics and Engineering, 2021, 385: 114008.

[34]

KIM S, SHIN H. Deep learning framework for multiscale finite element analysis based on data-driven mechanics and data augmentation[J]. Computer Methods in Applied Mechanics and Engineering, 2023, 414: 116131.

[35]

ANGELOPOULOS A, CAHOON J F, ALTEROVITZ R. Transforming science labs into automated factories of discovery[J]. Science Robotics, 2024, 9(95): eadm6991.

[36]

STRIETH-KALTHOFF F, HAO H, RATHORE V, et al. Delocalized, asynchronous, closed-loop discovery of organic laser emitters[J]. Science, 2024, 384(6697): eadk9227.

[37]

BENNETT J A, OROUJI N, KHAN M, et al. Autonomous reaction Pareto-front mapping with a self-driving catalysis laboratory[J]. Nature Chemical Engineering, 2024, 1(3): 240250.

[38]

DAI T W, VIJAYAKRISHNAN S, SZCZYPIŃSKI F T, et al. Autonomous mobile robots for exploratory synthetic chemistry[J]. Nature, 2024, 635: 890897.

[39]

SONG T, LUO M, ZHANG X L, et al. A multiagent-driven robotic AI chemist enabling autonomous chemical research on demand[J]. Journal of the American Chemical Society, 2025, 147(15): 1253412545.

[40]

MACLEOD B P, PARLANE F L, MORRISSEY T D, et al. Self-driving laboratory for accelerated discovery of thin-film materials[J]. Science Advances, 2020, 6(20): eaaz8867.

[41]

CURRAN R, RAGHUNATHAN S, PRICE M. Review of aerospace engineering cost modelling: The genetic causal approach[J]. Progress in Aerospace Sciences, 2004, 40(8): 487534.

[42]

AYKAN M, ÇELIK M. Vibration fatigue analysis and multi-axial effect in testing of aerospace structures[J]. Mechanical Systems and Signal Processing, 2009, 23(3): 897907.

[43]

KROLL N, ABU-ZURAYK M, DIMITROV D, et al. DLR project digital–X: Towards virtual aircraft design and flight testing based on high-fidelity methods[J]. CEAS Aeronautical Journal, 2016, 7(1): 327.

[44]

CATTAFESTA L, BAHR C, MATHEW J. Fundamentals of wind-tunnel design[M]//Encyclopedia of Aerospace Engineering. Hoboken: John Wiley & Sons Inc., 2010.

[45]

AZARHOOSH Z, ILCHI GHAZAAN M. A review of recent advances in surrogate models for uncertainty quantification of high-dimensional engineering applications[J]. Computer Methods in Applied Mechanics and Engineering, 2025, 433: 117508.

[46]

FORRESTER A I J, SÓBESTER A, KEANE A J. Engineering design via surrogate modelling: A practical guide[M]. Hoboken: John Wiley & Sons Inc., 2008.

[47]

SAMADIAN D, MUHIT I B, DAWOOD N. Application of data-driven surrogate models in structural engineering: A literature review[J]. Archives of Computational Methods in Engineering, 2025, 32(2): 735784.

[48]

KOCHENDERFER M J, WHEELER T A. Algorithms for optimization[M]. Cambridge: The MIT press, 2019.

[49]

KHURI A I, MUKHOPADHYAY S. Response surface methodology[J]. WIREs Computational Statistics, 2010, 2(2): 128149.

[50]

FERDOSIAN I, CAMÕES A. Eco-efficient ultra-high performance concrete development by means of response surface methodology[J]. Cement and Concrete Composites, 2017, 84: 146156.

[51]

RASMUSSEN C E, WILLIAMS C K I. Gaussian processes for machine learning[M]. Cambridge: The MIT press, 2006.

[52]

KAYMAZ I. Application of Kriging method to structural reliability problems[J]. Structural Safety, 2005, 27(2): 133151.

[53]

ZHOU C C, LI C, ZHANG H L, et al. Reliability and sensitivity analysis of composite structures by an adaptive Kriging based approach[J]. Composite Structures, 2021, 278: 114682.

[54]

GUTMANN H M. A radial basis function method for global optimization[J]. Journal of Global Optimization, 2001, 19(3): 201227.

[55]

SUN G Y, LI G Y, GONG Z H, et al. Radial basis functional model for multi-objective sheet metal forming optimization[J]. Engineering Optimization, 2011, 43(12): 13511366.

[56]

SMOLA A J, SCHÖLKOPF B. A tutorial on support vector regression[J]. Statistics and Computing, 2004, 14(3): 199222.

[57]

ZHAI X Y, CHEN M T. Machine learning aided design of Bi2WO6/MIL–53(Al) nanocomposites[J]. Computational Materials Science, 2024, 233: 112737.

[58]

WU Y C, FENG J W. Development and application of artificial neural network[J]. Wireless Personal Communications, 2018, 102(2): 16451656.

[59]

O’SHEA K, NASH R. An Introduction to Convolutional Neural Networks[J/OL]. (2015–11–26)[2025–03–15]. https://doi.org/10.48550/arXiv.1511.08458 (2015).

[60]

XIE T, GROSSMAN J C. Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties[J]. Physical Review Letters, 2018, 120(14): 145301.

[61]

CRESWELL A, WHITE T, DUMOULIN V, et al. Generative adversarial networks: An overview[J]. IEEE Signal Processing Magazine, 2018, 35(1): 5365.

[62]

QIAN C, TAN R K, YE W J. Design of architectured composite materials with an efficient, adaptive artificial neural network-based generative design method[J]. Acta Materialia, 2022, 225: 117548.

[63]

POPEL M, BOJAR O. Training tips for the transformer model[J]. The Prague Bulletin of Mathematical Linguistics, 2018, 110(1): 4370.

[64]

CHEN W, GAO Y, LI Y Y, et al. Broadband solar metamaterial absorbers empowered by transformer-based deep learning[J]. Advanced Science, 2023, 10(13): 2206718.

[65]

FUHG J N, FAU A, NACKENHORST U. State-of-the-art and comparative review of adaptive sampling methods for Kriging[J]. Archives of Computational Methods in Engineering, 2021, 28(4): 26892747.

[66]

FANG K-T, LIN D K J, WINKER P, et al. Uniform design: Theory and application[J]. Technometrics, 2000, 42(3): 237248.

[67]

MORRIS M D, MITCHELL T J. Exploratory designs for computational experiments[J]. Journal of Statistical Planning and Inference, 1995, 43(3): 381402.

[68]

MCKAY M D, BECKMAN R J, CONOVER W J. A comparison of three methods for selecting values of input variables in the analysis of output from a computer code[J]. Technometrics, 2000, 42(1): 5561.

[69]

BABALOLA A E, OJOKOH B A, ODILI J B. A review of population-based optimization algorithms[C]//Proceedings of 2020 International Conference in Mathematics, Computer Engineering and Computer Science (ICMCECS). New York: IEEE, 2020.

[70]

LAMBORA A, GUPTA K, CHOPRA K. Genetic algorithm- a literature review[C]//Proceedings of 2019 International Conference on Machine Learning, Big Data, Cloud and Parallel Computing (COMITCon). New York: IEEE, 2019.

[71]

DAS S, SUGANTHAN P N. Differential evolution: A survey of the state-of-the-art[J]. IEEE Transactions on Evolutionary Computation, 2011, 15(1): 431.

[72]

WANG D S, TAN D P, LIU L. Particle swarm optimization algorithm: An overview[J]. Soft Computing, 2018, 22(2): 387408.

[73]

DORIGO M, BIRATTARI M, STUTZLE T. Ant colony optimization[J]. IEEE Computational Intelligence Magazine, 2006, 1(4): 2839.

[74]

KNEIDING H, BALCELLS D. Augmenting genetic algorithms with machine learning for inverse molecular design[J]. Chemical Science, 2024, 15(38): 1552215539.

[75]

JIN P, YANG S, XU L J, et al. Particle swarm optimization for realizing bilayer thermal sensors with bulk isotropic materials[J]. International Journal of Heat and Mass Transfer, 2021, 172: 121177.

[76]

SHAHRIARI B, SWERSKY K, WANG Z Y, et al. Taking the human out of the loop: A review of Bayesian optimization[J]. Proceedings of the IEEE, 2016, 104(1): 148175.

[77]

SERLES P, YEO J, HACHÉ M, et al. Ultrahigh specific strength by Bayesian optimization of carbon nanolattices[J]. Advanced Materials, 2025, 37(14): 2410651.

[78]

ALVI S M A A, JANSSEN J, KHATAMSAZ D, et al. Hierarchical Gaussian process-based Bayesian optimization for materials discovery in high entropy alloy spaces[J]. Acta Materialia, 2025, 289: 120908.

[79]

DEB K, PRATAP A, AGARWAL S, et al. A fast and elitist multiobjective genetic algorithm: NSGA–II[J]. IEEE Transactions on Evolutionary Computation, 2002, 6(2): 182197.

[80]

ZITZLER E, LAUMANNS M, THIELE L. SPEA2: Improving the strength pareto evolutionary algorithm[R/OL]. Computer Engineering and Networks Laboratory of ETH Zurich, 2001. http://hdl.handle.net/20.500.11850/145755.

[81]

WANG J W, ZHOU L L, FAN C Z. A machine learning-based method for co-design and optimization of microwave-absorbing/load-bearing multifunctional structures[J]. Smart Materials and Structures, 2024, 33(4): 045023.

[82]

YANG J F, JIN H Y, TANG R X, et al. Harnessing the power of LLMs in practice: A survey on ChatGPT and beyond[J]. ACM Transactions on Knowledge Discovery from Data, 2024, 18(6): 132.

[83]

SCHULHOFF S, LLIE M, BALEPUR N, et al. The prompt report: a systematic survey of prompt engineering techniques[J/OL]. (2024–06–06)[2025–03–15]. https://doi.org/10.48550/arXiv.2406.06608.

[84]

WEI J, WANG X, SCHUURMANS D, et al. Chain-of-thought prompting elicits reasoning in large language models[J/OL]. (2023–01–10)[2025–03–15]. https://doi.org/10.48550/arXiv.2201.11903.

[85]

YAO S, YU D, ZHAO J, et al. Tree of thoughts: Deliberate problem solving with large language models[J/OL]. (2023–12–03) [2025–03–15]. https://doi.org/10.48550/arXiv.2305.10601.

[86]

AMATRIAIN X. Prompt design and engineering: Introduction and advanced methods[J/OL]. (2024–05–05)[2025–03–15]. https://doi.org/10.48550/ARXIV.2401.14423.

[87]

LIALIN V, DESHPANDE V, YAO X, et al. Scaling down to scale up: A guide to parameter-efficient fine-tuning[J/OL]. (2024–11–22)[2025–03–15]. https://doi.org/10.48550/arXiv.2303.15647.

[88]

HU E J, SHEN Y, WALLIS P, et al. LoRA: Low-rank adaptation of large language models[J/OL]. (2021–10–16) [2025–03–15]. https://doi.org/10.48550/ARXIV.2106.09685.

[89]

DETTMERS T, PAGNONI A, HOLTZMAN A, et al. QLoRA: Efficient finetuning of quantized LLMs[J/OL]. (2023–05–23)[2025–03–15]. https://doi.org/10.48550/arXiv.2305.14314.

[90]

DETTMERS T, LEWIS M, BELKADA Y, et al. LLM.int8(): 8-bit matrix multiplication for transformers at scale[J/OL]. (2022–11–10)[2025–03–15]. https://doi.org/10.48550/ARXIV.2208.07339.

[91]

RAJBHANDARI S, RASLEY J, RUWASE O, et al. ZeRO: Memory optimizations toward training trillion parameter models[C]//SC20: International Conference for High Performance Computing, Networking, Storage and Analysis. New York: IEEE, 2020.

[92]

DEEPSEEK-AI, GUO D, YANG D, et al. DeepSeek–R1: Incentivizing reasoning capability in LLMs via reinforcement learning[J/OL]. (2025–01–22)[2025–03–15]. https://doi.org/10.48550/ARXIV.2501.12948.

[93]

KANG Y, LEE W, BAE T, et al. Harnessing large language models to collect and analyze metal-organic framework property data set[J]. Journal of the American Chemical Society, 2025, 147(5): 39433958.

[94]

BAI X F, HE S, LI Y, et al. Construction of a knowledge graph for framework material enabled by large language models and its application[J]. NPJ Computational Materials, 2025, 11: 51.

[95]

GAO Y, XIONG Y, GAO X, et al. Retrieval-augmented generation for large language models: A survey[J/OL]. (2024–03–27)[2025–03–15]. https://doi.org/10.48550/arXiv.2312.10997.

[96]

BOIKO D A, MACKNIGHT R, KLINE B, et al. Autonomous chemical research with large language models[J]. Nature, 2023, 624(7992): 570578.

[97]

ZHANG D, LIU W, TAN Q, et al. ChemLLM: A chemical large language model[J/OL]. (2024–04–25)[2025–03–15]. https://doi.org/10.48550/arXiv.2402.06852.

[98]

GHAFAROLLAHI A, BUEHLER M J. AtomAgents: Alloy design and discovery through physics-aware multi-modal multi-agent artificial intelligence[J/OL]. (2024–07–13)[2025–03–15]. https://doi.org/10.48550/ARXIV.2407.10022.

[99]

弭光宝, 孙圆治, 吴明宇, . 机器学习在航空发动机钛合金研究中的应用进展[J]. 航空制造技术, 2024, 67(1/2): 6678.
MI Guangbao, SUN Yuanzhi, WU Mingyu, et al. Applications of machine learning on aero-engine titanium alloys[J]. Aeronautical Manufacturing Technology, 2024, 67(1/2): 6678.

[100]

JIANG L, FU H D, ZHANG Z H, et al. Synchronously enhancing the strength, toughness, and stress corrosion resistance of high-end aluminum alloys via interpretable machine learning[J]. Acta Materialia, 2024, 270: 119873.

[101]

HE M, LI Y, GHAFFARI B, et al. Machine learning-augmented modeling on the formation of Si–dominated Non–β″early-stage precipitates in Al–Si–Mg alloys with Si supersaturation induced by non-equilibrium solidification[J]. Acta Materialia, 2025, 282: 120454.

[102]

LIU T, PEI Z R, BARTON D, et al. Characterization of nanostructures in a high pressure die cast Al–Si–Cu alloy[J]. Acta Materialia, 2022, 224: 117500.

[103]

JUAN Y F, NIU G S, YANG Y, et al. Knowledge-aware design of high-strength aviation aluminum alloys via machine learning[J]. Journal of Materials Research and Technology, 2023, 24: 346361.

[104]

JIANG L, ZHANG Z H, HU H, et al. A rapid and effective method for alloy materials design via sample data transfer machine learning[J]. NPJ Computational Materials, 2023, 9: 26.

[105]

HARIHARAN A, ACKERMANN M, KOSS S, et al. High-speed 3D printing coupled with machine learning to accelerate alloy development for additive manufacturing[J]. Advanced Science, 2025, 12(17): 2414880.

[106]

ACKERMANN M, HAASE C. Machine learning-based identification of interpretable process-structure linkages in metal additive manufacturing[J]. Additive Manufacturing, 2023, 71: 103585.

[107]

TAKAHASHI R J, DE ASSIS J M K, FAZAN L H, et al. TBC development on Ti–6Al–4V for aerospace application[J]. Coatings, 2025, 15(1): 47.

[108]

LEE J A, PARK J, SAGONG M J, et al. Active learning framework to optimize process parameters for additive-manufactured Ti-6Al-4V with high strength and ductility[J]. Nature Communications, 2025, 16: 931.

[109]

MONTALBANO T, NIMER S, DAFFRON M, et al. Machine learning enabled discovery of new L–PBF processing domains for Ti–6Al–4V[J]. Additive Manufacturing, 2025, 98: 104632.

[110]

BAI J Y, YANG Y, WEN C, et al. Applications of magnesium alloys for aerospace: A review[J]. Journal of Magnesium and Alloys, 2023, 11(10): 36093619.

[111]

WANG Q H, QIN X, XIA S, et al. Interpretable machine learning excavates a low-alloyed magnesium alloy with strength-ductility synergy based on data augmentation and reconstruction[J]. Journal of Magnesium and Alloys, 2025, 13(6): 28662883.

[112]

LIU Z Y, WANG T Y, JIN L, et al. Towards high stiffness and ductility—The Mg–Al–Y alloy design through machine learning[J]. Journal of Materials Science & Technology, 2025, 221: 194203.

[113]

XUE D, WEI W, SHI W, et al. Optimization of stabilized annealing of Al–Mg alloys utilizing machine learning algorithms[J]. Materials Today Communications, 2023, 35: 106177.

[114]

WANG Y F, WANG K, ZHANG C. Applications of artificial intelligence/machine learning to high-performance composites[J]. Composites Part B: Engineering, 2024, 285: 111740.

[115]

KAUFMANN K, MARYANOVSKY D, MELLOR W M, et al. Discovery of high-entropy ceramics via machine learning[J]. NPJ Computational Materials, 2020, 6: 42.

[116]

XIAO J, GUO W J, YANG J G, et al. Analysis and regularity of ablation resistance performance of ultra-high temperature ceramic matrix composites using data-driven strategy[J]. Ceramics International, 2024, 50(18): 3193731945.

[117]

BORKOWSKI L, SKINNER T, CHATTOPADHYAY A. Woven ceramic matrix composite surrogate model based on physics-informed recurrent neural network[J]. Composite Structures, 2023, 305: 116455.

[118]

ZHOU J, LI L, LU L, et al. Machine learning-based quality optimisation of ceramic extrusion 3D printing deposition lines[J]. Materials Today Communications, 2024, 41: 110841.

[119]

KACZMAR J W, PIETRZAK K, WŁOSIŃSKI W. The production and application of metal matrix composite materials[J]. Journal of Materials Processing Technology, 2000, 106(1–3): 5867.

[120]

ZHONG Z Y, AN J, WU D, et al. A machine learning strategy for enhancing the strength and toughness in metal matrix composites[J]. International Journal of Mechanical Sciences, 2024, 281: 109550.

[121]

HASAN M S, WONG T, ROHATGI P K, et al. Analysis of the friction and wear of graphene reinforced aluminum metal matrix composites using machine learning models[J]. Tribology International, 2022, 170: 107527.

[122]

PIMENTA S, PINHO S T. Recycling carbon fibre reinforced polymers for structural applications: Technology review and market outlook[J]. Waste Management, 2011, 31(2): 378392.

[123]

MARR J, ZARTMANN L, REINEL-BITZER D, et al. Data-based prediction of the viscoelastic behavior of short fiber reinforced composites[J]. PAMM, 2023, 22(1): e202200085.

[124]

LI M Z, ZHANG H W, LI S R, et al. Machine learning and materials informatics approaches for predicting transverse mechanical properties of unidirectional CFRP composites with microvoids[J]. Materials & Design, 2022, 224: 111340.

[125]

THOMAS A J, BAROCIO E, PIPES R B. A machine learning approach to determine the elastic properties of printed fiber-reinforced polymers[J]. Composites Science and Technology, 2022, 220: 109293.

[126]

LU L, HOU J, YUAN S Q, et al. Deep learning-assisted real-time defect detection and closed-loop adjustment for additive manufacturing of continuous fiber-reinforced polymer composites[J]. Robotics and Computer-Integrated Manufacturing, 2023, 79: 102431.

[127]

LIU T, ZHANG T, CHEN Y, et al. Neural co-optimization of structural topology, manufacturable layers, and path orientations for fiber-reinforced composites[J/OL]. (2025–04–30)[2025–03–15]. https://doi.org/10.48550/arXiv.2505.03779.

[128]

SONG J, LEE J, KIM N, et al. Artificial intelligence in the design of innovative metamaterials: A comprehensive review[J]. International Journal of Precision Engineering and Manufacturing, 2024, 25(1): 225244.

[129]

NIAN Y Z, WAN S, AVCAR M, et al. Nature-inspired 3D printing-based double-graded aerospace negative Poisson’s ratio metastructure: Design, fabrication, investigation, optimization[J]. Composite Structures, 2024, 348: 118482.

[130]

WANG H F, ZHANG C, WANG C, et al. The application of negative Poisson’s ratio metamaterials in the optimization of a variable area wing[J]. Aerospace, 2025, 12(2): 125.

[131]

WANG M H, SUN S, ZHANG T Y. Machine learning accelerated design of auxetic structures[J]. Materials & Design, 2023, 234: 112334.

[132]

ZHAO M, LI X W, YAN X, et al. Machine learning accelerated design of lattice metamaterials for customizable energy absorption[J]. Thin-Walled Structures, 2025, 208: 112845.

[133]

FAN J X, SONG B, ZHANG L, et al. Structural design and additive manufacturing of multifunctional metamaterials with low-frequency sound absorption and load-bearing performances[J]. International Journal of Mechanical Sciences, 2023, 238: 107848.

[134]

ZHANG H J, LIU J W, MA W T, et al. Learning to inversely design acoustic metamaterials for enhanced performance[J]. Acta Mechanica Sinica, 2023, 39(7): 722426.

[135]

FENG M F, ZHANG K F, XIAO J J, et al. Material-structure collaborative design for broadband microwave absorption metamaterial with low density and thin thickness[J]. Composites Part B: Engineering, 2023, 263: 110862.

[136]

LIM D D, IBARRA A, LEE J, et al. A tunable metamaterial microwave absorber inspired by chameleon’s color-changing mechanism[J]. Science Advances, 2025, 11(3): eads3499.

目录