石墨烯对众多应用和基础研究均至关重要，但目前仍缺乏一种成本低廉的合成方法来进行大规模和高质量的生产。即使是目前最优的制备方法—化学气相沉积(Chemical Vapor Deposition, CVD)，仍然面临着许多挑战，如石墨烯缺陷控制、低温生长以及层数调控等。而理论上的认知不足将严重阻碍实验方案的设计与优化，为增强理论与实验研究的协同作用以解决前沿关键问题，本论文研究并发展了两种可应用于石墨烯生长体系的机器学习力场—DP-C/Cu和DP-C/Ni/Cu，以打破传统计算模拟方法精度和效率的两难困境。接近真实生长情形的准确模拟能够揭示原子尺度下石墨烯生长机制，为实验设计提供理论见解和决策参考。DP-C/Cu经过验证能充分学习并表达铜表面石墨烯生长行为。其对于C/Cu复合体系的力预测误差仅1.31×10-1 eV/?，能量预测误差仅5.36×10-3 eV/atom。利用该力场，本论文探索了石墨烯生长初期碳团簇稳定性及边缘生长路径，系统地研究了生长过程中支链碰撞成环以及生长前端缺陷愈合等关键过程。深入分析了不同反应温度及生长速率条件下碳结构变化趋势，为缺陷控制提供了理论指导。针对晶界控制及低温生长等挑战，本论文基于DP-C/Cu对特定化学气相沉积条件下石墨烯生长情形进行了分子动力学模拟。揭示了石墨烯在平台表面、越过表面台阶以及跨越基底晶界时，其晶界生成概率逐渐提升，基底表面状态将决定石墨烯不同的合并生长机制。进一步地，本论文通过调控生长温度、碳源供应速率以及沉积位点等生长条件模拟了相应石墨烯生长过程，为低温生长时石墨烯的低均匀性提供了理论解释，提出了控制碳源沉积位点并结合单晶粒外延生长方法予以改善的可行方法。针对石墨烯层数调控问题，本论文首次利用机器学习力场结合分子动力学模拟了铜镍合金表面双层石墨烯生长过程。结果表明除表面介导以及偏析生长机制外，同步生长机制在双层石墨烯生长过程中也起到了重要作用，为其生长机制提供了新的解释。此外，还研究发现了双层石墨烯生长模式对石墨烯堆垛顺序的重要影响。研究中使用的三元体系机器学习力场DP-C/Ni/Cu是基于DP-C/Cu发展而来的，其对C/Ni/Cu复合体系的力预测误差低至1.27×10-1 eV/?，对训练集外新体系的力预测误差仅1.39×10-1 eV/?。并且补充新数据仅占总数据的15.86%，可基于此提供一种计算成本低且性能优异的机器学习力场应用体系迁移方法。
Graphene is critical for a wide range of applications and fundamental research, but a cost-effective method for large-scale and high-quality production is still lacking. Chemical vapor deposition(CVD), as the current optimal preparation method, still faces many challenges, such as defect control, low-temperature growth, and layer number tuning. While the lack of theoretical knowledge will seriously hinder the design and optimization of experimental schemes, in order to enhance the synergy between theoretical and experimental research to address cutting-edge critical issues, this thesis researches and develops two kinds of machine learning force fields that can be applied to the graphene growth system, DP-C/Cu and DP-C/Ni/Cu, which breaks the dilemma of the traditional computational simulation methods in terms of accuracy and efficiency. Accurate simulation close to the real growth situation can reveal the graphene growth mechanism at the atomic scale and provide theoretical insights and decision-making references for experimental design.The machine learning force field DP-C/Cu is demonstrated to adequately learn and express graphene growth behavior on copper surfaces. The force prediction error for the C/Cu composite systems is only 1.31×10-1 eV/?, and the energy prediction error is only 5.36×10-3 eV/atom. By using this force field, this thesis explores the stability of the carbon clusters in the early stage of the graphene growth and the edge growth path, and systematically researches the processes of collision of the branched chains to form the rings and the healing of the defects in the growth front. The trend of carbon structure change under different reaction temperature and growth rate is deeply analyzed, providing theoretical guidance for defect control.To address the issues of grain boundary formation and low-temperature growth, molecular dynamics simulations of graphene growth under specific chemical vapor deposition conditions are carried out using DP-C/Cu. It is revealed that the probability of grain boundary formation in graphene is sequentially increased when graphene is on the terrace, crossing the surface step, and crossing the substrate grain boundary, and that the state of substrate surface will determine the growth mechanisms of graphene. Further, the corresponding graphene growth process is simulated by regulating the growth conditions such as growth temperature, carbon source supply rate and deposition site, which provides a theoretical explanation for the bad uniformity of graphene during low-temperature growth. And a feasible method for controlling the carbon source deposition sites and combining it with the single-grain epitaxial growth method is proposed to improve it.To address the problem of graphene layer number tuning, this thesis simulates the growth process of bilayer graphene on the surface of Cu/Ni alloy for the first time using machine learning force field combined with molecular dynamics. The results show that in addition to the surface-mediated as well as segregation growth mechanisms, the synchronic growth mechanism also plays an important role in the growth process of bilayer graphene, which provides a new explanation for its growth mechanism. In addition, the important influence of the growth behavior of bilayer graphene on the stacking order of graphene is also investigated. The machine learning force field for ternary system used in the study, DP-C/Ni/Cu, is developed based on DP-C/Cu, which has a force prediction error as low as 1.27×10-1 eV/? for the C/Ni/Cu composite systems, and only 1.39×10-1 eV/? for the new systems outside of the training set. And the supplemental new data is only 15.86% of the total data, thus, a low computational cost and high performance transfer method for machine learning force field application systems can be provided.
