How amorphous carbon breaks: atomistic models and machine learning

非晶态碳(a-C)具有许多工业应用，从电化学传感器到耐磨涂层。断裂是涂层性能下降的关键因素，涂层经常因剪切或弯曲裂纹而失效。这意味着，除了能够预测断裂韧性之外，了解拉伸和剪切混合载荷的响应并预测裂纹的轨迹也是至关重要的。在这个项目中，我们将建立在数据驱动的方法上，这些方法使用机器学习技术来以一小部分的成本产生量子力学精确的模型，并使用它们来产生对a-C中裂纹扩展的完整描述。只有原子模拟才具有真正预测的能力，因为更大规模的模型，如X-有限元、相场和其他总是包括经验裂纹扩展算法。该项目将涉及与弗莱堡大学的Lars Pastewka教授的合作，项目主管最近与他一起展示了原子模拟可以用来产生与实验非常一致的a-C断裂韧性的定量预测[1]。这项工作使用标准的连续体线弹性边界条件，因此需要较大的原子域，以防止扩展到裂纹路径选择或混合模式加载。该项目还将使用一种新的数值连续增强型灵活边界条件格式NCFlex，该格式最近由主管与Maciej Buze博士(伯明翰大学)[2]共同开发，后者也将参与该团队。该方法将材料建模技术与数值分析相结合，生成了裂纹的分叉图。a-C的断裂代表了一个“甜蜜点”，工艺区域是纳米级的，因此可以直接进行原子模拟，但仍具有直接的技术重要性。它是目前唯一一种可以用预测原子学方法研究断裂特性的各向同性材料。对于真正的预测模型，还需要提高原子间相互作用势的准确性。在这个项目中，我们将建立在数据驱动的方法上，例如[3]，它使用机器学习技术以一小部分的成本产生QM精确的力场，并超越拉伸加载模拟来产生关于混凝土中裂纹扩展的完整描述。

Amorphous carbon (a-C) has many industrial applications, from electrochemical sensors to wear-resistant coatings. Fracture plays a crucial role in the degradation of its performance, with coatings often failing by shear or flexural cracks. This means that as well as being able to predict fracture toughness, it is crucial to understand the response to mixed tensile and shear loads and predict the trajectory of cracks. In this project, we will build on data-driven approaches that use machine learning techniques to produce quantum mechanically accurate models at a fraction of the cost, and use them to produce a complete description of crack growth in a-C.Only atomistic simulations have the capability of being truly predictive, since larger scale models such as X-FEM, phase-field and others invariably include empirical crack growth algorithms. The project will involve collaboration with Prof. Lars Pastewka at the University of Freiburg, with whom the project supervisor has recently shown that atomistic modelling can be used to produce quantitative predictions of the fracture toughness of a-C in good agreement with experiment [1]. This work used standard continuum linear elastic boundary conditions, and thus required large atomistic domains, preventing extension to crack path selection or mixed-mode loading.The project will also employ a novel numerical continuation enhanced flexible boundary condition scheme, NCFlex, that has recently been developed by the supervisor with Dr Maciej Buze (University of Birmingham) [2] who will also be involved in the team. The approach fuses materials modelling techniques with numerical analysis to produce bifurcation diagrams for cracks.Fracture of a-C represents a "sweet spot" where the process zone is nanoscale and hence accessible to direct atomistic simulation, but still of immediate technological importance. It is currently the only isotropic material whose fracture properties can be studied with predictive atomistic methods. For truly predictive models, improved accuracy is also needed in the interatomic potential. In this project, we will build on data-driven approaches such as [3] that use machine learning techniques to produce QM-accurate force fields at a fraction of the cost, and go beyond tensile loading simulations to produce a complete description of crack growth in a-C.