Amorphous Materials by Design through Atomistic Simulations

新技术相关材料的发现和设计是物理科学的主要研究目标。大型超级计算机上的量子力学模拟使材料的计算设计触手可及：确定合适的成分，寻找稳定的晶体结构，指导和激发实验发现。然而，这一进展在很大程度上仅限于具有长程结构有序和相对较小的晶胞的晶体材料。相比之下，非晶（非结晶）状态一直是预测原子模拟的长期挑战，这严重限制了新材料的发现范围。我在这里建议克服这一重要挑战：通过开发机器学习（ML）驱动的方法，用于原子尺度建模，优化和设计具有所需特性的多组分非晶材料。这个雄心勃勃的项目将利用监督和无监督ML算法的力量来“学习”和导航结构空间。第一个目标是开发一种方法，用于创建ML原子间势的普遍适用的拟合数据库-从而使多组分非晶材料适合于现实的原子模型，与今天的结晶固体处理方式相同。第二个目标是开发一种新的深度学习模型，用于预测固态NMR位移，但对基于ML的性质预测具有更普遍的意义。最后，该方法将用于计算实践，并应用于关键材料系统。该项目将在结构建模和对非晶状态的理解方面开辟新的现实主义程度，提供丰富的公开可用的研究数据，并最终实现新的非晶材料的计算驱动设计。

The discovery and design of new technologically relevant materials is a major research goal in the physical sciences. Quantum-mechanical simulations on large supercomputers have brought the computational design of materials within reach: identifying suitable compositions, searching for stable crystal structures, guiding and inspiring experimental discoveries. However powerful, this progress has been largely limited to crystalline materials with long-range structural order and relatively small unit cells. In contrast, the amorphous (non-crystalline) state has been a long-standing challenge for predictive atomistic simulations, which has severely restricted the range of new materials to be discovered. I here propose to overcome this important challenge: by developing machine learning (ML) driven approaches for the atomic-scale modelling, optimisation, and design of multicomponent amorphous materials with desired properties. This ambitious project will leverage the power of both supervised and unsupervised ML algorithms to "learn" and navigate structural space. The first objective is to develop methodology with which to create universally applicable fitting databases for ML interatomic potentials - thereby making multicomponent amorphous materials amenable to realistic atomistic modelling, on par with how crystalline solids are treated today. The second objective is to develop a novel deep learning model, here to be used for predicting solid-state NMR shifts, but with more general implications for ML-based property prediction. Finally, the methodology will be used in computational practice and applied to key materials systems. This project will open up a new degree of realism in the structural modelling and understanding of the amorphous state, provide a wealth of openly available research data, and ultimately enable the computationally driven design of new amorphous materials.