CAREER: Machine Learning Enabled Study of Thermal Transport in Polycrystalline Materials from First Principles

具有超高、超低或各向异性导热系数的新材料的开发可能使新型能量存储和转换设备以及电子设备的有效热管理成为可能。晶界普遍存在于固体材料中，对热输运有重要影响，因此设计晶界对于开发具有所需热性能的新型材料至关重要。由于在获得实验数据方面的挑战，以及依赖于使用经验势的模拟研究的局限性，通过晶界的热传输还没有被很好地理解。这项拟议的研究将开发一种新的多尺度模拟框架，将机器学习技术和第一原理计算相结合。新的框架具有与直接第一原理计算相当的高精度，并且在计算上是可行的，使得能够发现跨越晶界的热传输过程的基本物理。还提出了几项教育活动，以提高公众意识，特别是关于机器学习技术如何改变基础科学和工程研究的意识。这一职业项目的目标是建立对具有第一原理高预测能力的各种类型晶界的热传输的定量理解。新的多尺度模拟框架有可能使计算成本比直接的第一原理计算低几个数量级。这是通过集成(I)在~1 nm尺度下局部原子势景观的原子间势的机器学习，(Ii)在10-100 nm尺度上声子通过晶界散射的原子格林函数方法，以及(Ii)在亚毫米尺度上的总声子输运的Peierls-Boltzmann输运理论来实现的。使用这一新的模拟框架，该项目将寻求对几种实际相关的2D和3D半导体多晶体中的声子输运获得决定性的了解。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The development of new materials with ultrahigh, ultralow, or anisotropic thermal conductivity can potentially enable novel energy storage and conversion devices, and effective thermal management of electronics. Grain boundaries, which commonly exist in solid materials, significantly affect thermal transport, and thus engineering grain boundaries is crucial in developing novel materials of desired thermal properties. Due to challenges in obtaining experimental data and limitations in simulation studies that rely on employing empirical potentials, thermal transport across grain boundaries is not well understood. The proposed research will develop a new multiscale simulation framework that combines machine learning techniques and first-principles calculations. The new framework has a high accuracy comparable to direct first-principles calculations and is computationally feasible, enabling the discovery of the underlying physics of thermal transport processes across grain boundaries. Several educational activities are also proposed to increase public awareness, particularly about how machine learning techniques transform the basic science and engineering research. The goal of this CAREER project is to establish a quantitative understanding of thermal transport across various types of grain boundaries with the high predictive power of first principles. The new multiscale simulation framework has the potential to keep the computational cost several orders-of-magnitude cheaper than the direct first-principles calculation. This is made possible by integrating (i) machine learning of interatomic potentials for local atomic potential landscape at ~ 1 nm scale, (ii) atomistic Green's function method for phonon scattering by grain boundaries at 10 to 100 nm scale, and (ii) the Peierls-Boltzmann transport theory for overall phonon transport at sub-mm scale. Using this new simulation framework, this project will seek to obtain a conclusive understanding of phonon transport in several practically relevant 2D and 3D semiconductor polycrystals.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.