Machine learning the thermodynamics of complex materials with ab initio accuracy

拟议项目Mach-Initio的总体目标是从头开始研究确定复杂材料基本热力学性质的基本物理有限温度激发机制及其相互耦合效应。为此，我们将机器学习领域和从头材料设计领域的PI的想法和知识整合到一个独特的框架中。PI在最近的过去成功合作，为Mach-Initio提供了坚实的基础。将他们的专业知识结合到一个单一的联合项目中，保证了雄心勃勃的目标的成功实现。从头算方法已经成功地应用于计算材料的零开尔文基态能量多年，然而，在有限温度下直接从头算计算激发在大多数情况下是昂贵的。PI最近的联合工作表明，基于机器学习势的有效哈密顿量，特别是矩张量势（MTP）和低秩势（LRP），可以用来大大减少计算工作量，从而实现从头计算精度的振动和构型激发的高效研究。本项目的目的是进一步推进这些有限温度从头计算方法的基础上的MTP和LRP。特别是，我们将开发用于计算高度精确的自由能表面的算法，包括与振动，构型熵，磁性及其相互耦合效应相关的所有相关激发机制。为此，将开发一种新型的MTP，该MTP占磁自由度（mMTP）。多边中期计划的概念和方法框架将由俄方制定。德国方面将继续将mMTP整合到基于从头算的热力学方法中，包括技术相关材料系统的应用和验证。

The general aim of the proposed project Mach-Initio is the ab initio investigation of the fundamental physical finite-temperature excitation mechanisms and their mutual coupling effects that determine the basic thermodynamic properties in complex materials. For that purpose, we leverage ideas and knowledge of the PIs from the field of machine learning and the field of ab initio materials design into a single and unique framework. The PIs have successfully collaborated in the recent past providing a solid basis for Mach-Initio. The combination of their expertise into a single joint project guarantees a successful fulfillment of the ambitious goals.Ab initio methods have been successfully applied for many years to calculate the zero Kelvin ground state energy of materials, however, direct ab initio calculations of excitations at finite temperatures are in most cases prohibitively expensive. Recent joint work of the PIs has shown that effective Hamiltonians based on machine-learning potentials, in particular moment tensor potentials (MTPs) and low-rank potentials (LRPs), can be utilized to reduce the computational effort drastically, thus enabling a highly efficient study of the vibrational and configurational excitations with ab initio accuracy. The aim of the present project is to further advance these finite-temperature ab initio approaches based on MTPs and LRPs. In particular, we will develop algorithms for the computation of a highly accurate free energy surface including all relevant excitation mechanisms related to vibrations, configurational entropy, magnetism, and their mutual coupling effects. To this end, a novel type of MTPs will be developed that accounts for the magnetic degrees of freedom (mMTPs). The conceptual and methodological framework of the mMTPs will be developed by the Russian side. The integration of the mMTPs into an ab initio-based thermodynamic methodology including the application to and validation for technologically relevant material systems will be pursued by the German side.