An active learning force field for the thermal transport properties of organometallic complex crystals

The accurate prediction of lattice thermal conductivity in organometallic thermoelectric materials is crucial for advancing energy conversion technologies. Methods based on molecular dynamics simulations can solve this problem well, but require force fields with sufficiently high accuracy. Due to the complexity of chemical bonding in organometallic complex materials, the development of force fields with high predictivity has been a long standing challenge, particularly when thermal transport is concerned which requires even greater accuracy. In recent years, the rapid advancement of machine learning force fields has offered substantial potential for addressing these issues. However, there remain challenges for materials with large organometallic complexes in one unit cell and both inter- and intra-molecular interactions. In this work, we employ an active learning approach combined with deep neural networks to develop a force field taking copper phthalocyanine as an example. The model utilizes a local environment descriptor for representation without explicitly characterizing the metal-organic coordination. The nonlinear mapping capabilities of deep neural networks enable the model to effectively capture higher-order many-body interactions. Furthermore, we utilized the Green-Kubo method to calculate the thermal conductivity of copper phthalocyanine, revealing a value of 0.49 W m^-1 K^-1 at 300 K, consistent with experimental findings (0.39 W m^-1 K^-1). This result significantly surpasses previous work with classical force fields. This work represents a significant advancement in demonstrating that machine-learning force fields can effectively characterize interactions in metal-organic complex systems and can significantly advance the development and discovery of organometallic thermoelectric materials.