Cover News trainee reporter Che Jiazhu
On October 23, the reporter learned from the University of Science and Technology of China that the research group led by Jiang Bin has made important progress in the research of the force field of the atomic neural network induced by the development field. The research results, titled "General Machine Learning Force Fields Describing the Response of Atomic Systems to External Fields", were published in Nature Communications on October 12, 2023.
Help with atomic simulation
Understand the world of chemistry on a microscopic level
Atomic simulations are a key tool for understanding the spectra, reaction kinetics, and energy and charge transfer processes of complex chemical, biological, and material systems at the microscopic level, with accurate and efficient high-dimensional potential energy surfaces (i.e., force fields).
In recent years, efficient and accurate molecular simulations based on machine learning interaction potentials at precise atomic centers have become a common practice. However, these models are often used to describe isolated systems, treating energy only as a function of atomic coordinates and atomic species, and cannot express the interaction between the outer field and the system. Exterior fields can trigger electron polarization, spin polarization, and changes in spatial orientation of systems through interaction with atoms, molecules, or solid materials.
This provides an important tool for changing chemical structures, controlling phase transitions of materials, and precisely manipulating chemical reactivity and selectivity in catalytic reactions. Therefore, there is an urgent need to develop machine learning models that correctly describe the interaction between the field and the system to achieve accurate and efficient simulation of complex reactions under the field.
Inspired by quantum chemistry
Research on high-precision machine learning force field methods
In addition, the reporter also learned that Professor Jiang Bin's research group has long been committed to the research of high-precision machine learning force field methods. Inspired by the concept of linear combination of atomic orbitals into molecular orbitals in quantum chemistry, the researchers proposed a recursively embedded atomic electron density descriptor, then treated the outer field as a virtual atom (see Figure 1), and introduced a linear combination of field-dependent atomic orbitals and coordinate-based atomic orbitals to obtain symmetry-adapted field-dependent embedded charge density, thus developing a field-induced symmetry-matched recursive embedding atomic neural network method.
Figure 1: Schematic diagram of a field-induced recursive embedding atomic neural network model
Compared with the research results in 2019, the research team improved the embedded atomic neural network method developed in the early stage, so that the orbital coefficients in the embedded charge density descriptor became chemically environment-dependent, and the recursive embedding atomic neural network method was proposed by updating the embedded charge density descriptor recursively (see Figure 2).
Figure 2: Schematic diagram of a recursively embedded atomic neural network model
Interestingly, this neural network approach is essentially the same as the physically less intuitive form of messaging neural network. The research team further demonstrated that more volume interactions can be introduced in the form of recursively updating orbital coefficients, deriving a complete description of a local chemical environment and determining the relationship between the number of iterations (number of messages passed) and the number of neighboring atoms. This method does not need to explicitly calculate higher-order interactions, greatly simplifies calculations, and explains the superiority of message-passing neural networks from the perspective of many-body interactions.
After several years of efforts, the latest results published recently, this method can accurately correlate various response properties such as dipole moment and polarizability with the energy changes dependent on the external field, and is suitable for the spectral and kinetic simulation of molecular and periodic systems in the presence of the external field. In particular, for periodic systems, this model only needs to train atomic force data to overcome the problem of polarization multi-value inherent in periodic systems. Through the kinetic simulation results of methylacetamide and liquid water (see Figures 3 and 4), the ability of this model to efficiently model various complex systems under strong external field conditions is verified.
Figure 3: Vibrational spectra of methylacetamide molecules
Figure 4: Infrared spectrum of liquid water
This research work correlates physical concepts with machine learning descriptors, providing new ideas for the development of more general machine learning models.
[If you have news leads, welcome to report to us, once adopted, there is a fee to thank you.] Report WeChat concern: ihxdsb, report QQ: 3386405712]