ИСТИНА

High-level quantum chemical potential energy surfaces (PESs) and dipole moment surfaces (DMSs) are indispensable for accurately simulating molecular spectra. Representing a global multidimensional PES or DMS based on the array of preselected ab initio points across the configuration space poses a difficult challenge. Machine learning, and neural networks in particular, have shown great promise for building flexible and computationally efficient models [1, 2]. In the current study, we undertake the development of machine-learned PESs and DMSs for weakly bounded complexes for simulating the thermophysical and spectral properties of atmospheric gases. A recently suggested permutationally invariant neural network (PIP-NN) approach [3] is refined and extended to represent the full-dimensional two-body components of the molecular pair energy and dipole. To ensure the asymptotic zero-interaction limit, as well as to accurately represent the long-range behavior, a tailored subset of the polynomial basis set is utilized. The suggested technique is showcased by fitting databases of N2-Ar and N2-CH4 interaction energies and N2-Ar induced dipoles calculated using CCSD(T)/CCSD(T)-F12 methods. The second virial coefficient, fully accounting for molecular flexibility, is calculated within the classical framework. A firstprinciple trajectory-based simulation [4] of the N2-Ar collision-induced absorption is then conducted both in the far- and mid-infrared ranges using the constructed PIP-NN PES and DMS. This work is supported by the Russian Science Foundation (project No. 22-17-00041)