Development of AI based Ab initio quality Potentials for large clusters of AStrOphysical iNtereSt – DIAPASONS

The DIAPASONS project aims at improving our knowledge of the role of (nano-)grains in the physics and chemistry of the interstellar medium via atomistic modeling using a state-of-the art methodology based on Deep Potential Molecular Dynamics (DPMD).
There is indeed a crucial lack of accurate theoretical data to characterize (i) structural, energetic and spectral properties of interstellar nanograins, (ii) dynamical processes occurring in space leading to their growth and destruction, -(iii)- gas-grain chemistry and reactive processes at the surface of grains. This lack is due to the large size of the systems, their complex potential energy surface (PES), and the need to describe long-time dynamic processes at low temperature. Moreover, performing extensive molecular dynamics (MD) simulations is required to be able to derive converged statistics. Furthermore, the reliability of the results depends on the quality of the PES, which has suffered from methodological limitations so far. However, such accurate data are deeply needed to complement experimental data and to feed astrochemical databases.
In this respect, the DIAPASONS project proposes the development of ab initio quality potentials and atomic dipoles based on Deep Neural Network (DNN) learning to study grains of astrophysical interest of various sizes and natures. The Deep-MD kit software will be used to train the DNN potentials (DNNPs). We will focus on modeling prototype clusters for icy grains such as (H2O)n, (CO)n and mixed H2On(CO)m(CO2)p clusters, and polycyclic aromatic hydrocarbons (PAHs)n clusters that are good candidates for interstellar carbon grains. Mixed (PAH)m(H2O)n clusters studied experimentally, will also be considered as well as clusters possessing an impurity (H+, OH-). In a first step, DNNPs of DFTB (density functional based tight binding) quality will be obtained using data issued from DFTB simulations as reference data for fitting. Nuclear Quantum effects (NQEs) will be investigated using reference data from Path-Integral MD simulations. The DNNPs will then be refined though iterative training to obtain reference data of DFT quality or higher. The expected results are pools of accurate DNNPs and DNNADs for the systems of interest heading towards a unique pool.
Once accurate DNNPs and DNNADs are obtained, properties resulting from extensive MD simulations using these potentials and these atomic dipoles will be retrieved. To achieve this goal, we will use the i-PI software interfaced with the Deep-MD kit. The exhaustive exploration of the PES will be achieved via Parallell Tempering (PT) MD and PT-PIMD simulations. Cohesion energies and heat capacities will be determined and compared to experimental data when available. MD and PIMD extensive simulations will be performed to determine finite-temperature infrared spectra and to investigate dynamical processes: destruction by thermal evaporation or collision induced dissociation, growth by low-energy collisions with determination of sticking coefficients, reactivity at the surface of the clusters simulated by gas-grain collisions. NQEs, particularly relevant at low temperature and for light atoms, will be investigated using the DNNPs trained from PIMD simulations.
This four-year project will be carried out by four researchers and one PhD student of the “Modeling, Aggregates, Dynamics” group of the Laboratoire de Chimie et Physique Quantiques (Toulouse, France). One thesis scholarship and one post-doctoral fellowship are requested to carry out this challenging project.
Carrying out this project is of paramount importance for our group, that has just started to become familiar with machine-learning (ML) techniques, expected to be powerful for the systems planned in this project. Further applications are planned for clusters of atmospheric interest and to make progress in the description of soil/pesticide interactions.