Constructing feed-forward artificial neural networks to fit potential energy surfaces for molecular simulation of high-temperature gas flows

Ab initio investigation of a hypersonic double cone experiment

This article presents a direct molecular simulation (DMS) of a reactive Mach 8.2 oxygen flow over a double cone geometry. The free stream conditions and article configuration generate a flow with thermal and chemical nonequilibrium, which are common attributes of hypersonic flight. This scenario was first studied experimentally at Calspan University of Buffalo Research Center's test facility. DMS is a particle method that uses quantum mechanically derived interaction potentials to simulate molecular collisions within a flow field. Since these interaction potentials are the only modeling inputs used in the simulation, all flow features can solely be attributed to the ab initio potential energy surfaces. Hence, providing a comparison of a hypersonic ground test and numerical data anchored to quantum mechanics. The fundamental nature of DMS is leveraged to investigate molecular level mechanisms prevalent in the flow, and comparisons with lower fidelity simulations are presented to highlight the role of these first principles calculations as benchmark solutions.

Exhaustive State-to-State Cross Sections and Rate Coefficients for Inelastic N2-N2 Collisions using QCT Combined with Neural Network Models

Using the quasi-classical trajectory method, we systematically studied the state-to-state vibrational relaxation process of N2(v1) + N2(v2) collisions over a wide temperature range (5000-30,000 K). Different temperature dependencies of the single- and multiquantum VV and VT events in various (v1,v2) collisions are captured, with the dominant channel being related to the initial vibrational energy levels (vmax = 50). At a specified relative translational energy, there is a monotonic relationship of the VT cross sections with the vibrational energy level, particularly in high-energy collisions. Additionally, we constructed well-trained neural network models (R-values reaching 0.99) using limited quasi-classical trajectory (QCT) data sets, which can be used to predict the state-to-state cross sections and rate coefficients of the VV processes N2(v1) + N2(v2) → N2(v1 - Δv) + N2(v2 + Δv) and VT processes N2(v1) + N2(v2) → N2(v1 - Δv) + N2(v2) (Δv = ±1, ±2, ±3) for collisions with arbitrary initial vibrational states. This work not only significantly reduces computational resources but also serves as a reference for the study of the state-to-state dynamics of all four-atom collision systems in hypersonic flows.

Machine Learning Diffusion Monte Carlo Forces

Diffusion Monte Carlo (DMC) is one of the most accurate techniques available for calculating the electronic properties of molecules and materials, yet it often remains a challenge to economically compute forces using this technique. As a result, ab initio molecular dynamics simulations and geometry optimizations that employ Diffusion Monte Carlo forces are often out of reach. One potential approach for accelerating the computation of "DMC forces" is to machine learn these forces from DMC energy calculations. In this work, we employ Behler-Parrinello Neural Networks to learn DMC forces from DMC energy calculations for geometry optimization and molecular dynamics simulations of small molecules. We illustrate the unique challenges that stem from learning forces without explicit force data and from noisy energy data by making rigorous comparisons of potential energy surface, dynamics, and optimization predictions among ab initio density functional theory (DFT) simulations and machine-learning models trained on DFT energies with forces, DFT energies without forces, and DMC energies without forces. We show for three small molecules─C₂, H₂O, and CH₃Cl─that machine-learned DMC dynamics can reproduce average bond lengths and angles within a few percent of known experimental results at one hundredth of the typical cost. Our work describes a much-needed means of performing dynamics simulations on high-accuracy, DMC PESs and for generating DMC-quality molecular geometries given current algorithmic constraints.

Potential energy surfaces for high-energy N + O2 collisions

Potential energy surfaces for high-energy collisions between an oxygen molecule and a nitrogen atom are useful for modeling chemical dynamics in shock waves. In the present work, we present doublet, quartet, and sextet potential energy surfaces that are suitable for studying collisions of O₂(³Σ_g ^-) with N(⁴S) in the electronically adiabatic approximation. Two sets of surfaces are developed, one using neural networks (NNs) with permutationally invariant polynomials (PIPs) and one with the least-squares many-body (MB) method, where a two-body part is an accurate diatomic potential and the three-body part is expressed with connected PIPs in mixed-exponential-Gaussian bond order variables (MEGs). We find, using the same dataset for both fits, that the fitting performance of the PIP-NN method is significantly better than that of the MB-PIP-MEG method, even though the MB-PIP-MEG fit uses a higher-order PIP than those used in previous MB-PIP-MEG fits of related systems (such as N₄ and N₂O₂). However, the evaluation of the PIP-NN fit in trajectory calculations requires about 5 times more computer time than is required for the MB-PIP-MEG fit.