Inefficient intramolecular vibrational energy redistribution for the H + HO2 reaction and negative internal energy dependence for its rate constant

Abstract

Quasiclassical trajectories (QCT) and newly constructed global potential energy surfaces are used to compute thermal and nonthermal rate constants for the H + HO2 reaction. The thermal QCTs rate constants are up to 50% smaller than transition state theory (TST) rate constants based on the same level of electronic structure theory. This reduction is demonstrated to result from inefficient intramolecular vibrational energy redistribution (IVR) in the transient H2O2 well, with a significant fraction of trajectories that reach the H2O2 well promptly dissociating back to reactants instead of via the heavily statistically favored 2OH channel. The nonstatistical reduction factor, κIVR, that quantifies this effect is shown to increase in importance with temperature, with κIVR = 0.81 at 300 K and 0.47 at 2500 K. Finally, we show that inefficient IVR causes H + HO2 rate constants mediated by H2O2 to depend inversely on the initial vibrational excitation of HO2.

Quasi-Classical Trajectory Calculation of Rate Constants Using an Ab Initio Trained Machine Learning Model (aML-MD) with Multifidelity Data

“…The higher-accuracy data consist of energies for the same geometries calculated using the MRCI+Q method employing a complete basis set (CBS) limit extrapolation of aug’-cc-pVDZ and aug’-cc-pVTZ results, as described previously . The MRCI+Q/CBS level of theory for this system was benchmarked against ANL1 energies in ref and generally exhibits higher accuracy than DFT, albeit at a much higher computational cost. , Such calculations are particularly important for geometries where multireference effects are important, such as the interactions between two radicals at large separations. For example, Liu et al employed both MRCI and ωB97X-D functionals to calculate the ionization energy of a Boron cluster, finding that the error of MRCI functional was only a quarter of that of ωB97X-D when compared to experimental results.…”

“…The promise of the transfer learning methodology lies in the use of existing data sets with different accuracies. In this case, the higher-accuracy energy data generated at the MRCI+Q/CBS level of theory were also used during the rate constant predictions for H + HO 2 → H 2 + 3 O 2 using DiNT . Hence, this higher-accuracy data set is a good test-case to assess the capabilities of transfer learning methodologies.…”

“…The initial ML-MD model trained using the DeePMD-kit framework, hereafter referred as the aML-MD model, was trained using DFT calculations based on geometries from Jasper et al 16 Briefly, these geometries were generated in batches for geometries surrounding the unimolecular stationary points (via expansions about the triplet saddle points) and bimolecular reaction channels (reactants H + HO 2 and…”

“…In this case, the higher-accuracy energy data generated at the MRCI+Q/CBS level of theory were also used during the rate constant predictions for H + HO 2 → H 2 + 3 O 2 using DiNT. 16 Hence, this higher-accuracy data set is a good test-case to assess the capabilities of transfer learning methodologies. Before performing the lower accuracy ωB97X-D functional-based calculations, we also undertook a transfer learning case study using a combination of a classical MD force field and QM calculations.…”

“…Trajectory based rate constant calculation packages such as DiNT, VENUS , or internal codes can be used to train these PESs and calculate the rate constants using trajectories. Jasper et al used DiNT along with permutationally invariant polynomial (PIP) expansions to calculate singlet and triplet rate constants for the H + HO 2 reactions. Jiang et al developed a permutation invariant polynomial neural network (PIP-NN) approach to train the PES for calculating the rate constants.…”