Unveiling MgC2 reaction dynamics using a permutation-invariant-polynomial neural-network potential energy surface constructed via the trust region framework

Recognized as an astrophysical molecule, MgC2 has garnered significant attention. In this study, we employed the DW-SA-CASSCF/icMRCI+Q/AV(Q+d)Z (Dynamic-weighting-state-averaged-complete active space self- consistent field/Internally contracted multi reference configuration interaction with Davidson correction based on augmented valence quadruple zeta with added diffuse and polarization functions) method to compute high- precision ab initio energy points. A neural-network approach was subsequently utilized to fit the global potential energy surface of the MgC2 molecule in its X 1A1 state. In contrast to previous methods, the optimization of neural-network parameters in this study was carried out using the trust region framework (TRF). The TRF algorithm exhibits superior performance in parameter optimization and demonstrates reduced dependency on neuron parameters compared to the Levenberg-Marquardt algorithm. Utilizing the obtained potential energy surface, three reaction pathways for MgC(X 3E-)+C(3Pg) -> Mg+C2 were identified, confirming that this reaction proceeds as a barrierless exothermic process. To further verify the accuracy of the fitted potential energy surface, the reaction cross section was calculated under the initial condition v = 0, j = 0, 20, 50, 80, 120 for this reaction. By integrating the cross section with the relative translational energy, rate coefficients were determined under various initial conditions. Under the initial condition v = 0, j = 0, the rate coefficient is at its maximum, and the trend of the rate coefficient with temperature first decreases and then increases. At a temperature of 100 K, the rate coefficient reaches its minimum value. At low temperatures, the rate coefficient is very high, which is consistent with the characteristics of a barrierless exothermic reaction.