Uncertainty-driven Automatic Construction Of Potential Energy Surfaces Of Gas-surface Reactions And Its Applications

In this thesis,we propose an efficient uncertainty-driven active learning sampling algorithm that is adapted for neural networks(NNs)to automatically construct full-dimensional PESs.The basic principle of the algorithm is to define energy-dependent negative of squared difference surfaces with two different NNs as uncertainty space representation of PESs.Optimization algorithm is used to search for local minima to add new ab initio data points to achieve efficient sampling.We first tested the algorithm by constructing full-dimensional PESs for gas-phase H3 and OH3 reaction systems,which could reproduce their quantum reaction probability curves well with very little data points.The algorithm was found to have the advantages of fewer parameters,high stability,high efficiency,and transferability.Gas-phase molecule-surface reaction systems contain numerous surface atoms,which makes it challenging to construct full-dimensional PESs that consider surface atomic degrees of freedom(DOFs),regardless of ML or data sampling methods.We linked the algorithm to Vienna Ab initio Simulation Package(VASP)and embedded atom neural network(EANN)methods to achieve fully automated construction of elementary reaction potential energy surfaces for gas-phase surface systems.In addition,we proposed an efficient method suitable for high-dimensional systems to identify equivalent symmetric local minima using atomic descriptors.We successfully constructed a global PES for H2 dissociation on the Ag(111)surface using only～150 data points with PBE functional,demonstrating that the algorithm still exhibits high adaptability and efficiency in high-dimensional gas-phase surface reaction systems.The EANN PES accurately reproduces the experimental sticking probability So of D2(v=0,j=2)and D2(v=1,j=2)on Ag(111)surface with varying translational energies at 570 K.Considering surface DOFs,So is more consistent with the experimental results in the low translational energy region than quantum dynamics results,demonstrating the importance of accurately describing surface atomic motion.The equilibrium kinetics of highly exothermic adsorbates have been widely studied in both experimental and theoretical fields.The so-called hot atom diffusion on metal surfaces,which is induced by the large kinetic energy gained from strongly exothermic reactions,can significantly affect subsequent reaction processes.Due to the broad configurational space involved in this process,it poses a significant challenge for theoretical investigations.In particular,recent work by Meyer and Reuter revealed the sensitivity of this dynamic process to supercell size using the QM/Me model.In this thesis,the uncertainty-driven active learning algorithm is used to automatically construct an EANN PES that includes the hot atom diffusion channel.The atomic-representation EANN method accurately learns the atomic environment in a supercell with edge lengths and heights greater than twice the cutoff radius,making the PES capable of accurately predicting potential energy in larger supercell.Using this EANN PES,the hot atom diffusion of O2 on the Pd(100)and Pd(111)surfaces after dissociation is explored,demonstrating the sensitivity of the energy dissipation to supercell size.The classical molecular dynamics simulations reproduce the equilibrium oxygen atom distance distribution in the O2/Pd(111)system,which matches well with the STM experimental data of Rose et al.This can be explained by a random-walk type diffusion mechanism proposed by Bukas and Reuter.In contrast,the statistical results of the O2/Pd(100)system indicate that the ballistic diffusion obtained by Meyer and Bukas using the QM/Me model is actually a special case of the molecular-surface configuration under ideal initial conditions.We obtained equilibrium distance distributions of oxygen atom pairs of 1-5 times the surface lattice constants(SLCs),with 2 SLC accounting for～40%.Accurately simulating the initial conditions of experiments is of great significance for properly describing and understanding the mechanism of hot atom diffusion.