Construction Of Potential-energy Surfaces For Large Molecular Systems Based On Neural Networks

Molecular dynamics is an important molecular simulation method,which is playing an increasingly important role in many fields,such as physics,chemistry,biology and pharmacology.Potential-energy surface is the foundation of performing molecular dynamics simulation,from which the energy and gradient of the system can be acquired.Approaches to the construction of potential-energy surfaces include fitting empirical and semi-empirical formulas,interpolation and neural networks,etc.Among all these approaches,neural networks receive much attention due to its flexible functional forms and high fitting accuracy.In chapter 2,we introduce the development and theories of neural networks and how we can construct a potential-energy surface using neural networks.In chapter 3,we construct a potential-energy surface for linear H3 using empirical formula and neural networks respectively.Our fitting dataset consists of 5151 configurations whose energy is calculated using completely renormalized coupled cluster CR-CC(2,3)method.Regarding the fitting method using empirical formula,London potential is adopted.The fitting error is 0.426 kcal mol-1 and the saddle point lies at 1.762 a0.As for the method using neural networks,a(2-10-1)network is adopted.The fitting error is 0.0327 kcal mol-1 and the saddle point lies at 1.758 a0.The results indicate that neural networks yield smaller fitting error and saddle point closer to that of previous work(1.757 a0).This shows that neural networks is able to construct potential-energy surfaces with higher accuracy.In chapter 4,we introduce a low scaling quantum chemistry program(LSQC),which implements the generalized energy-based fragmentation(GEBF)and.the cluster-in-molecule(CIM)algorithm.LSQC is capable of computing the energies of large systems at a low scaling accurately.The program first divides the target system into several sub-systems according to the GEBF and the CIM algorithm.Then conventional calculation is performed for all the sub-systems.Finally results for the target system are acquired by combining those of the sub-systems.In chapter 5,we construct a potential-energy surface for the water cluster(H2O)48 using neural networks.We perform a molecular dynamics simulation for the cluster and 100000 configurations in the trajectories are sampled to form the fitting dataset.First,a potential-energy surface is constructed using a neural network with a 20-neuron hidden layer.The fitting error is 2.464 kcal mol-1.To achieve higher accuracy,we propose a method called the partition method.We set-400 kcal mol-1 as the boundary and divide the whole configuration space into two partition.For each partition a neural network potential-energy surface is constructed.Then a neural network classifier is trained to label any given configuration.The final potential-energy surface is acquired by combing these three neural networks,of which the fitting error is 2.221 kcal mol-1.This suggests that the partition method can effectively improve the accuracy of the potential-energy surface.