The Construction Of Water Models Based On BP Neural Network And Genetic Algorithm

Water is a ubiquitous substance in nature.With the development of micro/nano-technology,a large amount of water has been found inside the micro/nano-scale confinements,such as the soil,bones,leaf stems and other natural materials.The relevant properties and behaviors of water inside the small confinement are significantly different from those at macroscale.Numerical simulation of the confined water is rather important for understanding the relevant laws and mechanisms.The high-precision water model keeping comprehensive balance among the precisions on the various physical properties of water is the fundamental guarantee for the validity of simulations.However,the construction workflow of the existing water model is relatively complicated and often requires an iterative update with numerical simulations which results in an expensive cost of computation.Meanwhile,it is also a great challenge to maintain a comprehensive balance between the various physical properties of water under limited model parameters due to the simple structure of water molecule.Aiming at the above problem,in this dissertation,two high-precision water models are proposed on the basis of molecular dynamics simulation with the combination of BP neural network and genetic algorithm.In terms of the reparametrized strategy,based on the four-site rigid water model TIP4P,the charge q,the distance d from the massless point to the oxygen atom,and the parametersσandεof the Lennard-Jones potential are set as the optimized variables.The mean absolute percentage errors of the physical property densityρ,the enthalpy of evaporationΔH_vap,the self-diffusion coefficient D and viscosityηis the optimization objective.First of all,about 6000groups with random model parameters q,d,σandεwithin the specified range and the physical properties of pure water at 1 atm and 273 K,283 K,298 K,323 K,343 K,373 K are comprehensively conducted as the training data for constructing BP neural network.Then,an effective mapping is constructed between the parameters and the four crucial physical properties of water.Without additional time-consuming MD simulations,this mapping could result in sufficient and accurate data for genetic algorithm to optimize the model parameters as well as possible.Based on the proposed parameterizing strategy,a conventional four-site water model TIP4P-BG and an advanced model TIP4P-BGT with temperature-dependent parameters are finally established.In terms of the simulation results,the physical properties of the two constructed models are calculated,which are also compared with the results from the TIP4P-Ew,TIP4P/2005 and TIP4P/εmodels.For the density,vaporization enthalpy,self-diffusion coefficient and viscosity of the above five models,it is found that the mean absolute percentage errors of the TIP4P-BG and TIP4P-BGT at 298 K are 3.53%and 3.08%,respectively,while the mean absolute percentage errors of the TIP4P-Ew,TIP4P/2005 and TIP4P/εare 6.08%,4.98%and 4.17%,respectively.For the trend of density with temperature,the TIP4P-BGT exhibits closer results with the experimental values relative to the TIP4P/2005 and TIP4P/εmodels,but the TIP4P-BG slightly underestimates the density at the temperature high than 313 K,which is similar to TIP4P-Ew model.For the vaporization enthalpy and self-diffusion coefficient,the TIP4P-BG and TIP4P-BGT models are close to the experiment and are more excellent than the other three models.For the viscosity,the TIP4P-BG and TIP4P-BGT models underestimate the experimental values and are better than TIP4P-Ew model,but the differences decrease gradually with the increase in temperature.Moreover,the temperature of maximum density of the two constructed models are close to the experimental value of 277 K,the isothermal compressibility and the thermal expansion coefficient are close to the experimental values of45.3×10^-6 atm^-1 and 2.572×10^-4 K^-1,respectively.The radial distribution function and surface tension are also in good agreement with the experimental values.In summary,the TIP4P-BG and TIP4P-BGT models in this dissertation show excellent performance,and not only have a reasonable balance between the simulation accuracy of the four crucial physical properties but also agree well with the experimental values of the other five physical properties.The research provides high-precision models for the numerical simulation of confined water.In addition,the proposed procedure provides an effective and easy-to-implement method for the parameterization of molecular dynamics force field models of other materials.