Calculation Of The Fluids' Radial Distribution Functions And Vapor-Liquid Equilibrium By Using Artificial Neural Networks

Radial distribution functions (RDF), which is a key function for describing the structure of fluids, plays an important role in determining equilibrium and non-equilibrium thermodynamic properties of fluids. It is also the primary linkage between macroscopic thermodynamic properties and interaction forces among the molecules in pure fluid and fluid mixture. The theory of RDF is one of the most active research aspects in chemical thermodynamics and the most precise section among the theory of liquids as well. How to obtain the RDF is the key of the theory. At present, molecular simulation, integral methods and micro-phase equilibrium method are the main calculation ways to obtain the RDF. The results of RDF obtained from molecular simulation are best, but the calculating time is very long. Both integral methods and micro-phase equilibrium method have some disadvantages. In this paper, RDF of fluids was calculated by use of BP neural network. RDF data by Monte Carlo (MC) simulation method were chosen as training samples of BP network to design hard-sphere (HS), square-well (SW) and LJ potential functions' RDF calculating models. Some RDF data for these potential models were used to test the prediction ability of BP networks, and the results were better than that of integral methods. The calculating time was less than that of any other ways. Results show BP neural network is an effective method in prediction of RDF. Similarly, MC data of Helmholtz free energy, compressibility factor and internal energy of the LJ fluid been chosen as training samples of BP network, the calculation models of these thermodynamic properties were built. BP networks can predict free energy, compressibility factor andinternal energy of the LJ fluid very well.Distillation is perhaps the most widely used separation process in chemical industry. The correct design of the distillation column requires the availability of accurate vapor-liquid equilibrium (VLE) data. In the paper, BP neural network was used to calculate the VLE of multicomponent systems. Choosing experimental VLE data of multicomponent systems as training samples of BP neural network, liquid mole fractions as input, and vapor mole fractions and equilibrium temperatures or pressures as output respectively, BP neural network prediction model of VLE was built. The predicted results of BP networks are all better than that of activity coefficient models for concerned systems. Been used in the systems that activity coefficient models couldn't give well description, BP networks also get satisfactory results. When the training samples are as many as test samples, the prediction results by BP networks are as good as that by activity coefficients models. When the training samples are three times more than prediction samples, the results by BP networks are better than that by activity coefficients models.Choosing experimental VLE data of three constituent binary systems as training samples, BP neural network prediction models of VLE for ternary systems could been built. The networks predict VLE of the ternary systems quite well, and the results are as good as traditional thermodynamic models. Similar to ternary system, the network prediction models of VLE for quaternary systems were built through choosing experimental VLE data of four constituent ternary systems or six constituent binary systems as training samples. Some VLE data for the quaternary system were calculated respectively, and the results were both satisfactory, and the deviations as little as that by activity coefficient models.Network model was simpler than traditional thermodynamic models because it needn't consider the mixing rules and calculate the interaction parameters.The results show that BP neural network is an effective method in RDF and VLE calculation.