Optimization and analysis of force field parameters by combination of genetic algorithms and neural networks

In search of a force field description for tripod metal templates, tripodM [tripod = RC(CH2X)(CH2Y)(CH(2)Z); X, Y, Z = PR'R "] force field parameters, p, were optimized by the use of genetic algorithms (GA) with the structure s of ten compounds, tripodMo(CO)(3), serving as the database. It was found that the evaluation of the fitness criterion, based on the root-mean-square deviation (rms) between observed and calculated structures by force field methods, is actually the time-consuming step under this optimization protoc ol. It is shown now how this time-consuming step may in part be substituted by using a trained neural network (NN) as the evaluating function. The net work is trained on the basis of parameter vectors that have been evaluated previously with respect to their corresponding rms values during several pr eceding generations of a GA run. The network function, rms = f( p), thus bu ilt up is able to calculate the rms corresponding to a specific parameter v ector within milliseconds, whereas obtaining the same result by molecular m echanics methods takes several minutes for the problem at hand and with the equipment used. Therefore, significant time savings may be expected using a combination of GA optimization and NN simulation In addition, the simulat ed function, rms = f( p), allows for insights into the dependence of the rm s value on specific parameters or combinations thereof. Kohonen mapping is used as a tool to visualize such dependence. (C) 1999 John Wiley & Sons, In c.