Catalysis of Organic..

312 Catalyst Library DesignCatalyst Library Design ProceduresThe construction of experimental holograms by HRS is described in detail elsewhere[23,24]. In the two-dimensional representation of a multi dimensional experimentalspace the discrete concentration levels of components and modifiers are representedby lines (see Fig.1). The level of each component increases gradually till it reaches itsmaximum then it decreases gradually again. This mode of representation leads towavelike arrangement of levels (see Fig. 1).The elements of symmetry of the experimental space is used to create the initialcatalyst library resulting in 16-48 different catalyst compositions [23,24]. The designof forthcoming generations by HRS has been described in detail in our previousstudies [23,24].Information MiningArtificial neural networks have been used for information mining ANNs provide thequantitative relationship between composition and catalytic performance. ANNsdescribing the objective functions in the given experimental space were previouslytrained with data obtained during HRS optimization. For appropriate formation ofANNs and for evaluation of their predictive ability the available data of each catalystlibrary have been divided into three sets: (i) training, (ii) validation and (iii) testing inthe following ratios: 70:15:15, respectively. The networks are trained with resilientback-propagation algorithm [26]. Training is stopped if the validation error increasesfor more then two consecutive epochs. Nineteen different network architectures wereinvestigated to achieve acceptable model accuracy [26]. Every neural networkarchitecture has been trained 1000 times (each training has been initialized withdifferent, random node-to-node weights) [26]. According to the average mean squareerrors (MSE) the resulted 19000 networks were ranked. The best 100 networks havebeen involved into Optimal Linear Combination [27], during which so called OLCnetworkhas been created. The resulting OLC-network has been applied in this studyfor "virtual" catalytic tests."Virtual" catalytic testsTwo optimization tools can be used for "virtual" catalytic experiments: (i) HRS andGenetic Algorithm (GA). We have recently demonstrated [28] that HRS is a fasteroptimization tool than the GA. The only advantage of GA with respect to HRS is thatGA uses a continuous experimental space, while HRS makes use of levels.In "virtual" catalytic experiments the objective function determined by ANNs isused for optimization, i.e. for finding compositions or experimental parameters withoptimum performance. In "virtual" catalytic experiments "virtual" catalyst libraries arecreated and just using computational methods several catalyst libraries can be virtuallytested, while in the virtual experimental space we are moving towards the virtualoptimum determined by the given objective function. Having found the virtualoptimum one new "real" catalyst library is created in the neighborhood of virtualoptimum. In this way it is possible to accelerate the process of optimization of a givencatalyst library.