The Development of Descriptors for Solids: Teaching'Catalytic Intuition' to a Computer

Autor: Catharina Klanner, David Farrusseng, Mourad Lengliz, L. Baumes, Claude Mirodatos, Ferdi Schüth
Rok vydání: 2004
Předmět:
Zdroj: Angewandte Chemie. 116:5461-5463
ISSN: 1521-3757
0044-8249
DOI: 10.1002/ange.200460731
Popis: High-throughput experimentation has become an accepted and important strategy in the search for novel catalysts and materials. However, one of the major problems is still the design of libraries, especially, if vast numbers of catalysts are to be explored. On the other end of the work flow, after catalysts have been tested, data mining and the search for trends is equally demanding. Several methods based on expert systems have been proposed to support the development of solid catalysts. Also the correlation of performance with catalyst composition, evaluated by neural networks, has been used for the optimization of catalysts. For such optimization programs in catalysis, evolutionary algorithms were found to be helpful as well. However, in these approaches the scope was usually very limited, and an attempt to include a wide range of properties to describe the solids was not made. A more integrated “knowledge extraction engine” has been proposed by Caruthers et al. for propane aromatization which is, however, focused on the reaction engineering aspects. There is a great need for software-based methods to plan the design of libraries based on chemical knowledge, in addition to the statistical tools which are implemented in some of the commercial software packages. QSAR (quantitative structure–activity relationship) is one of the most powerful methods used in drug discovery to design libraries and to extract knowledge from tests on possible drug libraries. Such structure–activity relationships are discovered by computer programs, for which molecules need to be represented in computer data bases by so-called descriptors. The descriptors can, for instance, be two-dimensional fingerprints, such as absence or presence of certain chemical functional groups, or can be pharmocophores, which relate to the relative spatial arrangement of three selected chemical functional groups, or physico–chemical properties, or many others. Whole journals are by now devoted to this topic. However, owing to the different nature of the problem, a transfer of descriptor concepts to solids has not been possible to date. In contrast to molecules, a solid can not easily be represented in a computer, since no structural formula can be given and encoded. If only the composition of a solid would be used, many important factors would be lost, since properties of a catalyst, for instance, are also very much dependent on the synthesis and the conditions of the reaction itself. There is one preceding study in which descriptors have been used to correlate structural features of zeolites with the ring size in the structures. Recently, we suggested a methodology to apply to heterogeneous catalysis which was expected to work in a similar manner to the molecular descriptors. It can to some extent be considered to be a multidimensional version of the volcano principle known in catalysis for decades. Herein, we show that these concepts can indeed be implemented. The descriptors thus developed have predictive power and the concept can therefore be considered as the transfer of “catalytic intuition” to a computer. The method, in short, consists of the creation of a library of solids, testing of the performance in a catalytic reaction, description of the solids in terms of a multitude of attributes which are available either from the synthesis of the solid or from tabulated physico–chemical data, and, finally, the identification of a set of those attributes which allow discrimination between different catalytic performance. A highly diverse library was synthesized, consisting of 467 different catalysts. Diversity in this case was judged by chemical intuition, based on the accumulated knowledge in the field. The library included binary oxides, multinary mixed oxides, supported catalysts on different support materials with various supported compounds, zeolites, and many other types. All the catalysts of this library were tested in the oxidation of propene with oxygen (O2:C3H6= 5:1, that is, slightly above stoichiometric for total oxidation) in a 16-fold parallel reactor which was a more advanced stainless steel version of the system described by Hoffmann et al. Products were analyzed sequentially by GC, which allowed the detection of about 30 products. Each catalyst was measured twice, which also allowed its temporal behavior to be analyzed, at five temperature levels (200, 250, 300, 400, and 500 8C). In this way 120 parameters, that is, conversions, selectivities to 21 products, temporal behavior, and carbon mass balance, all at each temperature, were generated for every catalyst. This set of data is too vast by far for a meaningful attempt at a correlation. We have thus classified the catalytic performances into distinct groups with respect to an analysis of the 120 output parameters, using principal components analysis and then clustering techniques based on euclidian distance. Figure 1 shows as an example the results of a tree cluster analysis. Each of the classes can be identified with a specific catalytic performance of the solids, as given in the legend to Figure 1. The other major task was the encoding of the solids. For a virtual screening, only such attributes are useful, which are either derived from a possible synthesis method or are tabulated, so that they do not need to be measured. For each catalyst we have created a set of 3179 attributes, which include the concentrations of 60 elements from the periodic table, 19 attributes which are related to the synthesis method, and 3100 attributes which are taken from tabulated data. These are, for instance, enthalpies of formation of different oxides, possible coordination numbers of the atoms, ionization energies, electronegativities, averages of such values for [*] Dr. C. Klanner, Prof. Dr. F. Sch th Max-Planck-Institut f r Kohlenforschung Kaiser-Wilhelm-Platz 1 45470 M lheim (Germany) Fax: (+49)208-306-2995 E-mail: schueth@mpi-muelheim.mpg.de
Databáze: OpenAIRE
Externí odkaz: