Deutsch
 
Hilfe Datenschutzhinweis Impressum
  DetailsucheBrowse

Datensatz

DATENSATZ AKTIONENEXPORT

Freigegeben

Hochschulschrift

Evaluation of Descriptors for Solids

MPG-Autoren
/persons/resource/persons58693

Klanner,  Catharina
Research Department Schüth, Max-Planck-Institut für Kohlenforschung, Max Planck Society;

Externe Ressourcen
Es sind keine externen Ressourcen hinterlegt
Volltexte (beschränkter Zugriff)
Für Ihren IP-Bereich sind aktuell keine Volltexte freigegeben.
Volltexte (frei zugänglich)
Es sind keine frei zugänglichen Volltexte in PuRe verfügbar
Ergänzendes Material (frei zugänglich)
Es sind keine frei zugänglichen Ergänzenden Materialien verfügbar
Zitation

Klanner, C. (2004). Evaluation of Descriptors for Solids. PhD Thesis, Ruhr-Universität Bochum, Bochum.


Zitierlink: https://hdl.handle.net/11858/00-001M-0000-000F-9689-1
Zusammenfassung
In this thesis, a first concept to derive descriptors for solid catalysts has been evaluated. The concept consists of the following steps:
(1) design of an initial diverse library of solids,
(2) description of the materials with a large number of attributes,
(3) catalytic test of all solids,
(4) data analysis of performance resulting in grouping of the solids in clusters, each contains solids exhibiting similar performance, and
(5) modeling of the correlation between selected attributes and performance clusters, to derive a set of predictive attributes, a “descriptor vector”.
In the first step, a library of 467 solids that was diverse concerning preparation methods and the constituent elements has been designed. Subsequently, this library has been synthesized within collaborations by conventional and high throughput techniques.
Then, a relational data base was evaluated to store the preparation methods and the composition of the materials. In addition, properties of elements, element oxides, and element ions have been inserted into the data base. Furthermore, a concept to compute comparable attributes for all catalysts has been developed and eventually approximately 3000 attributes have been retrieved from the data base. 75 attributes have been selected by chemical intuition to be fed in the correlation procedure.
In parallel, all materials have been tested for the oxidation of propene with oxygen in a high throughput set-up, consisting of a 16-fold plug-flow reactor and gas chromatography as analysis method. Each catalyst was tested at 5 temperatures in the range of 200 °C to 500 °C under identical conditions.
The resulting performances have been analyzed using two different approaches. In the first approach the data from only one temperature has been taken into account (300 °C), while the second approach accounted for all measurement temperatures. With both approaches analogous data analyses have been performed to compute clusters of catalysts exhibiting similar performance. The data treatment consisted of first a feature selection process based on chemical intuition, followed by a principal components analysis, and finally cluster analyses. The “one temperature approach“ yielded two very good classifications, dividing the library into 6 and 4 clusters, respectively. In contrast, the “all temperatures approach“ resulted in a division into 6 or 5 clusters.
In the last step, the 75 selected attributes were correlated with the clusters that had been obtained from both approaches in the performance data analysis. The modeling has been performed applying two different techniques, artificial neural networks and classification trees. First, probabilistic neural networks were used to model the relationship between attributes and clusters, but also to discriminate redundant and useless attributes. With this method the number of attributes could be decreased to approximately 45 (the number varies in the different approaches). Then, multilayer perceptrons were employed, using the selected attributes for each analysis. In all artificial neural network analyses, the prediction quality of the models was better than a random prediction. Finally, classification tree analysis was applied based on the attributes that had been selected by multilayer perceptrons. The second modeling procedure confirmed the results obtained by artificial neural network analyses. All resulting classification tree models were of comparable quality. Considering all approaches, with the ‘all temperatures approach’ (giving 5 clusters) the best models could be obtained. In the end, the attributes used in all modeling analyses were inspected, and 8 attributes of major importance could be identified that were selected preferentially by nearly all models.
Hence, all identified sets of attributes have predictive power and can thus be called descriptor vectors. The predictive power is acceptable as it is clearly above a random prediction, but for an absolutely reliable model the selection of attributes has still to be optimized, and possibly also the list of attributes needs to be expanded.
In a subsequent project, the identified descriptors will be validated by subjecting them to a virtual screening procedure. For this purpose, a new very large library of solids has to be designed in silico. Then, the descriptor vector for all solids in the virtual library will be computed to represent the solids in the descriptor space. Afterwards, clustering techniques will be used to identify clusters of similar virtual solids. From each cluster, a subset has to be selected. These solids will be synthesized and tested in the catalytic oxidation of propene. If the concept works, it is expected that catalysts from the same cluster would have roughly similar performance, whereas catalysts from different clusters should perform differently. Thus, the diversity of the large library should be represented in the selected subset. Using this technique, it could be avoided to test several similar catalysts and the parameter space could be screened more efficiently. Including this tool in the high throughput experimentation loop (see also Figure 3) would dramatically optimize the design stage for discovery libraries.
In addition, it should be tested whether the identified descriptor vector is also valid for other alkene oxidation reactions, for example, the oxidation of butene or maybe even alkane oxidations, to reveal whether the descriptor vector is applicable for oxidation reactions in general. The results of this thesis demonstrate that it is possible to teach “chemical intuition” to a computer program. As a broader and broader data base becomes accessible, better and better predictive power is expected. The toolbox to generate this broad data base is available with the high throughput experimentation technology, and application of the design principles laid out in this thesis may make the high throughput approach in catalysis even more efficient.