Development of high performance catalysts for CO oxidation using data-based modeling

Abstract

This paper presents a model-aided approach to the development of catalysts for CO oxidation. This is in contrast to the traditional methodology whereby experiments are guided based on experience and intuition of chemists. The proposed approach operates in two stages. To screen a promising combination of active phase, promoter and support material, a powerful “space-filling” experimental design (specifically, Hammersley sequence sampling) was adopted. The screening stage identified Au–ZnO/Al₂O₃ as a promising recipe for further optimization. In the second stage, the loadings of Au and ZnO were adjusted to optimize the conversion of CO through the integration of a Gaussian process regression (GPR) model and the technique of maximizing expected improvement. Considering that Au constitutes the main cost of the catalyst, we further attempted to reduce the loading of Au with the aid of GPR, while keeping the low-temperature conversion to a high level. Finally we obtained 2.3%Au–5.0%ZnO/Al₂O₃ with 21 experiments. Infrared reflection absorption spectroscopy and hydrogen temperature-programmed reduction confirmed that ZnO significantly promotes the catalytic activity of Au.

Introduction

CO is a highly noxious gas that needs to be removed, usually through catalytic oxidation, in many applications, such as gas masks, indoor air quality control systems and automobile exhaust treatment devices. CO is also known to deactivate the electrocatalysts in hydrogen fuel cells [1]. In addition, CO oxidation is regarded as a general reaction for investigating the activity of oxidation catalysts [2]. Low-temperature activity is preferred in this study, since low-temperature reaction is environmentally friendly and cost effective in terms of equipment requirements.

The catalysts that have been developed for CO oxidation so far may be classified into two categories: (1) noble metal catalysts and (2) transition metal oxide catalysts [3]. Noble metal catalysts often use Pt, Ru, Rh, Pd or Au as active phase, with support materials ranging from Al₂O₃, SiO₂, zeolite and CeO₂. A few catalysts were demonstrated to present high activity, such as Ru/Al₂O₃ [4], Rh/Al₂O₃ [5], Au–ZnO/SiO₂ [6], Au/TiO₂ [7] and Pt/Al₂O₃ [8]. Transition metal catalysts have great potential due to the low cost, though their performance is generally inferior to that of the noble metals. Co₃O₄ [9], MoO₃/CeO₂ [10] and CuO/CeO₂ [11] are the most important transition metal catalysts with demonstrated high activity. In addition, CuO was shown as a cost-effective active phase in CuO–CeO₂/Al₂O₃ [12] and CuO–ZnO/TiO₂ [13] catalysts.

Currently, the development of catalysts is usually based on experiments under the guidance of experience and intuition of chemists. In the presence of multiple factors that affect the activity, usually one factor is varied with other factors being fixed for conducting experiments. Subsequently, this procedure is repeated for each factor to search for the “optimal” performance. This one-factor-at-a-time method has long been recognized as ignorant of the correlation between factors, resulting in ineffective exploration of the factors’ space [14]. As a result, a more rational and systemic approach, based on mathematical models, is needed for the design and optimization of catalysts.

Data-based models (also termed empirical models) play an important role in model-aided catalyst development and interpretation [15], [16], [17], [18]. They are developed purely based on experimental data, with the possibility to incorporate prior knowledge (though not compulsory). Due to limited experimental resources, the procedure for developing a new catalyst usually includes two stages: screening for metal/support combinations and then adjusting the loading of the components. The statistical method of design of experiments (DoE) can be applied to the screening step with limited experiments. At the second stage, DoE, data-based modeling and mathematical optimization approaches can be applied to adjust the loading of catalyst components (and sometimes the reaction conditions). This model-aided technique is also termed response surface methodology (RSM) [14].

The general model-aided process design has been investigated in recent years with emerging applications in catalysis [19], [20], [21], [17], [22], [23], [24], [25], [26], [27]. The main components in these previous reports include quantitative property activity relationship (QSAR), genetic algorithms (GA) and artificial neural network (ANN), generally based on high-throughput experimentation (HTE). When HTE is not available (as in our laboratory) and thus nor are the large amount of data, the principle of QSAR and the use of ANN may be questionable [28].

In the current study, we integrate several state-of-the-art computational methods to search for optimal catalysts for CO oxidation. The catalyst screening stage was facilitated by the application of Hammersley sequence sampling, a space-filling DoE method that has been shown to provide better coverage of design space than traditional DoE methods [29]. The screening experiments suggested to focus on Au–ZnO/Al₂O₃, a catalyst that has not be intensively investigated previously for CO oxidation. To further optimize the catalyst performance, a Gaussian process regression (GPR) model is developed from experimental data to relate the CO conversion to the loadings of Au and ZnO. GPR has been shown to attain both accurate prediction and reliable quantification of its own prediction uncertainty (in terms of variance) [30], [31]. The latter property is especially important for our model-aided optimization strategy, namely maximization of expected improvement (EI), since a large variance suggests that the experimental data around this point are not sufficient to give a reliable prediction and thus more experiments should be allocated. The criterion of EI jointly considers the predictive mean and variance leading to a theoretically guaranteed global optimum [32]. The effectiveness of the optimization methods, in particular the use of GPR and EI, has been demonstrated elsewhere through both experiments and computer simulations [30], [31], [32]. This work is the first to report the application of such an integrated framework for the optimization of catalysis systems. The proposed approach successfully identified a high-performance catalyst 4.9%Au–5.0%ZnO/Al₂O₃. The other contribution of this paper is to demonstrate that the GPR model can also be used to help reduce the loading of Au with marginal deterioration of catalyst activity, a strategy to reduce the cost of catalyst. This is a powerful tool to aid the decision with regard to the compromise between performance and cost in practice.