Self-optimisation and model-based design of experiments for developing a C–H activation flow process

Self-optimisation and model-based design of experiments for developing a C–H activation flow process Alexander Echtermeyer1,2, Yehia Amar2, Jacek Zakrzewski2 and Alexei Lapkin*2,§ Full Research Paper Address: 1Aachener Verfahrenstechnik – Process Systems Engineering, RWTH Aachen University, Aachen, Germany and 2Department of Chemical Engineering and Biotechnology, University of Cambridge, Cambridge, United Kingdom Email: Alexei Lapkin* - * Corresponding author § + 44 1223 334796 Keywords: automated reaction system; C–H activation; design of experiments; flow chemistry; process modelling; self-optimisation Open Access Beilstein J. Org. Chem. 2017, 13, 150–163. doi:10.3762/bjoc.13.18 Received: 31 October 2016 Accepted: 05 January 2017 Published: 24 January 2017 This article is part of the Thematic Series "Automated chemical synthesis". Guest Editor: I. R. Baxendale © 2017 Echtermeyer et al.; licensee Beilstein-Institut. License and terms: see end of document. Abstract A recently described C(sp3)–H activation reaction to synthesise aziridines was used as a model reaction to demonstrate the methodology of developing a process model using model-based design of experiments (MBDoE) and self-optimisation approaches in flow. The two approaches are compared in terms of experimental efficiency. The self-optimisation approach required the least number of experiments to reach the specified objectives of cost and product yield, whereas the MBDoE approach enabled a rapid generation of a process model. Introduction The development of manufacturing processes to produce functional molecules, such as pharmaceuticals or fine chemicals, often relies on experience and trial-and-error, rather than on mechanistic process models [1]. The only reason for this is the complexity of chemistry and the duration of time required for the development of good mechanistic models. A game changer in this area is the recently emerged field of automated continuous-flow experiments driven by algorithms for sequential design of experiments (DoE), which significantly reduce the effort in running routine reactions and generating data for opti- misation of reaction conditions [2-7]. An illustration of the concept is shown in Figure 1. Mainly, self-optimisation experimental platforms are used to rapidly obtain optimal reaction conditions using either flow [8-10] or batch experiments [11]. In these cases, the optimisation is driven by the global or target optimisation towards the selected performance criteria. This is rather different from the objectives of model development. In the case of model development, the key criterions are the ability of a model to describe 150 Beilstein J. Org. Chem. 2017, 13, 150–163. Figure 1: A framework of closed-loop or self-optimisation combining smart DoE algorithms, process analytics, chemoinformatics and automated reactor systems. the observed experimental data and to predict process performance under unseen conditions. Thus, experiments required for model development are frequently what would be considered as ‘bad’ experiments in the case of optimisation. A model-development framework has been demonstrated on the basis of an automated microreactor experimental system for several complex reactions [8,12,13]. The framework uses factorial design of experiments to obtain an initial data set for parameter estimation, followed by an iterative search with online model discrimination and parameter estimation, guided by D-optimal design. In a different approach, transient data from continuous-flow experiments were used to identify parameters of a known mechanistic scheme to discriminate between several alternative model structures and to identify model parameters, but no specific design of experiments method was used [14]. The framework proposed in the present publication is using a model-based design of experiments method (MBDoE) [15-17], which incorporates the model with its parameters, as well as details of the experimental setup, such as measurement accuracy and experimental limitations, to design the most informative experiments. This approach requires some model structures to be known a priori which may restrict the methodology to reactions with known mechanism, or to empirical parametric models. A discussion of how a priori knowledge of chemistry, i.e., reaction mechanisms, is included in self-optimisation and model-development frameworks is not well documented in the literature. Very recently we have shown that a priori knowledge in the form of density functional theory level (DFT) mechanistic calculations can be used to propose process models and to perform in silico design of novel flow processes [18]. In this publication, we present an extension of this methodology, in which an initial process model is developed through a MBDoE methodology coupled with an automated self-optimisation flow system. This approach was tested on the Pd-catalysed C–H activation reaction of 1 resulting in the formation of an aziridine 2 (Scheme 1) [19]. The reaction was recently discovered [20] and its mechanism studied [21] and later proven [18]. A simplified mechanism is shown in Scheme 2. In the reaction of interest, the starting material 1, an aliphatic secondary amine, is converted into an intermediate species B in a catalytic first step and Scheme 1: Catalytic reaction scheme showing C–H activation of an aliphatic secondary amine 1 to form the aziridine product 2 [19,20]. Orange rings show C–H and C–N bonds in the substrate and the product, respectively, indicating the location of the C–H activation. 151 Beilstein J. Org. Chem. 2017, 13, 150–163. Scheme 2: A simplified reaction mechanism based on literature [21], showing intermediate B and the side reaction compounds 1∙HOAc and A. The key step includes the C–H activation. 1: starting material, 1∙HOAc: coordinated starting material, Pd(OAc)2: catalyst, 2: product, PhI(OAc)2: oxidant. consecutively transformed to product 2 in the second step, which comprises the C–H activation. In addition to the main reaction pathway, B can form the relatively unreactive resting state complex A, and compound 1 can also form a coordinated species 1∙HOAc upon protonation with a molecule of acetic acid. This limits the formation of A due to reduced concentration of 1. Table 1 gives an overview of the a priori knowledge used in this study. Fast reaction steps were lumped into a single one, containing the critical C–H activation, and described by reaction rate constant k3 in Scheme 2. Empirical information provided constraints of process conditions, such as temperature and concentration ranges, whereas initial values of kinetic parameters were estimated based on a DFT model. Further details can be found in Supporting Information File 1. Here we demonstrate an MBDoE approach on the basis of the model structure and the initial model parameters from DFT calculations and using automated flow experiments. We then use the obtained process model to develo (...truncated)