6. Reducing Material Consumption in Automated Flow Optimisations
7.1 Future Work
In the examples in this thesis, there was not strict control over the quenching of reactions and it was assumed that the small injection volume into the HPLC mobile phase would result in a low enough concentration to effectively stop the reaction before components were separated on the column. However the work-up and isolation of a product is incredibly important and can be an area where impurities are created, increased or purged. The ease at which multiple sets of reactors and downstream equipment can be combined in continuous reactors means that it should be easy to include work-up procedures as variables in an optimisation. For example, the flow rate of a quench pump can change the molar equivalents of quenching reagent and physical mixing of that quench,
which in turn can change the impurity profile of the product. In other processes it could be possible that a reaction is carried out in batch then purified using continuous distillation or purification and there doesn’t yet exist any automated optimisation of these processes. In summary, it would be highly advantageous to the pharmaceutical industry to see more example of self-optimised unit operations and not just reactions.
Solubility is a problem that haunts many flow chemists and the requirement for homogeneity can drive researchers away from particular reactions or substrates.95 There have been reactors designed to incorporate solid formation and can result in niche reaction conditions, not achievable in batch.261 However for these reactors to be adopted in self-optimisation, there needs to be techniques to sample reaction slurries for online analysis.
Another area of improvement is in the area of discrete variable optimisation. Continuous variables (e.g. concentration, time, temperature) are easy to optimise for by changing pump flow rates and reactor temperature. However it is difficult to screen which is the best catalyst, solvent, base or a combination of all using optimising algorithms because they need to correlate to a number. In work by Reizman and Jensen, multiple FEDs were carried out with each different discrete variable then steepest descent algorithm optimised the continuous variables.20, 21, 185 However there was little intelligent initial screening of the discrete variables and it relied heavily on high throughput screening, where multiple reactions of different discrete variables are carried out then the results compared.227, 262
Principal component analysis (PCA) is a technique that separates the properties of multiple discrete variables into two or three principal components.263 Murray et al. used PCA to transform two Pd sources, nine phosphine ligands, four bases and nine solvents into a three dimensional experimental space.264 With a 3-dimensional space, a series of DoE designs were used to hone in and then find the best combination of catalyst, base and solvent for a Buchwald-Hartwig sulfamidation, reducing the number of experiments from a possible 51 million to only 78. It is surely sensible to use an optimising algorithm to concurrently search the discrete variables and optimise the continuous variables with a tandem PCA-algorithm technique.
The difficulty in screening multiple discrete variables is the reactor setup. If the reactor setup in this thesis was adopted, where each continuous variable is controlled by a piece of equipment,
then each solvent and reagent would need its own pump or a series of switching valves connected to a single pump, assuming everything is liquid. For solid reagents, all combinations of reagent in solvent would need to be prepared, requiring a very large number of pumps. E.g. four solid reagents in four different solvents would require 16 separate pumps. Assuming that there is the physical laboratory space to connect all this equipment, the costs associated would be very high, probably too high for an academic environment.
Therefore the sensible approach would include robotics, with solid and liquid metering systems to prepare pump solutions of various compositions in order to test in the reactor. The disadvantage with this approach is the technological expertise required to operate robotics. The majority of researchers in self-optimisation are chemists/chemical engineers that have learnt to code laboratory equipment as part of their project requirements. Introducing robotics opens up a whole new field of electronic engineering which can be daunting to an adept coder, let alone a novice.
This also fuels the final requirement – a commercially available self-optimisation platform. I.e. a software/hardware package that can be bought and used to integrate all sorts of pumps, reactors and analysis with intelligent program control and minimising algorithms. It has been mentioned a few times in this chapter how equipment is more readily integrated if it is commercially available and does not need specialist individual manufacture. This alone would contribute to the widespread use of automated flow reactors but would be very difficult to implement. There are several different companies that make analytical equipment, and they all use different software programs. This would require collaboration between companies in order to make it easy to integrate all their software together. Protection of intellectual property is paramount to a company’s success and integration of multiple forms of technology into one platform is not just a problem for chemical consumers.