Polyoxometalates and crystallisation space: an artificial intelligence-assisted exploration and comparison with human intuition

The discovery of new inorganic molecules is an interesting problem since it implies an extended understanding of two contingent events: the first one is the formation of a new molecule, and the second is its crystallisation. The reason for that is that in the case of the product formation (and in order to make a discovery) the conditions under which the building blocks assemble have to be found, while in the case of crystallisation the conditions under which the product aggregates into crystals (which can be a subregion of the potential synthesis conditions) need to be identified in order to be isolated and characterised. There are a vast number of combinations of the experimental conditions and the coordination modes of the transition metals taking part in the building blocks, which means that a full exploration of the chemical space of any given compound would be impossible. As a result, the intuition of highly trained and experienced chemists is required in order to design the appropriate experiments that will determine the right conditions for the isolation of any new products. Unfortunately, intuitions of the experimenters can be biased by both the current knowledge of the field and their frame of mind, which makes important discoveries difficult to achieve. The work presented in this thesis is focused on the field of polyoxometalate chemistry and is exploring a multidisciplinary approach to probe the interaction of artificial intelligence methods with the human intuition during the process of exploring the crystallisation space. Our fundamental difference with relative work in the field is the application of active learning methods (which consist of methodologies capable of deciding what experiments to perform next in order to collect data that will improve the understanding of our system) in contrast to the data mining methods and simulations that have been employed so far. This algorithm method is compared to how human experimenters approach the exploration of the crystallisation space, and their performances are evaluated in terms of prediction accuracies and volume coverage. In this case, the human experimenters are allowed to follow whichever exploration strategy they see fit in order to address the task at hand. Finally, the same algorithm method is extended into collaborating with the human experimenters, and we will study the way the inherent biases can affect the search of the experimental space. This interaction is accomplished with the algorithm suggesting a set number of experiments and the human experimenter selecting the ones they seem appropriate to complete their task