Improved Understanding of Experimental Campaigns in Catalyst Development through machine learning

The development of new catalysts is typically carried out by performing extended experimental campaigns of dynamic experiments through high-throughput miniature reactors in which the sequence of the experiment is often managed based on the experience of the scientists and the developers. In these systems, the sequential nature of experiments introduces complex effects that may propagate to successive experimental batches at different conditions which are difficult to model and interpret. Big amounts of data are typically collected from experimental campaigns, which provide the opportunity to develop data-driven models that extract valuable information on the system. In this study, we propose a new machine-learning methodology that allows the in-depth understanding of the experiment dynamics, associated with both the experiment batch itself and the catalyst history (namely, the sequence of multiple experiments performed in different conditions of temperature, composition, etc.). In particular, multiway multivariate latent variables techniques are used to capture the dynamic within the single experimental batch and the high auto-and cross-correlation between variables, two-dimensional dynamic modelling is used to deal with the dynamics of the catalyst history and orthogonalization is used to remove information redundancy. The methodology is validated in the case study of the development of catalyst for ammonia production. We show that the model captures the correlation between variables which describe the reaction kinetics and thermodynamics within each experimental batch, as well as the influence of catalyst history, especially in terms of feed composition. Furthermore, the model captures the contributions of both the dynamics of the single experimental batches and the catalyst history, ensuring very good predictive performance on the ammonia productivity.