Repository logo

Combining artificial intelligence and robotic system in chemical product/process design



Change log


Cao, Liwei 


Product design for formulations is an active and challenging area of research. The new challenges of a fast-paced market, products of increasing complexity, and practical translation of sustainability paradigms require re-examination the existing theoretical frameworks to include the advantages from business and research digitalization. This thesis is based on the hypotheses that (i) new products with desired properties can be discovered by using a robotic platform combined with an intelligent optimization algorithm, and (ii) we can the connect data-driven optimisation with physico-chemical knowledge generation, which will result in a suitable model for translation of product discovery to production, thus impacting on the process development steps towards industrial applications. This thesis focuses on two complex physicochemical systems as case studies, namely the oil-in-water shampoo system and sunscreen products.

Firstly, I report the coupling of a machine-learning classification algorithm with the Thompson-Sampling Efficient Multi-Optimization (TSEMO) for the simultaneous optimization of continuous and discrete outputs. The methodology was successfully applied to the design of a formulated liquid product of commercial interest for which no physical models are available. Experiments were carried out in a semi-automated fashion using robotic platforms triggered by the machine-learning algorithms. The proposed closed-loop optimization framework allowed to find suitable recipes meeting the customer-defined criteria within 15 working days, outperforming human intuition in the target performance of the formulations. The framework was then extended to co-optimization of both formulation and process conditions and ingredients selection.

Secondly, I report the methods for the identification of new physical knowledge in a complex system where a prior knowledge is insufficient. The application of feature engineering methods in sun cream protection prediction was discussed. It was found that the concentration of UVA and UVB filters are key features, together with product viscosity, which match with the experts’ domain knowledge in sun cream product design. It was also found that through the combination of feature engineering and machine learning, high-fidelity model could be constructed. Furthermore, a modified mixed-integer nonlinear programming (MINLP) formulation for symbolic regression method was proposed for identification of physical models from noisy experimental data. The globally optimal search was extended to identify physical models and to cope with noise in the experimental data predictor variables.The methodology was proven to be successful in identifying the correct physical models describing the relationship between shear stress and shear rate for both Newtonian and non-Newtonian fluids, and simple kinetic laws of chemical reactions.

The work of this thesis shows that machine learning methods, together with automated experimental system, can speed-up the R&D process of formulated product design as well as gain new physical knowledge of the complex systems.





Lapkin, Alexei


machine learning, closed-loop optimization, artificial intelligence, formulated product design, automated experimental platform, physical model identification, feature engineering


Doctor of Philosophy (PhD)

Awarding Institution

University of Cambridge