(288b) Optimisation of Formulations Using Robotic Experiments Driven By Machine Learning Doe

Liquid formulated products have a large number of applications in chemical, energy, and environmental industries, and include a wide range of products such as pharmaceuticals, food, fuels, cosmetics and personal care products, polymers etc. They are produced by mixing different ingredients together in order to achieve a set of desired product characteristics [1]. The nature of the formulated product design problem is diverse and multidisciplinary [2]. It includes many issues, such as product design, process design, economic and environmental analyses [3]. With the global business environment striving for shorter time-to-market for products, there is a need for faster development cycles in the chemical and formulation industries [2].

The automation of chemical experiments and advances in machine learning algorithms to guide automated experiments have recently emerged as a new paradigm for chemical R&D exemplified by several pioneering applications: robotic experimental platform for nanomaterial discovery [4], design of experiments for high-dimensional statistical learning [5], and discovery of materials and process conditions through machine learning [6].

In this work, we present such a closed-loop optimization system for the multi-objective optimization of a commercial formulated product. The target was to obtain a clear homogeneous formulation within a certain viscosity range, minimizing the cost of the adopted ingredients. A classification model based on Bayes reasoning was built, to predict the stability of the investigated formulation as a binary variable. Shannon’s entropy was introduced in the active sampling algorithm as a quantification of uncertainty. This method can lead to exploring the boundary area of the two classes, and therefore increase the stability prediction accuracy. A Gaussian process-based regression model was developed to predict the continuous performance criteria of the formulation, namely viscosity, turbidity, and price. The TSEMO algorithm [5] was used to solve a multi-objective optimization problem. The algorithm searched for optimal solutions globally, with the aim of identifying the Pareto front. Although the approach used did not guarantee global optimality, the use of large sample and generation numbers in the underlying NSGAII algorithm and of Thompson sampling ensured that the surrogate-based optimization algorithm searched for optimal solutions. The interaction between a classification algorithm and an efficient continuous-variables multi-optimization algorithm enabled to simultaneous meet discrete, i.e. stability, and continuous, i.e. viscosity, turbidity, and price, targets at the same time. The proposed methodology enabled to find suitable solutions within a relatively short time, using only a very little empirical pre-knowledge about the physical system to define the constraints of the input variables. This makes the proposed pipeline particularly suitable for the early stages of the product design and can be used by operators without any particular expertise. This would ideally lead to shortening the development and product release time with consistent increase in profits.

References

[1] K.Y. Fung, K.M. Ng, L. Zhang, R. Gani, A grand model for chemical product design, Comput. Chem. Eng. 91 (2016) 15–27. doi:10.1016/j.compchemeng.2016.03.009.

[2] L. Zhang, K.Y. Fung, C. Wibowo, R. Gani, Advances in chemical product design, Rev Chem Eng. 34 (2018) 319–340. doi:10.1515/revce-2016-0067.

[3] S.K. Chandrasegaran, K. Ramani, R.D. Sriram, I. Horváth, A. Bernard, R.F. Harik, W. Gao, The evolution, challenges, and future of knowledge representation in product design systems, Comput. Des. 45 (2013) 204–228. doi:10.1016/j.cad.2012.08.006.

[4] H.N. Miras, G.J.T. Cooper, D.-L. Long, H. Bögge, A. Müller, C. Streb, L. Cronin, Unveiling the transient template in the self-assembly of a molecular oxide nanowheel., Science. 327 (2010) 72–4. doi:10.1126/science.1181735.

[5] E. Bradford, A.M. Schweidtmann, A. Lapkin, Efficient multiobjective optimization employing Gaussian processes, spectral sampling and a genetic algorithm, J. Glob. Optim. (2018) 1–33. doi:10.1007/s10898-018-0609-2.

[6] C. Houben, A.A. Lapkin, Automatic discovery and optimization of chemical processes, Curr. Opin. Chem. Eng. 9 (2015) 1–7. doi:10.1016/J.COCHE.2015.07.001.