Accelerating Liquid Formulation Design using Lab Automation and Machine Learning

Liquid formulations are ubiquitous yet have lengthy product development cycles owing to the complex physical interactions between ingredients, making it challenging to tune formulations to customer-defined property targets. Methods to accelerate liquid formulation design are desired to address changing customer preferences, supply chain/regulatory pressures, and a drive to develop more sustainable products. This thesis focuses on using lab automation to develop a high-throughput liquid formulation workflow and machine learning (ML) to build property prediction models for a system of shampoo formulations. The methods developed in this work are generalisable to other surfactant-based products in the personal care industry. ML and optimisation have expedited a broad range of product and process development but are critically dependent on the volume and quality of data available. I focused on developing a high-throughput formulation workflow comprised of modular unit operations. Formulation development involves working with viscous materials and often challenging and labourious processing or characterisation steps, e.g., pH adjustment or rheology measurement. Within this work, I present (i) an automated viscous liquid handling protocol using a retrofitted Opentrons OT-2 robot, (ii) a self-driven pHbot for automated titration of viscous liquid formulations, (iii) computer vision for stability prediction, and (iv) a proxy viscometer for Newtonian fluids. I used the developed workflow to collect a dataset of over 800 liquid formulations with phase stability, turbidity, and viscosity measurements. The formulations prepared were a binary mixture of surfactants, a conditioning polymer, and a thickener; selected from a choice of 18 industrial formulation ingredients. Formulation design typically results in a high-dimensional, mixed nominal-continuous design problem for which there was no suitable design of experiments. I developed a weighted space-filling design using Maximum Projection Designs with Quantitative and Qualitative Factors (MaxProQQ). The weighting was from a phase stability classifier trained within an active learning cycle for difficult-to-formulate (unstable) formulation sub-systems to guide them to regions of stability. Finally, I used the generated dataset to develop phase stability, turbidity, and viscosity models. Previous work developed these models based only on the concentration of ingredients, which would not extrapolate to new ingredients. Therefore, I introduced a featurisation based on the surfactant functional groups as an initial step towards generalisation. I additionally explored a set of surfactant molecular descriptors selected based on domain knowledge to improve the quality of the viscosity model. However, I conclude that system-level descriptors are required instead.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2024

Author dc:creator

• Chitre, Aniket

Advisor dc:contributor.advisor

• Lapkin, Alexei

• formulation design
• machine learning
• lab automation
• property prediction
• design of experiments

Licence

• https://www.repository.cam.ac.uk/bitstreams/2ec6dc20-6093-4b31-b3f8-13a4e8329ca7/download
• https://www.rioxx.net/licenses/all-rights-reserved/

Language dc:language

eng

DOI dc:identifier.doi

https://doi.org/10.17863/CAM.113555

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/376242