A dynamic knowledge graph approach to self-driving chemical laboratories

The contemporary design of self-driving laboratories faces difficulties in scalability and interoperability when it comes to the vision of a globally connected research network. This is due to heterogeneous data formats and resources as an obstacle to holistic integration. This thesis investigates a potential solution to the interoperability problem in chemical experiments by utilising a dynamic knowledge graph to unify the representation of data, software, hardware, and workflow. The developed approach is applied to a few selected case studies. To realise a self-driving chemical laboratory, the design-make-test-analyse cycle is first reformulated as the process of propagating information through a dynamic knowledge graph by means of a chain of actions. Our approach utilises ontologies to capture the data and material flows involved in the experimentation, and employs autonomous agents as executable knowledge components to carry out both computational and physical tasks. The iterative workflow is automatically managed by a derived information framework with data provenance semantically preserved following the FAIR principles – Findable, Accessible, Interoperable and Reusable. The derived information framework is also applied to an automated flood impact assessment in the smart cities domain to demonstrate its generalisability. On the computational front, the dynamic knowledge graph approach is applied to automate the calibration of a kinetic reaction mechanism, which demonstrated a reduction of calibration time from months when done manually to days while an increase in accuracy measured as a 79% decrease in objective function value. In the wet lab, we demonstrate the practical application by linking two robots in Cambridge and Singapore to achieve a collaborative closed-loop optimisation for an aldol condensation reaction in real time. The two robots effectively produced a Pareto front for the cost-yield optimisation problem over the course of three days of operation. The dynamic knowledge graph approach is also applied to optimise two Suzuki coupling reactions for efficient synthesis of challenging molecules, obtaining 82 mg of the final product from a 40 mL scale-up that would be otherwise difficult to access without extensive process redesign and manual synthesis efforts.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2023

Author dc:creator

• Bai, Jiaru

Advisor dc:contributor.advisor

• Kraft, Markus

• AI for science
• Data provenance
• Digital twins
• Dynamic knowledge graph
• FAIR data
• Multi-agent systems
• Ontologies
• Self-driving laboratories
• Workflow management systems

Licence

• https://apollo8-f-pro.lib.cam.ac.uk/bitstreams/12885087-0174-4b84-8e11-652ab0c534d2/download
• https://creativecommons.org/licenses/by/4.0/

Language dc:language

eng

Author Identifier

0000-0002-1246-1993

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/364775