Bayesian optimisation of a bio-based biodegradable polyester thermoset

Glass fibre-reinforced polymer composites are widely used in the wind turbine manufacturing industry, and are commonly manufactured using general-purpose unsaturated polyester resin thermosets. However, these resins are obtained from oil-derived aromatic monomers making them both unsustainably sourced and difficult to recycle. Bio-based and biodegradable alternatives exist but have not yet seen widespread commercial success due to their relatively high cost and poor mechanical properties. This paper seeks to address the question of whether bio-based branching thermoset polyesters can be optimised using Bayesian optimisation such that properties correlated with mechanical performance, including extent of polymerisation, might be improved upon. This is achieved by selecting a bio-based and biodegradable polymer (poly(glycerol citrate itaconate)), identifying a desirable objective variable (mass loss under esterification as proxy for extent of polymerisation) and some tunable predictor variables (monomeric stoichiometry), and then applying sequential Bayesian techniques to uncover well-performing samples in an experimentally efficient manner. A Bayesian technique employing the probability of improvement acquisition function discovered the experimental global optimum stoichiometry producing a mass loss of 20.8% in curing polyester samples, which represents 72% of the theoretical maximum, indicating a high extent of polymerisation. Bayesian techniques employing both expected improvement and probability of improvement acquisition functions were simulated 10 000 times across five different emulations of the practical experiment's parameter space. For one particular method (GP-Mat-05), the estimated probabilities of reaching the defined region of global optimality were 65.0% and 23.8%, respectively. We conclude that Bayesian optimisation techniques represent a practically deployable and sample-efficient means by which desirable properties of bio-based thermosets may be improved.