Designing formulations of bio-based, multicomponent epoxy resin systems via machine learning
Rodrigo Q. Albuquerque, Florian Rothenhäusler, Holger Ruckdäschel
Abstract
Abstract Petroleum-based epoxy resins are commonly used as a matrix in fiber-reinforced polymer composites. Bio-based epoxy resin systems could be a more environmentally friendly alternative to conventional epoxy resins. In this work, novel formulations of multicomponent, amino acid-based resin systems exhibiting high or low glass-transition temperatures ( $$T_{{\text{g}}}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:msub><mml:mi>T</mml:mi><mml:mtext>g</mml:mtext></mml:msub></mml:math> ) were designed via Bayesian optimization and active learning techniques. After only five high- $$T_{{\text{g}}}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:msub><mml:mi>T</mml:mi><mml:mtext>g</mml:mtext></mml:msub></mml:math> experiments, thermosets with $$T_{{\text{g}}}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:msub><mml:mi>T</mml:mi><mml:mtext>g</mml:mtext></mml:msub></mml:math> already higher than those of the individual components were obtained, pointing out the existence of synergistic effects among the amino acids used and confirming the efficiency of the theoretical design. Linear and nonlinear machine learning (ML) models successfully predicted $$T_{{\text{g}}}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:msub><mml:mi>T</mml:mi><mml:mtext>g</mml:mtext></mml:msub></mml:math> with a mean absolute error of 3.98 $$^{\circ }{\text{C}}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:msup><mml:mrow/><mml:mo>∘</mml:mo></mml:msup><mml:mtext>C</mml:mtext></mml:mrow></mml:math> and $$R^2$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:msup><mml:mi>R</mml:mi><mml:mn>2</mml:mn></mml:msup></mml:math> score of 0.91. A price reduction of up to 13.7% was achieved while maintaining the $$T_{{\text{g}}}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:msub><mml:mi>T</mml:mi><mml:mtext>g</mml:mtext></mml:msub></mml:math> of 130 $$^{\circ }{\text{C}}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:msup><mml:mrow/><mml:mo>∘</mml:mo></mml:msup><mml:mtext>C</mml:mtext></mml:mrow></mml:math> using an optimized formulation. The LASSO model provided information about the dependence of $$T_{{\text{g}}}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:msub><mml:mi>T</mml:mi><mml:mtext>g</mml:mtext></mml:msub></mml:math> on the number of active hydrogen atoms and aromaticity. This study highlights the importance of Bayesian optimization and ML to achieve a more sustainable development of epoxy resin materials. Impact statement This article shows how the sustainability of epoxy resin systems (ERSs) can be significantly improved by combining experimental and theoretical strategies. First, amino acids are used as curing agents in multicomponent formulations to produce bio-based ERSs. Second, the number of trial-and-error experiments required to obtain formulations with high or low glass-transition temperatures ( T g ) is greatly reduced using machine learning (ML) strategies to design all experiments. Not only is it shown how T g can be maximized in only five new theoretically designed formulations, but the economic advantages of the proposed approach are also discussed. The trends between T g and the type of optimized biocomponents are discussed based on the unambiguous interpretation of the best-trained ML model. The results presented in this study pave the way for the theoretical design of more sustainable polymeric materials. Graphical abstract