An adaptive framework to accelerate optimization of high flame retardant composites using machine learning

Extensive machine learning methods consist of linear and nonlinear algorithms have heralded a sea change in the areas of metals, catalyst, polymers, and so on. However, most of these prevalent researches in polymer fields are focused on molecule design of polymers itself or simulation instead of composition exploration of functional polymer-based composites. The incorporation efforts of machine learning into functional polymer-based composites (in this case, flame retardancy) remain at an elementary stage. Herein, we designed an adaptive framework combining domain knowledge and machine learning to accelerate optimization of high flame retardant composites. Data resources in the adaptive framework were divided into three approaches including experiments, handbooks, and published papers, which were used to train, feedback, or predict ingeniously. The comprehensive feature engineering of flame-retardant polymer-based composites was displayed and classified detailly. Four machine learning methods consist of conventional linear regression (Lasso and Ridge), nonlinear artificial neurol networks (ANN), and their combination of Lasso, Ridge, and ANN (L-ANN) were contrasted in the run of the adaptive framework. Models of limit oxygen index (LOI) by L-ANN method were suggestive of higher accuracy in twice runs, navigating new experiments with high flame retardancy and effective prediction across different flame retardants to tackle intuition-driven trail-and-error problem. The final optimized models from the adaptive framework might be further helpful for machine intelligence of engineering of flame-retardant polymer-based composites. The proposed adaptive framework can be extended hopefully for machine intelligence design of other functional polymer-based composites.