Since we have stepped into the 21st century, new materials and technologies play more and more significant roles in the construction of human civilization. The flame-retardant polymer composites (FRPCs) have found various applications in human society due to the versatility, ease of processing, customizability, cost-effectiveness and safety. Nowadays, traditional research workflow has been integrated with modern computational technology. It has widely spread and emerged in various traditional disciplines that utilizing high-performance computers to construct the digital models for complex humanities and social sciences, physical and chemical systems, as well as engineering technologies. In the field of materials science, by using machine learning (ML) models, the data-driven method has been used to predict the thermal, mechanical and combustion-related properties of different materials, and recognize the type of polymers and phase of physical/chemical processes, which effectively enhances the research process, reduces the cost and lower environmental impact.

In this study, we employed various ML algorithms to develop accurate, robust, and generalizable models for predicting the results of cone calorimeter test (CCT) and mechanical performance of FRPCs. Initially, we compiled and collected 828 data of related formulations from approximately 150 publications, involving the types of raw materials, selection of processing parameters, and final properties. Based on different flame-retardant (FR) mechanisms, we divided FRPCs data by the FR additives: inorganic metal hydroxide (MH), metal-organic framework (MOF) and layered double hydroxide (LDH), marked as different FR-systems. Mainly, the Random Forest (RF), Support Vector Machine (SVM) and their linear combination were used to build ML frameworks to predict the target properties of FRPCs in different FR-systems.

In the MH-system, RF models achieved coefficients of determination (R2) of over 0.81 in all test sets. Feature screening indicated that the distribution and amount of MH in polymer composites had the greatest impact, followed by the content of effective FR elements in other synergists. The models in MOF-system had predictive accuracy generally exceeded 80%. Model interpretation showed that not only the structure and type of MOFs significantly influenced the performance of FRPCs, but also the properties of polymer matrix and incorporation methods of FR additives were important. At last, LDH models exhibited good predictive accuracy, particularly in predicting Time to Ignition (TTI), where the R2 reached 0.9 and the Area Under Curve (AUC) was also 90%. The feature analysis revealed the key features as organic modification and interlayer spacing of LDH strongly affecting the heat release in CCT. Besides the prediction of separate values, the simulation of HRR development presented the potential of ML framework to replace the physical experiments. The simplified HRR curves containing important development tendencies have been predicted accurately, restored the heat release of FRPCs in CCT qualitatively and quantitatively. Meanwhile, the predictive performance of final ML models involving all kinds of FR additives, including MH, MOF, LDH, intumescent FRs, and other organosilicon, nitrogen-containing chemicals, exhibited the potential of the construction of a comprehensive, all-FR-systems-contained and prediction-oriented big-data model.