Functional polymer composites are applied in various fields in industry and daily lives. To overcome the main weakness of polymer materials, the high sensitivity to fire and other high temperature sources, flame retardancy is requested to the polymer composites. Generally, it is to find the appropriate formulations consisting of different polymer matrix and kinds of flame retardant additives, which may vary in amount and type, to realize the good fire performance of polymer composites. Conventionally, the sample preparation and properties characterization are carried out in lab, which are very time-consuming and cost-ineffective. As an advanced method, the data-driven approach is already widely used in materials developing. Here we use the machine learning algorithms to analysis the collected experimental data, feature their descriptors and build validated models to predict the peak heat release rate and time to ignition of polymer composites in cone calorimeter tests. Details will be discussed in the content. But both models have reached relatively high R2 scores (over 0.7) and low mean absolute errors in testsets. The predicted value of pHRR of a flame retardant HDPE composite is 332.34 , which is comparable to 324.26 from the CCT result. Although the prediction of TTI has relatively bigger error due to the imperfections in data pre-processing and feature engineering, but the potential of applying machine learning in predicting fire performance of polymer composites is still deep and attractive.