Applying Machine Learning for Firebrand Production Prediction

Abstract

This article presents a machine learning (ML) based metamodeling framework for firebrand production prediction. This framework was implemented to predict the firebrand areal mass density (FAMD) and firebrand areal number density (FAND) of landing firebrands using a large set of data from full-scale laboratory firebrand production experiments. The independent variables used in our ML models to predict the dependent variables FAND and FAMD were landing (or travel) distance, wind speed, and fuel type (structural and vegetative fuels). It was demonstrated that the non-linear non-parametric ML model, K-nearest neighbors (KNN), works the best for this purpose. The KNN model predicted discrete FAND and FAMD values with an accuracy higher than 90%. The current ML model can be used to predict locations with high risk of spotting ignition potential. This research is a small step towards the bigger goal of creating a numerical firebrand production simulator.