A machine-learning framework for rapid adaptive digital-twin based fire-propagation simulation in complex environments

摘要

The objective of this work is to illustrate how to algorithmically integrate Machine-Learning Algorithms (MLA's) with multistage/multicomponent fire spread models. In order to tangibly illustrate this process, this work develops a framework for a specific model problem combining: (I) a meshless discrete element "submodel" that tracks the trajectory of airborne hot particles/embers, subject to prevailing wind velocities and updrafts, (II) a topographical "submodel" of the ambient combustible material whereby airborne embers that make contact are allowed to start secondary fires (if conditions are appropriate), combined with ground-based surface spread and burn rates for generating new embers, new updrafts (due to hot air), etc., and (III) a Machine-Learning Algorithm to rapidly ascertain the multi-submodel system parameters that force the overall model to match observations. The submodels compute both ground and airborne hot-ember driven fire propagation, as well as subsequent distribution of debris/soot, which is important for air-quality assessment. The overall framework is designed for use in digital twin technology, which refers to an adaptive digital replica of a physical system, whereby model updates are continuously in near real-time. This necessitates a rapid simulation paradigm that can easily interface with telecommunications, cameras and sensors. The presented framework is designed to run quickly on laptops and hand held devices, with the guiding principle being to make it potentially useful for first-responders in real-time. (C) 2020 Elsevier B.V. All rights reserved.