Testing and training data sets for: A novel representation of time-resolved particle emissions from pyrolyzing wood

Biomass burning is responsible for emitting 90% of total primary organic
aerosols into the atmosphere. Products from pyrolysis are released to the
atmosphere when gas-phase reactions are unable to oxidize them, and even
narrow windows of release can produce a large fraction of the overall
particle emissions. This work introduces an approach to predict particle
emission during high-emitting periods of biomass burning. We trained a
regression-based machine learning model to predict particle emission using
data from pyrolysis experiments covering seven wood types under controlled
conditions. The model considered experimental mass-loss rate (MLR) of the
wood, wood density, and heating conditions as features for prediction of
measured particle-phase real-time emissions. After training, the
experimental mass-loss rate was replaced with modeled MLR from a
two-dimensional finite volume model of pyrolysis to predict particle
emission rate using only wood properties and the boundary conditions of
pyrolysis. The hybrid model explains 80% of the variance in particle
emission and has comparable error metrics with the machine learning model
that relies on experimental MLR. Errors are greatest during the initial
transient where pyrolysis is occurring near the surface.