A hybrid PMF-machine learning approach for resolving urban black carbon sources from multi-wavelength absorption measurements
Abstract. Source apportionment of urban black carbon (BC) is challenging because co-emitted combustion sources share overlapping optical and chemical signatures, and conventional Aethalometer approaches require prescribed absorption Ångström exponents (AAEs). Here we combined positive matrix factorization (PMF/ME-2) with a physically constrained machine-learning model to resolve urban BC into policy-relevant emission sectors. At an urban background site in Dublin, Ireland, PMF/ME-2 separated total equivalent BC (eBC) into fossil-fuel (eBC-FF) and solid-fuel (eBC-SF) factors directly from 7-wavelength observations, yielding AAEs of 1.1 and 2.0 without prior assumptions. A neural-network–linear mixed model then apportioned eBC-FF into traffic and oil-heating contributions, reproducing observed eBC-FF (R² = 0.79) with a coefficient “k” = 0.730 ± 0.036 that shifted by at most 18 % across the predictor sets tested without reordering the three sectors. Traffic, oil heating and solid fuel accounted for 59.2 %, 17.2 % and 23.5 % of total eBC. Collocated ACSM factors for peat and wood burning correlated far more strongly with eBC-SF (r = 0.83, 0.76) than with traffic (0.42, 0.44), independently supporting the solid-fuel assignment of eBC-SF. Solid-fuel BC therefore far exceeded the ~2 % of Dublin households reporting solid fuel as their primary heating fuel, showing that energy statistics alone underestimate this source; the oil-heating estimate remains conditional on the assumed proportionality between the two heating fuels.