Particulate matter removal by precipitation under normal and wildfire conditions across the United States using double machine learning

Precipitation scavenging is an important sink for particulate matter (PM) and can therefore have direct consequences for human and ecosystem health. Controlled experiments and modeling studies quantifying precipitation-driven removal have been conducted in the past, but the literature lacks large-scale observational-based estimates of precipitation influence on PM concentrations. This gap has become more important as wildfire smoke now contributes a growing share of PM pollution across the United States, and climate change drives rainfall pattern shifts. We analyzed hourly PM2.5 and PM10 observations across the contiguous United States from 2014 to 2023 at 760 (PM2.5) and 451 (PM10) Environmental Protection Agency monitors. These records were merged with NLDAS-2 meteorology and NOAA HMS wildfire smoke classifications. We used a machine learning-based causal inference framework (DoubleML) to separate the effect of precipitation from confounding meteorological variables. Particle removal increased with rainfall intensity, particle size, and initial concentration, but declined as wildfire smoke loading increased. Removal rate became increasingly negative up to about 2-3 mm h⁻¹ of rainfall, then declined at higher rainfall intensities even though total removal (i.e., total change in PM concentration from one hour to the next) increased in magnitude. We parameterized the results obtained from the DoubleML model into power law washout coefficients for PM2.5  (a = 0.0404, b = 0.5951) and for PM10 (a = 0.0812,b = 0.4508) to provide observation-based parameters that can be used in chemical transport models. The methodology and parameters can be adopted for human-PM exposure assessment, air quality model calibrations, and PM deposition to land, vegetation, and surface waters.