A physics-aware digital twin for correcting moisture-induced material artifacts in PM2.5 sensors

Abstract Deploying LCS sensor networks in moist maritime environments is fundamentally constrained by the existence of non-linear optomechanical interference. Humidity causes inherent systematic bias towards overestimating light-scattering signals, rendering data inaccurate. In order to overcome this problem, a new calibration approach based upon the fundamentals of Physics, which has been tested through the use of a pair of PurpleAir PA-II devices alongside an EPA Federal Equivalent Method (FEM) gravimetric monitor, is presented in this research work. Through combining the principles behind $$\kappa $$ -Köhler hygroscopic particle growth with the Lorentz-Lorenz refractive index mixture law, we model the micro-physical interactions of particles within the lens of the LCS sensors. We show that a relative humidity of 76.16% increases the optical intensity by 59.5%. This effect, known as “Optical Bloom,” can be measured using a Structural Mass-Intensity Divergence Factor ( MIDF ) of 4.33. Essentially, moisture in the air causes the optical signal to overestimate the mass increase due to accumulation of moisture around the core of the aerosol. In order to correct for such an artifact in real-time, a Polynomial-Enhanced Random Forest Regression (PERFR) approach is applied within a multi-dimensional Interaction Manifold ( $$\Gamma $$ ). Based on our findings obtained from a 150-day longitudinal field study of LCS data within an extremely industrialized area (EPA site ID: 220330009), PERFR is successful in removing systematic +17.55% bias present in raw hardware data by bringing the calibrated Normalized Mean Bias (NMB) to 0.13% and achieving $$R^2 = 0.9010$$ .