Low-cost light-scattering sensors can only measure PM2.5 accurately after calibration against co-located regulatory monitors, and calibrations built in one metropolitan region degrade when applied in regions with different particle composition and meteorology (Zusman et al., 2020). Later work produced a single United States-wide correction for one sensor family that works under typical and smoke-impacted North American conditions (Barkjohn et al., 2021), but the same corrections transfer poorly across aerosol regimes — failing by factors of 5–6 for dust (Jaffe et al., 2023) and losing accuracy outside North America, where region-specific calibrations remain superior (Ginsburg et al., 2026). The open problem is to develop and validate a general calibration framework — models plus transfer/adaptation procedures with bounded local co-location effort — that delivers epidemiology-grade PM2.5 accuracy from low-cost optical sensors in target regions with distinctly different aerosol composition, relative humidity, and concentration regimes, including episodic high-concentration events, with the trade-off between adaptation effort and achieved accuracy quantified.

Background

Fine particulate matter (PM2.5, particles with aerodynamic diameter below 2.5 micrometers) is a leading environmental health risk, and environmental epidemiology needs PM2.5 exposure measurements at spatial and temporal resolutions that sparse regulatory networks cannot provide. Low-cost optical PM sensors promise this coverage: a nephelometer-type sensor (e.g. Shinyei PPD42NS) measures the overall near-forward light scattered by particles in a small chamber, while an optical particle counter (e.g. Plantower PMS series, used inside the widely deployed PurpleAir monitors) converts a laser scattering pattern, via Mie theory (the classical theory of light scattering by spheres), into particle counts per size bin and then a manufacturer-reported mass estimate. Because the scattering-to-mass conversion depends on particle size distribution, refractive index, and density, and because particles take up water at high relative humidity (hygroscopic growth), the raw sensor output is biased in a way that varies with aerosol type and weather. Sensors are therefore calibrated empirically by co-locating them with regulatory reference instruments and fitting a regression — typically with temperature and relative-humidity terms. The regulatory references themselves differ: FRM (Federal Reference Method) monitors give 24-hour integrated gravimetric PM2.5, while FEM (Federal Equivalent Method) instruments such as the TEOM (Tapered Element Oscillating Microbalance) and the BAM (Beta Attenuation Monitor), both continuous automated mass-measurement instruments, give quasi-continuous readings that agree with FRM only approximately.

The multi-city study that generated this question (Zusman et al. 2020) co-located Plantower PMS A003 and Shinyei PPD42NS sensors at regulatory sites in seven US metropolitan regions for roughly two years. Region-specific multivariate calibration models for the Plantower sensor performed well (e.g. cross-validated R² = 0.96, RMSE (root-mean-square error) = 1.15 µg/m³ for daily PM2.5 in Seattle), while the Shinyei sensor was too imprecise to calibrate well (R² = 0.40). However, applying the Seattle model in the other regions gave clearly worse predictions, and the authors concluded that climatological differences between regions limit calibration transferability, recommending region-specific calibration based on at least three co-located reference sites. Episodic high-concentration periods (July 4 fireworks, an August 2018 wildfire episode) had to be excluded from model fitting, and differences in fitted coefficients between regions and sites were attributed to particle composition and density differences that the undocumented manufacturer count-to-mass conversion cannot handle.

Subsequent work sharpened both sides of the problem. A physics-based framework using Mie calculations (Hagan and Kroll 2020) showed mechanistically that the mass error of low-cost optical sensors depends jointly on relative humidity, aerosol optical properties, size distribution, and the mismatch between the aerosol used to calibrate the sensor and the aerosol being measured — explaining why calibrations do not transfer between aerosol regimes. On the empirical side, Barkjohn et al. (2021) fitted a single simple correction (linear in the raw signal with a relative-humidity term) from almost 12,000 co-located daily averages across 16 US states; it reduced RMSE from 8 to 3 µg/m³ nationwide and performed acceptably under smoke-impacted conditions, and it now runs operationally on the US EPA AirNow Fire and Smoke Map. Yet the same evaluation machinery shows the limits of such generalizations: an evaluation across 50 high-concentration pollution events found the US corrections accurate for urban and smoke aerosol (with a roughly 20% underestimate above 600 µg/m³) but underestimating dust-impacted PM2.5 by a factor of 5–6 (Jaffe et al. 2023), and a six-continent assessment of 1,013 co-located PurpleAir sensors found that US-based corrections have limited transferability outside North America, that a fitted global model helps some regions and underperforms in others, and that regional calibrations consistently perform best (Ginsburg et al. 2026). Domain-adaptation and machine-learning approaches that transfer a calibration to a new location with only a short local co-location exist as local demonstrations (e.g. Jha et al. 2021; Dey et al. 2025), and reviews of the field conclude that there is still no established standard calibration procedure and that method choice depends on regional conditions (Zhang et al. 2025). What does not exist is a validated general methodology that makes low-cost sensor calibrations portable across regions with genuinely different aerosol composition and meteorology at a bounded, quantified adaptation cost.

Problem Statement

Develop and validate a general calibration framework under which PM2.5 measurements from low-cost optical particle sensors (nephelometer- or optical-particle-counter-based) reach epidemiology-grade accuracy in target regions whose particle composition, size distribution, density, and meteorology (especially relative humidity) differ distinctly from the region(s) where calibration data were collected. Epidemiology-grade accuracy means accuracy sufficient for the calibrated measurements to be usable, with caution, for exposure measurement in epidemiological studies — the standard contemplated by the source study — as judged against co-located regulatory reference measurements (FRM/FEM) with standard performance metrics (e.g. RMSE, R², and bias) reported over the concentration range relevant to the target region. Consistent with Zusman et al. (2020), the choice of evaluation criteria depends on the purpose of calibration and no single universal pass/fail threshold is prescribed; accordingly, the metrics, the reference class, and the concentration range over which the accuracy target holds must be stated explicitly by the framework and applied consistently across all validation sites, rather than selected post hoc.

The framework must specify (i) the calibration model class, (ii) a transfer or adaptation procedure that, from local co-location data in the target region — a stated number of co-location sites and a stated co-location duration — produces a calibration for that region, and (iii) a quantified trade-off between that local adaptation effort and the achieved accuracy. The deliverable is one general procedure applicable across target regions; it may yield region-specific calibrations, and the local effort it requires may depend on how distinct the target aerosol regime is from the regime(s) where the calibration data were collected, but that dependence must itself be part of the quantified trade-off. Bounded local effort means short local co-location clearly smaller than a full from-scratch regional calibration campaign of the kind conducted in the source study (multiple reference sites co-located over roughly two years); a framework whose declared local effort simply reproduces such a campaign does not answer the question, and the trade-off must be characterized at more than one effort level, so that accuracy is reported as a function of the local co-location budget rather than as a single point. This generalizes the observation of Zusman et al. (2020) that calibrations are region-bound and transferable only under similar meteorological and environmental conditions, and extends the within-US generality demonstrated by Barkjohn et al. (2021) for one sensor family beyond the scope in which a single US-wide correction was shown to apply.

Accuracy must be demonstrated on held-out data against co-located regulatory reference measurements in multiple target regions that are distinct in the aerosol-regime sense used above — differing particle composition, size distribution, and relative-humidity and concentration regimes — at daily resolution and, where the available references support it, hourly resolution. The reference class must be fixed in advance for each validation site: FRM where available, otherwise FEM, with the reference used at each site reported, since FRM and FEM instruments agree only approximately and the accuracy claim must not depend on selecting whichever reference is more permissive at a given site. The demonstration must include episodic high-concentration events (wildfire smoke, dust, fireworks), since these events were excluded from the source study's fitting and are known to stress such corrections (5–6× underestimate for dust under US-wide corrections).

The question is what framework achieves this and with what quantified adaptation cost.

References

  1. Zusman, M., Schumacher, C. S., Gassett, A. J., Spalt, E. W., Austin, E., Larson, T. V., Carvlin, G., Seto, E., Kaufman, J. D., Sheppard, L. (2020). Calibration of low-cost particulate matter sensors: Model development for a multi-city epidemiological study. Environment International 134, 105329. DOI: 10.1016/j.envint.2019.105329. https://pmc.ncbi.nlm.nih.gov/articles/PMC7363217/
  2. Barkjohn, K. K., Gantt, B., Clements, A. L. (2021). Development and application of a United States-wide correction for PM2.5 data collected with the PurpleAir sensor. Atmospheric Measurement Techniques 14(6), 4617–4637. DOI: 10.5194/amt-14-4617-2021. https://amt.copernicus.org/articles/14/4617/2021/
  3. Jaffe, D. A., Miller, C., Thompson, K., Finley, B., Nelson, M., Ouimette, J., Andrews, E. (2023). An evaluation of the U.S. EPA's correction equation for PurpleAir sensor data in smoke, dust, and wintertime urban pollution events. Atmospheric Measurement Techniques 16(5), 1311–1322. DOI: 10.5194/amt-16-1311-2023. https://amt.copernicus.org/articles/16/1311/2023/
  4. Hagan, D. H., Kroll, J. H. (2020). Assessing the accuracy of low-cost optical particle sensors using a physics-based approach. Atmospheric Measurement Techniques 13(11), 6343–6355. DOI: 10.5194/amt-13-6343-2020. https://amt.copernicus.org/articles/13/6343/2020/
  5. Sreekanth, V., Ajay Bhargav, R., Kulkarni, P., Puttaswamy, N., Prabhu, V., Agrawal, P., Upadhya, A. R., Rao, S., Sutaria, R., Mor, S., Dey, S., Khaiwal, R., Balakrishnan, K., Tripathi, S. N., Singh, P. (2022). Inter- versus Intracity Variations in the Performance and Calibration of Low-Cost PM2.5 Sensors: A Multicity Assessment in India. ACS Earth and Space Chemistry 6(12), 3007–3016. DOI: 10.1021/acsearthspacechem.2c00257. https://doi.org/10.1021/acsearthspacechem.2c00257
  6. Jha, S. K., Kumar, M., Arora, V., Tripathi, S. N., Motghare, V. M., Shingare, A. A., Rajput, K. A., Kamble, S. (2021). Domain Adaptation-Based Deep Calibration of Low-Cost PM2.5 Sensors. IEEE Sensors Journal 21(22), 25941–25949. DOI: 10.1109/JSEN.2021.3118454. https://doi.org/10.1109/JSEN.2021.3118454
  7. Dey, S., Arora, V., Tripathi, S. N. (2025). Leveraging Unsupervised Data and Domain Adaptation for Deep Regression in Low-Cost Sensor Calibration. IEEE Transactions on Neural Networks and Learning Systems 36(4), 7277–7286. DOI: 10.1109/TNNLS.2024.3409364. https://doi.org/10.1109/TNNLS.2024.3409364
  8. Jain, S., Zimmerman, N. (2024). Exploration of intra-city and inter-city PM2.5 regional calibration models to improve low-cost sensor performance. Journal of Aerosol Science 177, 106335. DOI: 10.1016/j.jaerosci.2024.106335. https://doi.org/10.1016/j.jaerosci.2024.106335
  9. Ginsburg, E., Adler, G., Boose, Y., Rudich, Y. (2026). Global and regional calibration of low-cost PM2.5 sensors (PurpleAir): assessing the transferability of U.S.-based corrections worldwide. Aerosol Science and Technology 60(10), 1157–1178. DOI: 10.1080/02786826.2026.2676293. https://doi.org/10.1080/02786826.2026.2676293
  10. Pruthi, D., Xu, K., Liu, Y. (2026). Calibration model for purple air monitored PM2.5 across the United States. Environmental Pollution 408, 128935. DOI: 10.1016/j.envpol.2026.128935. https://doi.org/10.1016/j.envpol.2026.128935
  11. Hossain, C. A., Delwar, S., Prapti, D. R., Bari, M. A., Marshall, J. D., Saha, P. K. (2025). Performance Evaluation and Calibration of Low-Cost PurpleAir PM2.5 Sensors in South Asian Conditions: Dhaka, Bangladesh. ACS ES&T Air 2(9), 1904–1915. DOI: 10.1021/acsestair.5c00105. https://doi.org/10.1021/acsestair.5c00105
  12. Zhang, J., Bai, L., Li, N., Wang, Y., Lv, Y., Shi, Y., Yang, C., Xu, C. (2025). Low-Cost Particulate Matter Mass Sensors: Review of the Status, Challenges, and Opportunities for Single-Instrument and Network Calibration. ACS Sensors 10(5), 3207–3221. DOI: 10.1021/acssensors.4c03293. https://doi.org/10.1021/acssensors.4c03293