AecroFormer: Fast, Noise-Robust Aerosol Microphysical Retrieval for Multiwavelength Raman Lidar

Aerosol microphysical parameters (e.g., size distributions and complex refractive index) control scattering and absorption and underpin quantitative estimates of aerosol radiative effects and aerosol–cloud interactions. Retrieving them from multiwavelength Raman lidar is inherently ill-posed: measurement noise and systematic uncertainties quickly erode multi-channel constraints under weak signals, and conventional LUT/iterative inversions are too slow (seconds to minutes per profile) for network-scale or high-throughput processing.

We propose AecroFormer, an end-to-end regression model that incorporates multi-head attention to learn cross-wavelength coupling and deliver physically coherent, range-resolved vertical-profile retrievals with improved stability under real-world SNR and noise. For channel combinations such as 3β+2α, AecroFormer achieves an inference speed of 7.4×10⁻⁵ s per range gate on an NVIDIA GeForce RTX 5080, delivering orders-of-magnitude acceleration relative to LUT/iterative schemes that typically operate from minute-level down to sub-second per range gate (e.g., Müller et al., 1999; Wang et al., 2022). Noise robustness tests show that the model maintains practical accuracy as noise increases: even at 20% noise, it remains stable with MAE(mᵣ) ≈ 0.0758 and MRE(rₑ) ≈ 32.9%.

Focusing on the two important application-critical profile products—effective radius (rₑ) and aerosol volume concentration—we assessed real-world applicability through  an observation-based consistency check using operational measurements from the Aksu site (Xinjiang, China) in January 2024, selecting four days for validation. Retrieved aerosol volume concentrations were converted to 0–2 km boundary-layer mean PM₂.₅ using an empirical density assumption and matched against surface air-quality observations (n = 28). The comparison yields a PM₂.₅ bias of 4.69 ± 26.87 µg/m³ and a relative bias of 3.29%, indicating that the method reproduces both the magnitude and variability observed by ground monitoring in a network-operational setting.

Overall, AecroFormer substantially reduces the computational cost while preserving noise-robust retrieval performance, enabling a practical transition from offline, slow microphysical inversions to near-real-time, high-throughput, and deployable processing. It also provides a reusable algorithmic foundation for future extensions under more realistic bimodal forward assumptions and tightly controlled uncertainty constraints.