Рақамли технологияларнинг назарий ва амалий масалалари 8-том 3-сан (2025) · 130-141-беттер
Integration of Machine Learning Methods into Lagrangian Particle Models of Atmospheric Transport
Туркменова, Р.Т.
Аннотация
This review systematizes how machine learning (ML) can be integrated into Lagrangian particle dispersion models (LPDM: FLEXPART, HYSPLIT, STILT, NAME) as a physics-consistent augmentation rather than a replacement of the core dynamics. We highlight four strands: (1) physics-informed surrogates for parameterizations (notably PBL/vertical diffusion and wet removal), (2) variance-reduction and Monte-Carlo acceleration (importance sampling, control variates), (3) bias correction and probabilistic calibration of ensembles (EMOS/quantile mapping) using proper scoring rules (CRPS, reliability), and (4) mass-conserving super-resolution of concentration fields without re-integrating trajectories. Inversions based on footprint matrices (Bayesian, variational, and ensemble approaches) and observation-network requirements are discussed separately. The methodology includes a formalized literature search (2000–2025; WoS/Scopus/Scholar) and reproducibility practices (fixed meteorological drivers, reporting protocols). We argue that robust ML gains depend on two disciplines: preserving physical invariants (mass balance, well-mixed, non-negativity) and methodological rigor (RMSE/correlation/CSI-POFD/CRPS, reliability diagnostics). A “minimal viable” operational pipeline is proposed: fix ERA5/WRF or GDAS/GFS forcings and LPDM configs; add mass-conserving bias correctors and probabilistic calibration; apply a sum-preserving SR module; optionally use PI surrogates for nocturnal PBL regimes, with mandatory validation on real episodes.
лагранжево-частичные моделиFLEXPARTHYSPLITSTILTNAMEмашинное обучениесупер-разрешениевероятностная калибровкаCRPSинверсии источников
Метадайындар булагы: журналдын OAI-PMH архиви · Sindex толук текстти сактабайт, булакка шилтеме берет.