Emerging investigator series: a machine learning approach to quantify the impact of meteorology on tropospheric ozone in the inland southern California

Abstract

The role of meteorology in facilitating the formation and accumulation of ground-level ozone is of great theoretical and practical interest, especially due to emissions shifts and changing global climate.

Performance of machine learning for ozone modeling in Southern California during the COVID-19 shutdown

“…Those included neural network, support vector machine, k-nearest neighbors, and random forest. 21 Here, we selected random forest regression (RFR), as RFR is the most suitable ML algorithm for predicting ozone concentrations in SoCAB. We also conducted a 10-fold cross-validation over the training data to ne tune the training RFR model in the previous study.…”

“…We also conducted a 10-fold cross-validation over the training data to ne tune the training RFR model in the previous study. 21 The random forest (RF) algorithm is a supervised learning method employing a tree-based ensemble approach. Each decision tree is derived from training data and represents a subset of the training data.…”

Graphics Processing Unit Assisted Computation for a Gas-Phase Chemical Solver in a Regulatory Chemical Transport Model

“…In regard to runtime, for example, a 12 km horizontally spaced, two-way coupled WRF-CMAQ simulation with 34 layers of variable thickness and a domain size of 279 x 251 grid cells requires over three hours of wall clock time for one simulated day when utilizing 32 CPU cores (24). In a preceding study, we simulated five months (1 May -30 Sep 2017) of ozone concentrations for Southern California using 4 km horizonal spacing and a domain size of 156 x 102 grid cells, and the wall clock time was 10 days when utilizing 16 MPI processes (i.e., ~1.6 hours per simulated day with 8 CPU cores) (5).…”

“…We used OBSGRID (47) to improve meteorological analyses, incorporating the observed surface and upper air to correct the NAM data corresponding to the ds461(48) and ds351 (49) datasets, respectively. Other specifications are summarized in a preceding study (5). Figures S1 to S16 Tables S1 to S3 Legends for Software S1 SI References…”

“…Recently, with the rapid growth of machine learning (ML) accessibility, empirical AQMs have become more useful for research purposes by speeding up analyses, and modeled outcomes rely fully on empirical relationships between historical data (5). However, ML models have not been standardized for regulatory decision-making, and conclusions drawn from ML applications are best grounded with the first principles relationships that govern deterministic AQMs.…”