Exposure to diesel-related air pollution, which includes black carbon (BC) as a major component of the particulate matter emitted in engine exhaust, is a known human health hazard. The resulting health burden falls heavily on vulnerable communities located close to major sources including highways, rail yards, and ports. While air quality in the United States has improved, racial disparities in exposure to pollution have persisted. Concentrations of directly emitted air pollutants like BC vary on fine spatial scales, but measurements of pollutant concentrations are limited. Modeling studies have historically used spatial resolutions that are too coarse to resolve elevated concentrations near major sources such as highways. Determination of source contributions to the overall pollution burden is challenging due to collinearity in the exhaust composition profiles for relevant sources including heavy-duty diesel trucks, railroad locomotives, cargo-handling equipment, and marine engines. Additionally, the impact of each source depends not just on the magnitude of emissions, but on its location relative to pollution receptors as well as on meteorology. Understanding exposure to diesel exhaust requires knowledge of how relevant pollutant concentrations vary in space and time, ideally with the relative contribution from different sources identified.

Source-resolved BC concentrations in West Oakland, California, are modeled at high (50 and 150 m) spatial resolutions using the Weather Research and Forecasting (WRF) model as well as a more computationally-efficient neural network-based model. West Oakland is located next to the Port of Oakland and is surrounded by highways. The WRF modeling uses large-eddy simulation in the highest-resolution innermost model domains, such that sub-grid scale turbulence is appropriately parameterized, and larger turbulent eddies are explicitly modeled. BC emissions are mapped and tracked separately in WRF for eight source categories: (1) heavy-duty diesel trucks, (2) light-duty gasoline vehicles, (3) railroad locomotives, (4) cargo-handling equipment, (5) port-related and (6) non-port ocean-going vessels, and (7) port-related and (8) non-port commercial harbor craft.

The ability of the models to predict hourly and 24-h average BC concentrations is evaluated for a 100-day period in summer 2017, when BC was measured at 100 sites within and near the community. A centrally-located monitoring site in West Oakland is found to correspond well with population-weighted average BC concentrations in the community. Major contributing sources to BC in West Oakland include on-road diesel trucks (44±5%) and three off-road diesel sources: ocean-going vessels (19±1%), railroad locomotives (16±2%), and harbor craft such as tugboats and ferries (11±1%).

Even with a relatively high spatial resolution of 150 m, the WRF model cannot capture high BC concentrations and steep concentration gradients near major highways with high volumes of diesel truck traffic. Running atmospheric models such as WRF over long (i.e., seasonal to annual) time scales with high spatial resolution is computationally expensive, due to the small model timestep that is needed to maintain numerical stability. Typically, trade-offs must be made between resolution and the spatial extent of the modeling domain to manage computational costs, resulting in a loss of valuable information. Machine learning-based modeling approaches have not been used extensively in meteorological and air quality modeling. In this research, an alternative neural network-based model for primary air pollutants is developed and evaluated.

Predicted 24-hour average maps of BC concentrations derived from two weeks of 50-m resolution WRF modeling are used to train a neural network to predict high-resolution BC concentrations using a combination of low-resolution (4 km) modeled wind speeds and planetary boundary layer height and high-resolution (50 m) source-resolved BC emission maps. The neural network predicts similar absolute BC concentrations and relative source contributions as the full WRF model simulations. The trained neural network predicts source-apportioned BC concentrations at 50 m resolution for the whole summer (100 days) with an 80% reduction in computational costs.

Activity in the freight sector in the United States is projected to increase at a rapid pace in coming years, exacerbating negative air quality impacts for communities living near associated goods movement infrastructure such as ports, highways, and railyards. Locomotives are a particularly troublesome component of the freight sector, due to delays in the rollout of stricter emissions standards and slow fleet turnover compared to diesel trucks. California has been successful in implementing stringent emissions control technology on heavy-duty trucks, but the railroad industry has contested local and state-level efforts to control emissions. The industry has prevailed in court with arguments that locomotive emissions can only be regulated by the federal government.

The impacts of future deployment of advanced emission control technologies and offsetting effects of growth in diesel engine activity are analyzed by scaling predicted BC concentrations for the baseline year (2017) to reflect future conditions. Total BC concentrations and relative source contributions were estimated for 2025 and 2030. In 2025, air quality in West Oakland is expected to improve relative to baseline conditions, as regulations on heavy-duty truck emissions are fully phased in. By 2030, however, growth in other freight-handling sectors (e.g. locomotives and marine vessels) will overtake advances in emission controls. As a result, BC concentrations in West Oakland are predicted to increase and by 2030 will climb back up to baseline (2017) levels. Assuming that currently-implemented emission control technologies on diesel trucks remain effective, the truck contribution to population-weighted average BC concentrations is forecast to drop from 44 to 8% by 2030. Meanwhile, rapid growth in activity and slow fleet turnover will cause the contribution from locomotives to increase from 16 to 31%.