Evaluating potentially hazardous effects of chemicals on ecosystems has always been an important research topic traditionally studied using laboratory or field experiments. Experiment-based ecotoxicity test results are only available for a limited number of chemicals due to the extensive experimental effort and cost. Given the ever-increasing number of chemicals involved in the modern production process and products, rapidly characterizing chemical ecotoxicity at lower costs has become critical for guiding technology and policy development for chemical risk management. In this study, artificial neural network models are developed to predict chemical ecotoxicity (HC 50 ) based on experimental data to fill data gaps in a widely used database (USEtox). To reduce the manual tuning effort on optimal network architecture, a genetic algorithm is investigated to automatically search and configure the network architecture. The resulting neural network model reached an average test R 2 of 0.632 and had a trivial difference with the global optimal regarding validation MSE. The findings of this study can rapidly predict the ecotoxicity of chemicals and further help to understand the potential risk of chemicals and develop strategies for risk management.

No abstract

Lacking unit process data is a major challenge for developing life cycle inventory (LCI) in life cycle assessment (LCA). Previously, we developed a similarity-based approach to estimate missing unit process data, which works only when less than 5% of the data are missing in a unit process. In this study, we developed a more flexible machine learning model to estimate missing unit process data as a complement to our previous method. In particular, we adopted a decision tree-based supervised learning approach to use an existing unit process dataset (ecoinvent 3.1) to characterize the relationship between the known information (predictors) and the missing one (response). The results show that our model can successfully classify the zero and nonzero flows with a very low misclassification rate (0.79% when 10% of the data are missing). For nonzero flows, the model can accurately estimate their values with an R 2 over 0.7 when less than 20% of data are missing in one unit process. Our method can provide important data to complement primary LCI data for LCA studies and demonstrates the promising applications of machine learning techniques in LCA.

Spatially explicit urban air quality information is important for developing effective air quality control measures. Traditionally, urban air quality is measured by networks of stationary monitors that are not universally available and sparsely sited. Mobile air quality monitoring using equipped vehicles is a promising alternative but has focused on vehicle-level experiments and lacks fleet-level demonstration. Here, we equipped 260 electric vehicles in a ride-hailing fleet in Beijing, China with low-cost sensors to collect real-time, spatial-resolved data on fine particulate matter (PM 2.5 ) concentrations. Using this data, we developed a decision tree model to infer the distribution of PM 2.5 concentrations in Beijing at 1 km by 1 km and 1 h resolution.Our results are able to show both short-and long-term variations of urban PM 2.5 concentrations and identify local air pollution hotspots. Compared with a benchmark model that only uses data from stationary monitoring sits, our model has shown significant improvement with the coefficient of determination increased from 0.56 to 0.80 and root mean square error decreased from 12.6 to 8.1 μg/m 3 . To the best of our knowledge, this study collects the largest mobile sensor data for urban air quality monitoring, which are augmented by state-of-the-art machine learning techniques to derive high-quality urban air pollution mapping. Our results demonstrate the potential and necessity of using fleet vehicles as routine mobile sensors combined with advanced data science methods to provide high-resolution urban air quality monitoring.

No abstract