(60j) Early-Stage Process Design and Sustainability: The Role of Machine Learning in Predicting Environmental Impacts

Early-stage process design is critical in ensuring that products and processes are designed to be environmentally sustainable. However, assessing the environmental impacts of chemicals at this stage can be challenging, as the information available is often limited and uncertain (Papadokonstantakis et al., 2016). This situation arises because the design of the products or processes is often not fully defined, making it difficult to identify all the potential impacts (Argoti et al., 2019). Furthermore, impact assessment requires collecting and analyzing a wide range of data, including information on the properties of the chemicals, their toxicity, and their fate in the environment, among others (Finnveden et al., 2009). The data collection can be extremely complicated, particularly for new or emerging chemicals. Additionally, the impact assessment involves interdisciplinary collaborations between scientists and engineers with expertise in various fields, which further adds to the complexity of the process (Chaplin-Kramer et al., 2017). Machine learning algorithms offer a promising solution to this challenge, as they can help predict the environmental impacts of these novel chemicals with high accuracy even when limited data is available (Karka et al., 2019). Thus, by using machine learning to assess the environmental impacts of chemicals and processes at an early stage of process synthesis, researchers and industries can make informed decisions about the design of sustainable products and processes, leading to a more sustainable design (Karka et al., 2019).

Methodology

This present study focuses on using machine learning to predict the endpoint environmental impacts of chemicals, specifically regarding human health impacts, ecosystem quality, climate change, and resource utilization and presents a novel approach to cradle-to-cradle life cycle assessment where the solvent is recovered for use at the disposal-phase. The eXtreme Gradient Boosting (XGBoost) algorithm was used due to its ability to handle large datasets with many features and handle non-linear relationships between the features and the target variables. A total of 350 data points were used, with 70% being used for training, 15% for validation, and the remainder for testing. SimaPro was used to acquire the data for the target variables. The feature set comprised 200 molecular descriptors and 23 thermodynamic properties for each chemical. A stepwise feature selection criterion was used to select the top 5 features from each feature set, reducing the number of features from 223 to 10. After tuning the six most important hyperparameters for the model, the performance was evaluated on the test set using the R-squared and Root-Mean-Squared- Error (RMSE). To demonstrate the application of the model, we investigate a cradle-to-cradle LCA of Isopropanol (IPA). We use our developed model to predict the LCA for the production of IPA from cradle-to-gate. We then model the use phase by considering a distillation problem where we try to separate a binary mixture of IPA and water. At the use phase (gate-to-gate), we consider emissions due to energy demands and fugitive emissions from the distillation column. At the end-of-life phase, we recover the solvent for reuse by implementing the solvent recovery framework proposed by (Aboagye et al., 2022) giving a gate-to-cradle assessment. The total LCA for the entire process is given by Equation 1. Figure 1 shows the implementation of this approach.

Results

Figure 2 shows the results from the ML model for the production-phase. The model is able to generalize both the human health and global warming potential metrics, giving an R² value greater than 0.7, while that on resource utilization has the least predictive performance with an R² value of 0.318 and a RMSE of 30. We expect to increase the size of this data set so that an accurate prediction can be made for ecosystem and resource utilization impact metrics.

References

Aboagye, E.A., Chea, J.D., Lehr, A.L., Stengel, J.P., Heider, K.L., Savelski, M.J., Slater, C.S., Yenkie, K.M., 2022. Systematic Design of Solvent Recovery Pathways: Integrating Economics and Environmental Metrics. ACS Sustainable Chem. Eng. 10, 10879–10887. https://doi.org/10.1021/acssuschemeng.2c02497

Argoti, A., Orjuela, A., Narváez, P.C., 2019. Challenges and opportunities in assessing sustainability during chemical process design. Current Opinion in Chemical Engineering 26, 96–103. https://doi.org/10.1016/j.coche.2019.09.003

Chaplin-Kramer, R., Sim, S., Hamel, P., Bryant, B., Noe, R., Mueller, C., Rigarlsford, G., Kulak, M., Kowal, V., Sharp, R., Clavreul, J., Price, E., Polasky, S., Ruckelshaus, M., Daily, G., 2017. Life cycle assessment needs predictive spatial modelling for biodiversity and ecosystem services. Nat Commun 8, 15065. https://doi.org/10.1038/ncomms15065

Finnveden, G., Hauschild, M.Z., Ekvall, T., Guinée, J., Heijungs, R., Hellweg, S., Koehler, A., Pennington, D., Suh, S., 2009. Recent developments in Life Cycle Assessment. Journal of Environmental Management 91, 1–21. https://doi.org/10.1016/j.jenvman.2009.06.018

Karka, P., Papadokonstantakis, S., Kokossis, A., 2019. Predictive LCA - a systems approach to integrate LCA decisions ahead of design, in: Computer Aided Chemical Engineering. Elsevier, pp. 97–102. https://doi.org/10.1016/B978-0-12-818634-3.50017-5

Papadokonstantakis, S., Karka, P., Kikuchi, Y., Kokossis, A., 2016. Challenges for Model-Based Life Cycle Inventories and Impact Assessment in Early to Basic Process Design Stages, in: Sustainability in the Design, Synthesis and Analysis of Chemical Engineering Processes. Elsevier, pp. 295–326. https://doi.org/10.1016/B978-0-12-802032-6.00013-X