(226f) Large-Scale Prediction of Life Cycle Inventories of Chemicals

The transition toward a sustainable chemical industry needs to assess the environmental impacts of chemical production accurately. A comprehensive approach to evaluate environmental impacts across multiple environmental categories is Life Cycle Assessment (LCA). However, LCA relies on detailed life cycle inventory (LCI) data (i.e., mass and energy flows). This LCI data is typically unavailable in early design phases. Consequently, current LCA databases cover only a limited number of chemicals, leaving substantial data gaps that must be filled through time-consuming manual data collection and analysis.

To overcome this lack of data, two types of methods have emerged for predicting the environmental impacts of chemical production: (1) machine learning-based models using molecular descriptors, which lack accuracy due to data scarcity for training the models and omission of reaction pathways [1,2]; and (2) stoichiometry-based approaches distinguish reaction pathways but rely on manual data input and are thus labor-intensive [3].

In this work, we propose a pathway-resolved framework to automatically predict the LCI of organic chemical production, enabling large-scale transparent environmental impact assessments. Specifically, from a target chemical’s molecular structure represented by its SMILES code, a machine-learning-based retrosynthesis tool predicts the reaction pathways and precursors to produce the target chemical. This retrosynthesis tool is further integrated with an optimization model to balance the reaction equations from which the feedstock demand is estimated. From the reaction equation, we predict the remaining gate-to-gate LCI data, e.g., energy and reagent demands, by recently developed decision trees. Ultimately, the framework links the target chemical to existing LCA databases via LCI data through its precursors. For precursors not covered by these LCA databases, their LCIA scores are predicted using a multi-task feedforward neural network. Due to the transparent modeling of pathways and inventories, the framework can be adapted for specific LCIA methods as well as system models and allows for easy refinement once better data becomes available.

We validated the predictive capabilities of our proposed framework by comparing the predicted results with a benchmark dataset of 136 organic chemicals, including industrially validated LCIs. The results demonstrate improved accuracy and broader scope in predicting environmental impacts across all categories compared to the existing literature. The fully automated framework helps bridge data gaps in LCA databases for early-stage process design, supporting a faster transition to a more sustainable chemical industry.

References

[1] J. Kleinekorte et al., 2023, APPROPRIATE Life Cycle Assessment: APROcess-Specific, PRedictive Impact AssessmenT Method for Emerging Chemical Processes. ACS Sustainable Chemistry & Engineering, 11(25), 9303-9319.

[2] D. Zhang et al., 2024, Enhanced deep-learning model for carbon footprints of chemicals. ACS Sustainable Chemistry & Engineering, 12(7), 2700-2708.

[3] T. Langhorst et al., 2023, Stoichiometry-Based Estimation of Climate Impacts of Emerging Chemical Processes: Method Benchmarking and Recommendations. ACS Sustainable Chemistry & Engineering, 11(17), 6600-6609.