(551d) Short-Cut Models Based On Molecular Structure for Life Cycle Impact Assessment of Biorefinery Products
Sustainable process design has been recognized as one of the key research challenges for process systems engineering. In fact, the experience gained so far has pointed out that the applicability of sustainability principles can be more advantageous in earlier phases of process design characterized by more degrees of freedom for decision making. Typical decisions made in these early process design phases involve the selection of chemical synthesis path, chemical auxiliaries, unit operation conditions up to a basic flowsheet design.
It is well established that life cycle analysis (LCA) can effectively assist policy makers to have an overall system-based perception of environmental impacts beyond the traditional plant emissions oriented one. LCA achieves these goals following a systematic procedure, well-defined in ISO-Norm 14040. One of the most prominent challenges to perform comprehensive LCA studies has to do with the material and energy flow analysis for cradle-to-gate life cycle inventories. Currently, there are a few databases (e.g., Ecoinvent) and tools (e.g., SimaPro) for obtaining “cradle-to-gate” data of a substantial number of chemicals based on data standardization techniques for evaluating multiple data sources. However, these databases are far from being complete. The direct impact of these data gaps for any process design project is that assumptions have to be made on the basis of process similarities, which requires expertise on the specific processes, results in time consuming modelling, is often case study specific, and, therefore, almost impossible to automate. Naturally, this phenomenon of data gaps becomes much more intense for new technologies and production concepts such as the biomass based production of chemicals. Therefore, especially for biorefienries, there is a need for developing short-cut models that estimate life cycle inventories or even life cycle impact assessment metrics (LCIAs) for a fast screening of different options in early phases of process design.
To this end, a novel approach for basic and fine chemicals based on fossil resources has already demonstrated that complementary life cycle impact assessment data can be obtained on the basis of molecular structures. The respective “black-box” models are based on similarity measures for an optimal set of molecular descriptors utilizing simple to advanced (e.g., from multiple linear regression to PCA and neural networks) correlations between these molecular descriptors and LCIAs such as cumulative energy demand (CED), global warming potential (GWP), and eco-indicator99 points (Wernet et al., 2009). This type of models usually target at modeling error around 30-40%, i.e., approximately at the same error level of cost estimations in early process design phases.
In this study we have developed short-cut models for estimating the CED and GWP of biorefinery chemical products based on their molecular structure and minimal process information. The short-cut models were developed using multi-linear regression (MLR) coupled with an optimization procedure for selecting the optimal set of predictors for the MLR model. In total, more than 20 predictors were considered and the best MLR model was derived based on a leave-one-out cross validation procedure and performance metrics penalizing for higher model complexity (e.g., Akaike metric, adjusted R2, etc.).
The data for training and validation of these models were derived from process simulations based on literature, experimental and pilot plant data. They include 25 chemical products produced from diverse biomass resources, process layouts and operating conditions. The resulting gate-to-gate material and energy flows were coupled with existing databases (e.g., Ecoinvent) for estimating cradle-to-gate life cycle inventories of the chemical auxiliaries and energy utilities used in the processes. The waste streams were also characterized using short-cut LCA models for waste treatment (i.e., mainly wastewater treatment and waste solvent incineration). The functional unit for the LCA values is 1 kg of chemical product, considering allocation and/or substitution for targeted multi-production or byproduct utilization, which is a typical case in biorefineries.
The resulting models show a satisfactory accuracy in validation terms (e.g., R2 values higher than 0.6), which is far beyond the statistical significance level for the number of available data and fitted parameters. Moreover, the sensitivity of model performance when selecting between diverse molecular structure and process layout descriptors is analyzed, highlighting the most problematic cases. Finally, a comparison to the equivalent LCA values for the production of these chemicals from fossil fuels is presented and future challenges are discussed.