Creating rapid, transparent, and updateable material flow analyses
The goal of this project is to improve the speed at which material flow analysis (MFA) is performed and to transparently communicate and reduce the uncertainty in the results, thus enabling a routine and updatable process to make MFAs at any scale (from factories to supply chains). The objective is to quantify the uncertainty of MFA parameters and network structures, and to conceive intelligent data collection strategies early in the MFA, thus decreasing the cost for high-confidence MFA results. MFAs are critical tools in the transition towards a circular economy. They reveal opportunities for material efficiency and symbiosis, as well as the system-level impacts of localized changes to the supply chain (e.g., the effect of increased electric vehicle deployment on lithium extraction and manufacturing emissions). Detailed MFA is currently a costly, time-intensive data collection process followed by a typically manual reconciliation of conflicting and missing data. Attendant uncertainties are typically invisible to the end-user. This research will develop an approach where computational methods are used to reduce data collection costs, improve output data, and better manage uncertainty.
First, Bayesian inference will be used to update uncertainty as new data is collected. This approach will be integrated with the principles of optimal experimental design (OED) to identify the next data records to collect that can lead to the largest uncertainty reduction, measured as the Shannon information gain. Second, the project will study the network structure uncertainty in MFAs by proposing a set of candidate network structures followed by Bayesian model selection to identify the most suitable structure. OED will also be adopted to plan data acquisition that reveals the best network structure. Third, the developed techniques will used to generate MFAs for historical years, and use error propagation to compute in-use stock levels and recycling rates in creating a time-dependent dynamic MFA. These contributions will be accessible through open-source code and demonstrated by studying U.S. steel flows and global polymer flows. The work will result in a guide for eliciting probability distributions for MFA variables from experts, moving the field away from the arbitrary allocation of probability distributions. The costs of improving MFA confidence will be reduced by performing targeted data collection using the principles of OED. Furthermore, network structure uncertainty (i.e., the existence or absence of nodes or flows between two nodes in an MFA) is pervasive and can severely undermine the reliability of the reconciled flow predictions, but has yet to be rigorously studied. This research will investigate network structure uncertainty by forming an ensemble of network structure candidates using both expert advice and randomized network structure generation. OED will also be developed to plan data acquisition that reveals the best MFA network structure while minimizing data collection costs. Through this work, new computational algorithms will be created, for example, an efficient categorical stochastic optimization procedure that incorporates domain knowledge about the MFA network structures, and an information-theoretic OED criterion for model-selection entailing a triple-nested Monte Carlo estimator. This research aims to enable companies, universities, and governments to use an inexpensive but statistically rigorous analysis of physical flows to inform policy, practice, and investment decisions that will increase resource efficiency. Reduced MFA costs and uncertainties will open the door to more powerful forecasting. For example, detailed annual MFAs across time can reveal the dynamics between industries (e.g., steel, aluminum, and cement nexuses) and policies (e.g., tariffs) that are needed for more reliable integrated assessment models. To promote impact, methods and findings will be presented at the Conference of the International Society of Industrial Ecology, produced codes and datasets will be deposited in the Industrial Ecology GitHub repository, and learning tools (e.g., on static and dynamic MFAs) posted to the online learning C-SED website. The project will build a sustained, cross-unit and multi-university Midwest collaboration infrastructure focusing on the topic of data science for sustainability decisions. An outreach program at Ypsilanti STEMM Middle College will integrate key lessons on sustainable materials with FIRST Robotics design principles, helping to empower URM students with the skills to pursue sustainable engineering careers.