This thesis develops an approximate, analytically based environmental assessment method that provides fast evaluations of product concepts. Traditional life-cycle assessment (LCA) studies and their streamlined analytical versions are costly, time-consuming, and data intensive. Thus, they are not practical to apply during early concept design phases where little information is available and ideas change quickly. Alternatives currently used are mostly qualitative, ad-hoc approaches that often provide overly simplistic assessments difficult to trade-off with other design objectives. The Learning Surrogate LCA method is an alternative approach that uses simple, high-level, and accessible descriptive information about a product to provide approximate, yet useful, analytical LCA results during early concept design stages. The method relies on a general artificial neural network (ANN) trained on high-level product descriptors and environmental performance data from pre-existing detailed life-cycle assessment studies or related data. To quickly obtain an approximate environmental impact assessment for a product concept, the design team queries the trained artificial model with new set of descriptors, without requiring the development of a new model. The predicted environmental performance, along with other key performance measures, can be used in tradeoff analysis and concept selection. Foundations for the approach were established by investigating: (1) model inputs in the form of a compact, and meaningful set of product concept descriptors; (2) ability to gather data and appropriately train an ANN-based surrogate LCA model. Proof-of-concept tests on life-cycle energy consumption showed that ANN-based surrogate models were able to: (a) match detailed LCA results within the accuracy of typical LCA studies; (b) predict relative differences of distinct product concepts; (c) correctly predict and generalize trends associated with changes for a given product concept. A product classification system based upon concept descriptors was developed to improve performance. The method was then applied to a case study with a heavy truck manufacturing company. A demonstration example was used to illustrate application scenarios for tradeoff analysis within DOME (Distributed Object-based Modeling Environment). The study suggested that high-level, customizable simulation interfaces of learning surrogate LCA models are likely to have a significant practical impact in the early decision making process.
Prof. David Wallace