Implementing Artificial Intelligence Techniques to predict Environmental Impacts: Case of Construction Products

Session
WE.1.A || Life Cycle Sustainability in Construction and Renovation of Buildings

Authors
Koyamparambath, Anish; Adibi, Naeem; Szablewski, Carolina; Adibi, Sierra A; Sonnemann, Guido

Abstract
Nowadays, decision-makers, including product designers, manufacturers and consumers, are interested in knowing the environmental impacts of their products, processes and services. Life Cycle Assessment (LCA) is a tool that assesses the environmental impacts of a product system over its life cycle. An Environmental Product Declaration (EPD) is one of the most common ways of communicating the results of an LCA. Conducting an LCA requires meticulous data sourcing and collection and is often time-consuming for both practitioner and verifier. This work proposes a new method of using Artificial Intelligence (AI) techniques to predict the environmental performance of a product to assist LCA practitioners and verifiers. This approach uses data from EPDs of construction products to train the machine-learning algorithm called Random Forest that constructs a multitude of decision trees randomly for the prediction. We trained the model with information on the product system and their environmental impacts using seven impact categories’ values. We verified the results using a testing dataset (20% of the data from EPD). Our results demonstrate that the model was able to predict the values of global warming potential, abiotic depletion potential for fossil and non-fossil resources, acidification potential and photochemical ozone creation potential with an accuracy (measured using R2 metrics, a measure to score the correlation of predicted values to real value) of 80%, 62%, 77%, 68% and 70%, respectively. Tuning the model’s hyperparameters using grid search and k-fold cross-validation method and transforming the values resulted in improved performance of the above indicators. However, ozone depletion potential (ODP) and eutrophication potential are exceptions with low accuracy, as the values are in the order of -10 to -16 that can be improved with a larger dataset.