Combining Learned and Analytical Models for Predicting CO2e Emissions in Textile Products

Sammanfattning: In recent years, there has been an increased attention to climate change and the effects it might have on human society. One of the main causes of climate change is increased carbon emissions due to different human activities and global scale industrialization. Various approaches exist to mitigate the effects of climate change. One of the largest sources of carbon emissions is the pro- duction of consumer goods. Therefore, by influencing the choices consumers make, we can help mitigate some of the effects of climate change due to carbon emissions.In this thesis, we focus on carbon emissions during the production phase of textile products. To be able to influence consumers in their choice of textile products, we must be able to provide them with detailed information on the emissions of individual products. Such information will allow consumers to compare the products and potentially influence them to choose environmen- tally friendly products. Such a product will hopefully push the market towards sustainable production.Our work focuses on providing methods that can be used to create a ma- chine learning model to predict the carbon emissions of textile products. We collaborate with IVL Swedish Environmental Research Institute (IVL) to build an analytical model to calculate the emissions of textile products. Further- more, we leverage the analytical model to design and compare three machine learning models. We focus on building models that benefit from the knowl- edge of the analytical model while being scalable with regards to new data. Moreover, we introduce a method to use knowledge in the form of analyti- cal models to bootstrap a machine learning model when labelled data is not readily available. In this way, the machine learning model can benefit from the existing knowledge of the analytical model while being adaptable to new labelled data.Finally, we compare the three proposed models and discuss the advantages and disadvantages of each model. We also mention the situations where each model performs best.