01
Earlier intervention
Evaluate environmental consequences before production decisions become fixed.
Master’s Dissertation
Machine Learning
2024
Designing more
sustainable fashion
materials with AI.
A machine learning framework designed to predict the environmental impact of fashion materials before production begins.
For my MSc dissertation at London College of Fashion, I developed a Random Forest regression model using Kering’s Environmental Profit & Loss data.
The research explored how predictive analytics could support more informed and sustainable material-selection decisions during fashion product development.
01
Overview
Can AI improve
material decisions?
Material selection shapes a garment’s environmental footprint long before it reaches a customer. Fashion brands must consider environmental impact alongside performance, cost, aesthetics, availability, and production requirements.
This research investigated whether machine learning could transform complex environmental data into practical decision support during product development.
Research Question
How can an AI-driven predictive model support more sustainable material selection in fashion product development?
02
Research Approach
From environmental
data to decisions.
I conducted a quantitative study using two datasets from Kering’s 2023 Environmental Profit & Loss reporting: Raw Material Intensity and EP&L Valued Result and EKPI.
The model evaluated six environmental key performance indicators to predict the total environmental impact associated with materials throughout Kering’s supply chains.
01
Frame
Defined the problem through Circular Economy and Life Cycle Assessment principles.
02
Prepare
Cleaned, transformed, and structured environmental data for predictive modeling.
03
Model
Built a Random Forest regression model to predict environmental impact.
04
Evaluate
Assessed performance through mean absolute error, cross-validation, and feature importance.
05
Translate
Converted model findings into recommendations for sustainable product development.
03
Environmental Data
Six measures of
environmental impact.
Kering’s EP&L methodology translates environmental harm into measurable ecological and societal costs. The predictive model evaluated six environmental dimensions.
01
Greenhouse Gas Emissions
Climate impact
02
Water Consumption
Resource scarcity
03
Land Use
Ecosystem impact
04
Air Emissions
Pollution and health impact
05
Water Pollution
Water-quality impact
06
Waste
Hazardous and non-hazardous waste
04
Model Development
Improving the
predictive model.
Random Forest regression was selected to capture nonlinear relationships between environmental indicators while also providing interpretable feature-importance results.
The model was refined through outlier removal, logarithmic transformation, hyperparameter tuning, and cross-validation.
01
→
Raw EP&L Data
02
→
Data Preparation
03
→
Random Forest
04
Model Evaluation
05
Results
58.07%
Feature importance attributed to greenhouse gas emissions.
0.723
Cross-validated mean absolute error after model optimization.
6
Environmental KPIs evaluated across Kering’s supply chains.
Greenhouse gas emissions emerged as the strongest predictor of overall environmental impact, indicating that emissions should remain a central consideration within sustainable material-selection strategies.
Cross-validation suggested that the model could generalize beyond its training data, while also identifying opportunities for improvement through larger and more granular datasets.
06
Business Impact
Sustainability as
decision support.
The project reframed sustainability data as an input to product development rather than simply a measurement reported after production.
A predictive framework like this could help material developers compare environmental tradeoffs earlier, identify the variables driving impact, and prioritize areas where intervention could create the greatest improvement.
02
Clearer priorities
Identify which environmental variables contribute most strongly to overall impact.
03
Scalable analysis
Apply predictive analytics across expanding material and supply-chain datasets.
07
Limitations
What the model
cannot answer yet.
The research relied on secondary, aggregated environmental data. The dataset did not capture every commercial consideration involved in material selection, including cost, durability, aesthetics, availability, and supplier constraints.
Future research could incorporate larger datasets, material-level comparisons, real-time supply-chain information, and additional business variables to create a more comprehensive decision-support tool.
Reflection
This project strengthened my ability to move between technical modeling, sustainability research, and business strategy—and communicate the result as a practical decision-making framework.
Explore the project