Measuring the Environmental Impact of AI Models

Overview

This project explores how different AI model design choices affect environmental sustainability. Instead of evaluating models only by accuracy, the project also measures energy consumption and CO₂ emissions during model training.

The goal is to demonstrate Responsible AI by Design by making environmental impact a measurable and comparable factor in AI development.

---

Problem Statement

AI models can be computationally expensive and energy-intensive. More complex models often achieve slightly better performance, but at the cost of significantly higher energy consumption and CO₂ emissions.

This project answers the question: Is higher accuracy always worth the environmental cost?

---

Use Case

A synthetic maintenance failure prediction dataset is used to simulate a realistic industrial AI scenario.

• Task: Binary classification (failure / no failure)
• Context: Predictive maintenance or operational analytics

---

Methodology

1. Data Preparation

• Generated a synthetic dataset with operational features
• Split data into training and testing sets
• Applied feature scaling

2. Model Training

Two models were trained on the same dataset:

• Simple model: Logistic Regression
• Complex model: Random Forest

3. Environmental Impact Measurement

• CO₂ emissions during model training were measured using CodeCarbon
• Each model training run was logged for reproducibility

4. Comparison

Models were compared using:

• Accuracy
• CO₂ emissions (kg)
• CO₂ emissions per unit of accuracy

---

Results

The comparison showed that:

• The complex model achieved slightly higher accuracy
• The simple model produced significantly lower CO₂ emissions
• Small accuracy improvements can result in disproportionately higher environmental cost

This highlights the importance of sustainability-aware model selection.

---

Responsible AI Perspective

This project follows Responsible AI principles by:

• Measuring environmental impact alongside performance
• Making sustainability trade-offs explicit
• Supporting transparent and reproducible experimentation
• Encouraging environmentally responsible design decisions

---

Tools & Technologies

• Python
• pandas, numpy
• scikit-learn
• CodeCarbon
• VS Code & Jupyter Notebooks

---

Project Structure

ai-environmental-impact/
├── data/
├── notebooks/
├── models/
├── metrics/
├── guidelines/
└── README.md

---

Key Takeaways

• Environmental impact should be considered early in AI development
• Simpler models can offer more sustainable solutions with acceptable performance
• Responsible AI requires measurable sustainability metrics

---

Author

Mainuddin Monsur Robin