AI-CARE

Carbon-Aware Reporting Evaluation Tool for AI Models

This repository provides the official implementation of AI-CARE, a reporting-centric evaluation tool for measuring and reporting predictive performance, energy consumption, and carbon emissions of machine learning models.

The repository accompanies the paper:

📄 AI-CARE: Carbon-Aware Reporting Evaluation Metric for AI Models
🔗 https://arxiv.org/abs/2602.16042
📥 https://arxiv.org/pdf/2602.16042

AI-CARE elevates energy and carbon emissions to first-class evaluation quantities and enables transparent, reproducible comparison of models under fixed experimental conditions.

---

Overview

AI-CARE evaluates machine learning models by executing training and inference under identical conditions, measuring empirical energy consumption, converting energy usage to carbon emissions using a fixed grid carbon intensity, and reporting accuracy–carbon tradeoffs through standardized metrics and visualizations.

The framework does not modify model architectures, optimization algorithms, or training procedures. It reports empirically observed performance, energy, and carbon metrics during both training and inference.

---

Models

The following representative model families are evaluated:

• Multilayer Perceptron (MLP)
• Convolutional Neural Network (CNN)
• Transformer-based classifier
• MLP-Mixer
• MobileNetV2
• ResNet-18 (ImageNet-100 only)

Model hyperparameters are held fixed across datasets to isolate architectural and dataset effects.

---

Datasets

Experiments are conducted on five vision benchmarks of increasing complexity:

• MNIST
• Fashion-MNIST
• CIFAR-10
• CIFAR-100
• ImageNet-100

Dataset properties (input resolution, number of channels, number of classes) are inferred automatically by the evaluation pipeline.

---

Experimental Setup

• Optimizer: Adam
• Learning rate: 1e-3
• Batch size: 64
• Training epochs: 10
• Hardware: CPU-only
• Grid carbon intensity: 400 gCO₂/kWh

A fixed number of epochs is used to emphasize relative accuracy–energy–carbon tradeoffs rather than absolute peak accuracy.

---

Reported Metrics

AI-CARE reports the following quantities:

• Task performance (accuracy)
• Training energy consumption and carbon emissions
• Inference energy consumption and carbon emissions
• Total carbon emissions (training + inference)
• Carbon–Accuracy Tradeoff Curves (CATC)
• Scalar Carbon-Aware Score (SCAS)

The scalar score integrates normalized task performance and total carbon emissions to support comparative ranking when a single decision criterion is required.

---

How to Run

pip install -r requirements.txt
python run_benchmark_experiments.py
python analyze_results.py

---

Outputs

The evaluation pipeline produces CSV files containing reported performance, energy, and carbon metrics, along with publication-quality visualizations of carbon–accuracy tradeoffs and per-dataset scalar carbon-aware scores.

---

Reproducibility

All experiments are executed under fixed hardware and software conditions with deterministic configurations and a fixed grid carbon intensity to ensure reproducible and comparable results across models and datasets.

---

License

This project is licensed under the Apache License 2.0 – see the LICENSE file for details.

Patent Notice

This project may be subject to patent protection. Certain components described in the associated publication are pending patent review.

---

---

Citation

If you use this work, please cite:

@misc{santosh2026aicarecarbonawarereportingevaluation,
  title={AI-CARE: Carbon-Aware Reporting Evaluation Metric for AI Models}, 
  author={KC Santosh and Srikanth Baride and Rodrigue Rizk},
  year={2026},
  eprint={2602.16042},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2602.16042}, 
}