Research Projects

The project, SustAInd addresses the need for environmentally and economically sustainable AI systems in India. As AI adoption rises across sectors, the energy demands of AI-enabled systems significantly impact sustainability goals. AI operations consume approximately 2% of the world's electricity, and training a single model can emit carbon equivalent to five cars over their lifetimes. Additionally, over 50% of ML systems fail to reach production due to cost and maintainability issues. To this end, this project aims to develop a software-centric framework that enables organisations, policymakers, and developers to design, develop and deploy AI-enabled systems that are sustainable from environmental, economic and technical perspectives. The primary scientific objective of SustAIndian is to create a software-centric framework that enables sustainable AI-based systems through optimized software architectures, adaptive runtime management and effective trade-offs. The hypothesis driving this research is that by embedding energy efficiency considerations within the architecture and runtime of AI-enabled systems, we can significantly reduce their environmental footprint without compromising performance. This framework will incorporate architectural design principles that consider energy a key factor, balancing system performance, cost and environmental impact. The project will proceed through three main experimental phases: 1. Design-Time Framework: This phase will focus on identifying key metrics that serve as sustainability indicators, establishing best practices for AI system architecture, including modular design patterns and lightweight model selection that optimize energy efficiency. A set of software tools will be developed to assess and quantify energy usage during AI system design. 2. Run-Time Self-adaptation Framework: In this phase, we will implement self-adaptation mechanisms to allow AI-enabled systems to adjust parameters based on current energy requirements dynamically. Approaches such as sustainable MLOps, selective activation of AI models, and runtime switching of models will be developed to enhance runtime sustainability. 3. Trade-off Analysis and Energy Audit Tool: To make informed sustainability decisions, we will analyze the trade-offs between energy consumption, system performance, and accuracy in AI models. A star rating system will also be introduced to provide standardized energy efficiency benchmarks for AI systems, along with ESG-compatible reports to guide organizations with sustainability goals. The SustAInd framework will contribute significantly to the field of Green AI, providing an actionable approach to reduce the environmental footprint of AI-enabled systems. The framework will apply to various domains, offering a toolset for developers and architects to build sustainable AI systems. Further, the project's output will contribute to India's sustainable development and green goals of Net Zero Emissions by 2070.