Green AI

Project: Energy-Efficient Inference of Large Language Models in Software Engineering

Third parties involved: Hasan Sozer – Ozyegin University
The project empirically analyses the energy consumption of large language models in continuous integration pipelines for software engineering tasks, with a focus on inference in code generation and analysis. It examines how hardware platforms, model configurations, benchmark datasets, and LLM-based agents influence energy use and computational cost. By producing systematic evidence on inference-related energy demands, the project aims to support more efficient, scalable, and sustainable use of LLMs in software engineering.

Project: Developing a framework for green AI practices

Third parties involved: Maria Ulan – RISE Research Institutes of Sweden
The project aims to define and standardize green AI metrics across the entire AI lifecycle. It will review existing metrics in sustainable software engineering, propose standardized calculations, and introduce new metrics where current ones are inadequate. These metrics will be evaluated and validated. The project will also outline the data requirements and functionalities for a benchmarking tool to minimize AI’s environmental impact. Expected outcomes include a robust open-source green AI metrics framework and its application to assess AI architectures for energy efficiency and performance. This will provide a foundation for European policymakers to develop regulations and eco-labels that promote environmentally responsible AI development and deployment.

Project: Information-theoretic analysis of generalization in deep learning

Third parties involved: Linara Adilova – Ruhr University Bochum
The project aims to apply information-theoretic analysis—specifically, information planes—to deep neural networks to predict their generalization and potentially optimize their performance and size. It will first select a suitable mutual information estimation approach for deep neural networks, addressing challenges of high dimensionality and computational complexity. Using this approach, the project will then assess the capability of information-plane analysis to explain a network’s generalization through a large, carefully curated set of experiments. Finally, it will propose and evaluate a method for optimizing network architectures for given tasks.

Project: Advancing World Model Learning with Neural Cloned-Structured Causal Graphs

Third parties involved: Tristan Manfred Stöber – Ruhr University Bochum
The project pioneers a neural implementation of Cloned-Structured Causal Graphs (CSCGs), an algorithm for learning world models from sequential observations. By reformulating CSCG as a deep learning problem, it demonstrates seamless integration of a neural CSCG module with a vector-quantised variational autoencoder (vqVAE) and its ability to produce rich latent representations from visual input. The flexibility of this neural approach enables joint training of the vqVAE and neural CSCG to test whether explicit world models can improve prediction of action consequences. It aims to deliver significant advances in autonomous, reliable, and data-efficient artificial intelligence.

Project: Tiny Models on the Resource Constrained Edge Devices

Third parties involved: Kasim Sinan Yildirim – University of Trento
The project explores the development of tiny AI models for self-powered edge devices operating under energy-harvesting conditions. These models will adapt their accuracy according to the available energy budget, enabling AI-based applications in devices powered exclusively by ambient energy. The project addresses the extreme resource constraints, intermittent operation, and frequent power failures caused by variable environmental power sources. By
supporting more efficient and adaptive AI at the edge, it contributes to sustainable and smart IoT applications for the next decade.

Project: Optimal hybrid AI strategies for battery management and state estimation

Third parties involved: Lluís Trilla, Andrés Bernabeu Santisteban – IREC (Catalonia Energy Research Institute)
This project analyzes hybrid AI strategies for battery management and state estimation by combining reduced-order physics-based models with machine learning. A key objective is to balance model granularity with computational efficiency for optimal estimation of SoC and SoH. Numerical studies will be complemented by experimental validation at IREC and KNOW Center. Expected outcomes include low-cost algorithms, efficient state estimation, and modular hybrid models that advance sustainable energy solutions for electric mobility and energy storage.

Project: Physics-Informed Neural Networks for Predictive Energy Management

Third parties involved: Sajjad Fattaheian, Mahyar Tofighi Milani – Aalto University
The project develops a physics-informed neural network framework to improve energy management in grids with high renewable penetration. By embedding power-flow equations and grid constraints into neural networks, this approach aims to enhance forecasting accuracy, robustness, interpretability, and physical plausibility. Trained and validated with data from records, sensors, SCADA systems, and simulations, it covers normal, peak-load, fluctuation, and fault
scenarios. The framework supports real-time decisions and contributes to reliable, scalable, and sustainable smart grid management.

Project: Think Before You Sample: Physics-Informed Active Learning for Green AI

Third parties involved: Bernd Schuscha – Materials Center Leoben
Forschung GmbH
The project develops physics-informed Bayesian Optimisation strategies by integrating partial physical knowledge to improve robustness, convergence, and
efficiency under real-world conditions, including noise and model misspecification. Combining physics-informed machine learning with active learning, it guides data sampling to reduce training data needs and computational complexity. The approach also evaluates trade-offs between convergence speed and predictive accuracy, adapting strategies to expected
noise levels.

Project: How Green is Physics-Informed Machine Learning for Solar Power Forecasting?

Third parties involved: Peter Pavlik – Kempelen Institute of Intelligent Technologies
Physics-informed machine learning (PIML) has emerged as a promising paradigm that integrates physical laws with data-driven models to improve accuracy and generalisation while promising reducing data and computation needs. This project examines the claim of environmental and computational efficiency of PIML in the context of solar power forecasting, a domain essential for maintaining grid stability amid the growing share of renewables. Using a real-world solar irradiance and power output dataset, we will develop and compare data-driven, physics-based, and hybrid forecasting models in terms of predictive accuracy, training energy consumption, and carbon footprint to guide responsible AI deployment in renewable energy forecasting.

Project: Towards Green AI adoption and policy for sustainable AI governance

Third parties involved: Anastasija Nikiforova – University of Tartu
The project examines Green AI adoption and develops a responsible framework to support sustainable AI governance and policy development. It combines
literature review, EU policy analysis, and interviews with public-sector representatives, policy experts, and practitioners in Estonia and Europe. Drawing on ESG and technology adoption theories, it identifies enablers, barriers, trade-offs, and sector-specific practices. The main outcome will be an AI lifecycle-based framework, validated with experts and translated into policy and regulatory recommendations.

Project: Hybrid AI Model for Mixed-Gas Sorption Upper Bounds

Third parties involved: Eleonora Ricci – University of Edinburgh
The project aims to build a hybrid modelling framework for mixed-gas sorption in polymer membranes for environmental applications, including carbon capture and hydrogen separation. It integrates physics-based and data-driven approaches. The project also evaluates multiple representation methods—such as molecular graphs and chemical language models—to develop machine-learning models that predict the parameters of a physics-based SAFT equation of state. The resulting integrated ML/SAFT framework will be released as open-source software and used to generate new mixed-gas performance plots. These plots will guide membrane scientists in materials selection, advancing membrane technology in Europe and beyond.