[ECSS 2025] Presentations of the Green ICT and ICT for Green Workshop

European Informatics Leaders Summit ECSS 2025 The European Voice of Informatics Research and Education Green ICT & ICT for Green Workshop Rennes, 29 October, 2025 Sustainable Computing for Fluid Dynamics via Physics-Informed Surrogates Gabriele Gianini w/Luigi Ciceri, Corrado Mio, Jianyi Lin Green ICT & ICT for Green Workshop, 29 October 2025 ECSS 2025 - European Informatics Leaders Summit, Rennes Computational Fluid Dynamics (CFD) •CFD is the field of numerical simulation of flow-related problems •It finds application in a number of diverse domains Large Eddy Simulation Turbo engine Digital Twins Key role in the design of efficient transportation, as well as urban landscapes design The task Numerically predicting behavior of fluids •Motion, forces, transport of heat/mass/momentum By solving the governing physical equations, e.g. Navier-Stokes’ under given •Geometry/domain •Physical params: ρ (density), μ(viscosity)… •Boundary conditions (BCs) •Initial conditions (ICs for transient problems) •Inlet fields (velocity/mass-flow, temperature) •Outlet fields (pressure/zero-gradient) •Walls (no-slip/slip, heat flux/temperature…) •Forcing (volumetric sources): •gravity/buoyancy, porous drag Numerics (implicit but important): discretization (FVM/FEM), mesh/resolution, time stepping, solver tolerances Navier-Stokes Partial Differential Equations •Often the target quantities are •Velocity vector field v(x,y,z), Pressure field p(x,y,z), •And the equations are the Navier-Stokes PDEs, e.g. for v and p Steady incompressible, no forcing case Space discreti zation FDE FVM FEM Depending on the use case the focus can be on the spatial resolution  complexity  1/3 Critical for CFD of porous structures such as tree canopies in urban settings Laminar vs. Turbolent flows These PDEs can become hard to solve in particular parameter ranges, especially with high → more turbolent systems, roughly Reynolds number comput. complexity  Re3-4 https://www.idealsimulations.com/resources/reynolds-number/ Re from 1 to 109 https://doi.org/10.1063/5.0217320 https://doi.org/10.1063/5.0199350 Order-of-magnitude estimated examples: footprint 104kg of CO₂ for single run •dependent on year, hw efficiency, PUE grid Our architecture: from PI-PointNet++ to PI-GANO++ If accuracy degrades at sharp interfaces/ geometry has fine, multi-scale details one one can move from PIPN to PIPN++ •PIPN++ upgrades the encoder to a hierarchical, locality-aware version, adding multi-scale neighborhood features and local conditioning We performed the same operation with PIGANO developing a PIGANO++ architecture, which in analogy to PIPN++ brings •hierarchical, locality-aware geometry into the neural-operator setting Yields better boundary/interface fidelity and generalization to fine geometric details—without giving up neural-operator flexibility across BCs, parameters, and meshes. Our work: - developed PIGANO++ - applied PIPN++ and PIGANO++ to hybrid freeflow and porous medium materials Example with 2D porous shape: PIPN++ Error landscape MAE in windbreak case studies Average Times Method Mesh / Collocation Training Time (x factor CFD) Inference Time (per case) OpenFOAM (Laminar) 86 k cells - 26 min 20.4 min PIPointNets 2500 pts 110 min ( 4.23) 12 ms PIGANO 3000 pts 301 min (14.75 ) 14 ms PIGANO++ 3000 pts 309 min ( 15.15) 17 ms HW: HPC cluster, 32 CPU cores + NVIDIA A100 (2D) ABC for PIPN, (3D) windbreaks for PI-GANO & PI-GANO++ MEDES 2025 - THE 17TH INTERNATIONAL CONFERENCE ON MANAGEMENT OF DIGITAL ECOSYSTEMS (HOCHI MIN CITY, VT) Questions? Sustainable Software Research @UNamur Gilles Perrouin, FNRS Research associate [email protected] What about Namur (Namen, Nameur) ? 2 FR Capital of the Walloon Region 55 Kms South-East of Brussels Population: 110 000 What about UNamur? 3 Students: 7300 Personnel: 1500 Belgium: 1830 UNamur: 1831 Faculties: 7 Research Institutes: 11 Collaborations: >55 countries Alliance: UNIVERSEH Faculty of Computer Science •Born in 1970, first in Belgium! •19 professors •1 FNRS Research Associate (tenured scientist, me J) •18 Teaching assistants (PhD students) •+/-40 researchers (grantees, etc.) •650 students 4 CNNGEN Domainspecific Heuristics ... Example 11 A more Diverse Search Space 12 5.2. On Datasets (a) Distribution of energy of 1 , 300 models generated with CNNGen on CIFAR-10. (b) Distribution of energy of 1 , 300 models generated with CNNGen on CIFAR-100. (c) Energy distribution of 1 , 300 models generated with CNNGen on the Fashion-MNIST. Figure 5.2: Distribution of energy for CIFAR-10, CIFAR-100, Fashion-MNIST while evaluating the performance of the 1,300 models after training for 36 epochs. 5.2 On Datasets We initially used the dataset presented in Section 5.1 to enable direct comparisons with existing benchmarks that will be presented in Section 8.1. Subsequently, we used the same dataset to train the predictors (see Chapter 6). In our previous work [Gratia et al., 2024], these predictors showed promising results, which led us to integrate them into the fitness function of the NSGA-II algorithm (as described in Chapter 7). While the predictors performed well in isolation (i.e., whenever being trained and tested on the 1.3k architecture dataset), their accuracy dropped significantly when used within the genetic algorithm. Specifically, the Mean Absolute Error (MAE) for performance prediction doubled, increasing from around 10% to 20%, and a similar degradation was observed for energy prediction. This led us to suspect that the predictors were either undertrained or overtrained so that they cannot generalise to new architectures generated by NSGA-II, particularly those created through crossover and mutation. To address this issue, we created a second dataset with a larger number of architectures to improve their generalisation and reliability. We initially began with a dataset of 1 , 300 architectures. We arbitrarily increased the size to 10 , 000. Based on observed performance improvements, we plan to continue expanding the dataset in steps of 10 , 000 architectures, stopping once the trade-off between performance gain and computational cost is no longer favourable. 49 §Models: oCNNGen:Generated models range from 9 to 270 layers (1,300 random models). oNASBench:Covers 15 to 85 layers across a larger set. §Accuracy Distribution: oCNNGen’s average accuracies are lower than NASBench’s. oHowever, the high variance indicates opportunities for targeted optimisation. §Energy Distribution: oCNNGen’s energy consumption spans a wide range. oUseful for building robust energy-prediction models. Greener Neuroevolution via Prediction 13 §Evolving the population with NSGA-II §Fitness function based on ML predictors (energy & accuracy) oNo need to train individuals (saves energy) oFaster (6x) oBut at the initial cost of predictor training on 40k models Energy consumption halved for a 2% performance decrease Contributions & Collaborators §Antoine Gratia, Hong Liu, Shin’Ichi Satoh, Paul Temple, Pierre-Yves Schobbens, Gilles Perrouin, “CNNGen: A Generator and a Dataset for Energy-Aware Neural Architecture Search”. ESANN 2024, Oct 2024, Bruges, Belgium. pp.173-178. §Gratia, A., Temple, P., Schobbens, PY., Perrouin, G. (2025). Energy-Aware Neural Architecture Search: Leveraging Genetic Algorithms for Balancing Performance and Consumption. Artificial Life and Evolutionary Computation Springer, 2025. §Antoine Gratia, Topological Architecture Exploration and Neuroevolution for Energy-aware Neural Architecture Search, PhD Thesis, September 2025. 14 Antoine Gratia Paul Temple Pierre-Yves Schobbens Shin’ichi Satoh Energy-efficient Neural Networks 15 Green ML / DL Parcimonious IA / green AI •challenge = AI is a huge consumer of data and energy •goal = reduce this consumption with no/small performance loss •research ideas •approximate the true objective function with something that is easier •create new architectures (e.g., based on conditional computation) •simplify models / focus on small models (e.g., SLMs) •create new, more efficient learning algorithms (w.r.t. data and energy) •links with practical applications (astronomy, sign language recognition, time series analysis, industry 4.0, etc.) 16 Application to CNNs DRAGUET, Noémie ; Frénay, Benoît. Making Convolutional Neural Networks Energy-Efficient : An Introduction. Proc. ESANN. 2025. 17 Understanding Software Energy Consumption 18 19 “We research emerging software quality properties through the lenses of software reliability and developers’ experience” https://snail.info.unamur.be/ •Green coding •Monitoring code energy consumption •Raising developers’ awareness •Studying the impact of a code change •https://snail.info.unamur.be/tag/energyconsumption/ Energy Codesumption (DevOpsSustain’25) 20 https://snail.info.unamur.be/publication/maquoi-2025/ Energy costs may be caused by the quantity and complexity of generated attributes inside constructors Energy Codesumption, Leveraging Test Execution for Source Code Energy Consumption Analysis Jérôme Maquoi [email protected] NADI, University of Namur Namur, Belgium Maxime Cauz NADI, University of Namur Namur, Belgium [email protected] Benoît Vanderose NADI, University of Namur Namur, Belgium [email protected] Xavier Devroey NADI, University of Namur Namur, Belgium [email protected] Abstract The software engineering community has increasingly recognized sustainability as a key research area. However, developers often have limited knowledge of e�ective strategies to reduce software energy consumption. To address this, we analyze energy consumption in software execution, aiming to raise developer awareness by linking energy consumption with each line of code. We rely on unit test executions to identify energy-intensive executions and manually analyze �ve hot and �ve cold spots to identify potentially energy-intensive source code constructs. Our �ndings suggest a link between the energy consumption of the source code and the number of objects’ attributes created within that code. These results lay the groundwork for further analysis of the relationship between object instantiation and energy consumption in Java. CCS Concepts •Hardware ! Power estimation and optimization;•Software and its engineering !Software testing and debugging. Keywords energy consumption, source code analysis, test execution, java ACM Reference Format: Jérôme Maquoi, Maxime Cauz, Benoît Vanderose, and Xavier Devroey. 2025. Energy Codesumption, Leveraging Test Execution for Source Code Energy Consumption Analysis. In 33rd ACM International Conference on the Foundations of Software Engineering (FSE Companion ’25), June 23–28, 2025, Trondheim, Norway. ACM, New York, NY, USA, 5pages. https://doi.org/10. 1145/3696630.3728707 1 Introduction Reducing the environmental impact of all aspects of IT, including energy consumption of running software, is becoming crucial to Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for pro�t or commercial advantage and that copies bear this notice and the full citation on the �rst page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior speci�c permission and/or a fee. Request permissions from [email protected]. FSE Companion ’25, Trondheim, Norway ©2025 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 979-8-4007-1276-0/2025/06 https://doi.org/10.1145/3696630.3728707 achieving a more sustainable society. In recent years, various aspects of software sustainability and green software engineering have been investigated by the research community [ 4 , 20 , 22 – 24 , 27 ]. In particular, [ 17 ] has shown that developers recognize, to a certain extent, the challenges associated with software energy consumption. However, they often lack knowledge of e�ective strategies to reduce their software’s energy footprint. To bridge this gap, we need (i) methods to measure and report source code energy consumption and (ii) identify code constructs that increase energy usage to develop appropriate tooling. This short paper lays the foundations for our research objective: gaining a deeper understanding of the underlying causes of energy consumption from the source code. More speci�cally, we answer the following research question (RQ): how does the source code of a Java project impact its energy consumption? For that, we rely on JoularJX [ 16 ], a state-of-the-art tool, to measure energy consumption at the source code level of Java projects (i.e., each execution branch is associated with its energy consumption). Based on the measurements, we identi�ed highand low-energy-intensive parts of the code and performed a manual analysis to identify recurring code constructs. 2 Background and related work A recent survey [ 13 ] highlights the need for improved skills among developers in energy-aware development. Developing such skills should be paired with raising developer awareness about the energy consumption of their code, which requires tools that can assess the energy consumption associated with the source code. Energy consumption assessment. The �rst method relies on theoretical models to estimate consumption without direct measurements. For example, the TEEC model estimates CPU, memory, and disk power usage during application execution through established mathematical expressions [ 1 ]. Alternatively, physical measurement utilizes a power meter connected to the hardware. This approach provides the most accurate representation of energy consumption during software execution. For instance, a framework that detects energy hotspots in Android applications employs a physical power meter for accurate detection of these energy-intensive areas [ 3 ]. More recently, hardware sensors with software interfaces have been able to measure components like CPUs, GPUs, RAMs, and disks. Manufacturers provide interfaces for power data, such as Intel’s COST Actions RAFAEL CAPILLA REY JUAN CARLOS UNIVERSITY (MADRID, SPAIN) [email protected] ECSS Annual Summit, October 2025 About myself 2 •Full professor at Rey Juan Carlos University of Madrid (Spain) •Adjunct professor at Lappeenranta University of Technology (LUT, Finland) •PhD. in Computer Science What are COST Actions ? •Interdisciplinary research network…but focused on networking rather than on research •Investigate a topic for 4 years with focus on innovation and excellence •A pan European environment to connect experienced and young researchers •Grow professional research networks and boost careers •Open to all scientific and technological fields •Gender (49% are women) and geographical inclusive •A career accelerator for young researchers 3 The network •At least 50% of the countries must be ITC (Inclusiveness Target Countries) •ITC: Bulgaria, Croatia, Cyprus, Czech Republic, Estonia, Greece, Hungary, Latvia, Lithuania, Malta, Poland, Portugal, Romania, Slovakia and Slovenia •EU members outermost regions: E.g. Canary Islands (Spain), Madeira (Portugal) •Full members not belonging to the E.U.: Albania, Armenia, Bosnia and Herzegovina, Georgia, Moldova, Montenegro, North Macedonia, Serbia, Türkiye, Ukraine •In principle, as many partners (institutions) and proposers you want •One Main Proposer (MC) but you can have Secondary Proposers (MC) that are allowed to attend the Action Management Committee meetings •Types of organizations: Universities and research centers, SME, large companies, EU institutions, EU agencies, European RTD organizations, Others 4 Main documents •Technical Annex: Strict limit of 15 pages •3 main sections: EXCELLENCE IN S&T AND NETWORKING, IMPACT, IMPLEMENTATION •Excellence in S&T: Define the goals and challenges of the network, objectives, rationale for funding, state-of-art, and critical mass •Impact: Impact regarding the objectives, stakeholders, and communication/dissemination •Implementation: Workplan and tasks, Deliverables, GANTT •New template since 2025 •COST policies document •Summary and references in text boxes •The network where all partners and proposers must be invited manually one by one using their name and email 5 Timeline •The submission is typically in the second half of October •They answer by May/June the year after 6 Success of proposals •Typically, 39% of the proposals are accepted •Threshold 35 points over 50 •Each item in the proposal is ranked from 1 to 5 points •We need items to be evaluated with 4/5 points 7 Funding •Funding this year has been reduced from approximately 150-165K per year to 125 K the first year and 150 K the other for years •Funding depends on the number of COST countries represented in the Working Groups •Funding covers expenses for networking rather than research (e.g. Short-term Scientific Missions, Training Schools, communication activities, grants to attend interesting international conferences, and virtual networking tools) 8 The GREEN-T COST Action proposal •A COST action network focused on GREEN IT/ICT technologies with focus on energy estimation metrics/technologies and sustainable goals •We aim to suggest metrics and indicators (KPIs) to measure carbon emissions and energy consumption and advance existing green measures and measurement techniques •We pursuit expanding existing ICT-based solutions for detecting deviations from green practices like greenwashing •Create a network of researchers in green practices 9 Origins •The action comes from the Informatics Europe GREEN ITC Working Group led by Marco Aiello 10 2 Roberto Verdecchia et. Al. -Mixed-method empirical study -Conducted in 2020/2021 -48 unique participants from the Netherlands 3 “What is the future of sustainable digital infrastructures ?” •What are the solutions to develop sustainable digital infrastructures? •What drives the adoption of sustainable digital infrastructure solutions? •What hinders the adoption of sustainable digital infrastructure solutions? •What are the open problems related to sustainable digital infrastructure solutions? Research Question 4 Study Process Results from the previous study by Roberto Verdecchia et. Al. 6 Roberto Verdecchia et. Al. Today Near future Possible future 9 Roberto Verdecchia et. Al. 11 Roberto Verdecchia et. Al. Today Near future Possible future European Study First insights 13 The European Study By leveraging the Green ICT WG of Informatics Europe, we aim to replicate and update the study from a European perspective. 1. Conducting the interviews across Europe 2. Focusing on how the landscape changed from 2022 to 2025 3. Review if the landscape is different across parts of Europe 14 The European Study North: •Denmark West: •Germany •The Netherlands South: •Spain •Italy Widening countries: • North : Czech Republic • West : Portugal • South : Greece 21 metrics, tools, measuring tools, design, awareness, regulation, legislation, fair competition, positive standards, sustainability consideration, environmental regulations, efficiency thinking, rethrinking data center proposition, mindset shift Discussed Adoption Factors based on semantic text embeddings, unsupervised clustering, and keyword extraction. 22 technology obsolescence, upskilling of people, cloud provider emissions, measurement data transparency, Fairness, cost of sustainable data centers, artificial intelligence, sustainable ai Discussed Open Problems based on semantic text embeddings, unsupervised clustering, and keyword extraction. Questions? Thank you [email protected] https://doi.org/10.1016/j.suscom.2022.100767 previous study Green ICT Working Group Green ICT White Paper ©Marco Aiello, 2024 Goal To articulate a shared European vision for Green ICT, grounded in state-of-the-art research and best practices, and to present a practical framework that serves as a blueprint for designing, developing, and governing sustainable digital systems in Europe. With the option of using it as a recommendation for the EU. ©Marco Aiello, 2024 Sections 1.Introduction 2.A Framework for Green Accountability 2.1.Green metrics 2.2.Measuring and Carbon Emissions 2.3.Programming Languages and framework choices 2.4.AI and Machine Learning 2.5.The Role of Clouds and Data Centers 2.6. The Framework 3. Related Works 4.Threats to validity 4.1.Virtualization and containerization 4.2.Precise accounting 4.3. Green-washing 5. Recommendation ©Marco Aiello, 2024 Status 1.Introduction 2.A Framework for Green Accountability 2.1.Green metrics 2.2.Measuring and Carbon Emissions 2.3.Programming Languages and framework choices 2.4.AI and Machine Learning 2.5.The Role of Clouds and Data Centers 2.6. The Framework 3. Related Works 4.Threats to validity 4.1.Virtualization and containerization 4.2.Precise accounting 4.3. Green-washing 5. Recommendation ©Marco Aiello, 2024 What Do We Mean by 'Framework'? • A structured model connecting principles, methods, and measurable actions for sustainable ICT systems. • It provides: 1.A common language for researchers, policymakers, and industry. 2. A reference structure to guide design, implementation, and evaluation. 3.A basis for standards and certification. ©Marco Aiello, 2024 Conceptual Layer Structural Layer Operational Layer Governance Lifecycle Management Resource efficiency Data Minimalism Transparency/ Traceability Circular ICT Economy Governance standards/ incentives Infrastructure (cloud/edge) Data (transfer/ storage) Software eaware design Design Principles Sustainable Data and SW Lifecycle Energy and carbon Metrics Synergy with Governmental Initiatives and Regulations Certification and Benchmarking Policies definition and monitoring PRINCIPLESDOMAINSACTIONS ©Marco Aiello, 2024 Discussion Next Steps