Deliverables

This deliverable reports early results from WP1 on ML task modeling. It lays the conceptual and technical groundwork for capturing user problems, structuring ML knowledge, and guiding energy-aware model design – all of which are essential for SustainML’s interactive, human-centric framework.

The report argues that existing ML taxonomies are often too abstract, topic-specific or disconnected from real applications to support practical task modeling. SustainML instead combines a top-down view of ML paradigms with bottom-up, application-oriented studies and prototypes grounded in user needs.

Main objectives of WP1 at this stage are to (i) develop a task-centric knowledge base that links user problem descriptions to ML techniques, (ii) prototype interactive tools for exploring and inspecting models, and (iii) account for hardware and lifecycle constraints already in the early conceptualization and model-design phases.

The scope covers:

• analysis of related literature and toolkits (LLMs, existing taxonomies, community hubs such as Hugging Face);
• design of an ML task modeling knowledge graph;
• development of an interactive latent-space exploration tool;
  initial work on HW-aware ML design.

The methods / approach are structured around the full ML lifecycle – conceptualisation, model design, training & optimization, deployment – with WP1 focused on the first two stages but informed by downstream hardware and deployment constraints.

For the knowledge base, SustainML uses RDF and Turtle syntax to build a graph-structured taxonomy that can represent complex, interconnected ML concepts more flexibly than traditional trees or tables. Nodes represent data types, tasks, algorithms, and goals, while predicates encode relationships (hierarchies, similarity, processing chains). The graph is populated from literature taxonomies and publication metadata, then refined with expert input. SPARQL queries (via rdflib in Python) extract pathways from user-level problem descriptions to candidate solution pipelines (e.g., temporal data → STFT → spectrogram → image-like representation → CNN-based classifier).

The report critically examines LLM tools like ChatGPT: while they can generate plausible high-level workflows, they lack explicit reasoning, consistent detail, and often hallucinate. This motivates a curated, explainable knowledge system that can nonetheless be interfaced with LLMs for interaction.

The interactive heterogeneous multimodal tool, the “Interactive Latent Space Explorer”, enables practitioners to probe trained neural networks by manipulating latent activations and visualizing effects via dimensionality reduction. Its modular architecture is presented as a candidate host for embedding the knowledge base and supporting interactive, explainable task modeling.

Initial work on HW resource-aware design links ML models and tasks to hardware choices, particularly for embedded and sensor-rich scenarios (e.g., human activity recognition on MCUs/FPGAs). This aligns WP1 with later work packages on hardware-aware optimization.

Key results include: a first version of the ML task knowledge graph; a working prototype for RDF querying; progress on interactive tooling; and three related research outputs already accepted or submitted to international venues (ISBI, IJCAI, ASAP).

Conclusions and impact: D1.1 demonstrates that structured knowledge graphs and interactive tools can bridge the gap between user-level problem descriptions and detailed ML design choices, while remaining compatible with energy and hardware constraints. These components will underpin SustainML’s future user interface and recommendation mechanisms, enabling more transparent, sustainable ML design.

This deliverable contributes to the energy-optimized ML toolkit by introducing EC-NAS, a tabular benchmark and surrogate-model toolkit that incorporates energy consumption and carbon footprint as first-class objectives in Neural Architecture Search (NAS). It provides methods and evidence for energy-aware model selection and optimization.

The report starts from the observation that energy consumption has become a primary bottleneck in scaling AI: model development involves substantial compute and energy costs during dataset curation, architecture search, training, and deployment. Traditional NAS focuses on accuracy and sometimes runtime but rarely on explicit energy and carbon metrics, which are increasingly critical for sustainable ML.

Main objectives are to (i) integrate energy efficiency as an explicit optimization criterion in NAS, (ii) provide an open benchmark (EC-NAS) containing energy, performance, and resource data for many architectures, (iii) develop a surrogate model to predict energy use, and (iv) demonstrate how multi-objective optimization can discover architectures that balance accuracy and energy use with minimal performance loss.

Scope is centered on image classification on CIFAR-10 using the NAS-Bench-101-style cell-based search space, extended with energy measurements. Architectures are represented as DAG cells with up to seven nodes and operations including convolutions and pooling, yielding ~423k unique architectures and about 1.6M datapoints across multiple training budgets (4, 12, 36, 108 epochs).

The methods / approach are:

• Measure training energy on a reference SLURM cluster (RTX 6000 GPU + Intel CPUs), using Carbontracker to track GPU, CPU, and DRAM power, and convert to kWh and kgCO₂eq.
• Train a four-layer MLP surrogate model that maps architectural features (adjacency matrix, operations, parameter count) to predicted energy. Using ~4,300 sampled architectures, the surrogate achieves high correlation (Kendall-Tau ≈ 0.90) between predicted and measured energy, enabling efficient estimation across the full space.
• Analyze the dataset to understand how graph size, parameter count, and operator choices affect accuracy, training time, and energy, and evaluate hardware consistency across different GPUs.
• Apply multi-objective NAS methods (evolutionary MOO algorithms such as SEMOA, SH-EMOA, MS-EHVI, plus random search baselines) to jointly optimize validation accuracy and training energy, identifying Pareto fronts and “knee” architectures.

Key results show that:

• Individual operator substitutions can significantly influence energy and accuracy without necessarily changing training time or parameter count, highlighting non-trivial design trade-offs.
• MOO algorithms discover architectures that substantially reduce energy consumption with only small accuracy drops compared to energy-hungry baselines. The Pareto analysis reveals architectures that are near-optimal in both dimensions, suitable for real-world deployment.
• The surrogate model enables large savings in GPU days and kWh when compared to brute-force measurement over the full search space.

The report also discusses Carbontracker itself—its capabilities for real-time monitoring, carbon-intensity awareness, reporting and workflow integration—and advocates using energy consumption rather than FLOPs or runtime alone as a robust, hardware-aware efficiency metric.

Conclusions and impact: EC-NAS provides a concrete tool and dataset for energy-aware NAS within SustainML and the wider community. It demonstrates that energy-efficient architectures can be systematically identified without sacrificing performance, and that integrating tools like Carbontracker into the development loop can raise awareness and steer practice toward lower-carbon ML.

This deliverable defines how SustainML will handle all research data over the project’s life: what is collected or generated, how it is stored and shared, how FAIR principles are applied, and how ethical and legal requirements are respected. It is the foundational policy document for data governance.
The DMP begins by restating SustainML’s goal of developing an energy-efficient ML design framework and toolkit, and notes that this will generate diverse data: ML task parametrizations, catalogs of efficient practices, hardware libraries, SustainML framework source code, and validation results. As part of the EU Open Research Data Pilot, the project commits to making data “as open as possible, as closed as necessary”, subject to GDPR, IPR, and security constraints.

Main objectives are to describe the data lifecycle, define open access policies, document dataset types and formats, ensure FAIR (Findable, Accessible, Interoperable, Re-usable) data, allocate resources for data management, and address security and ethics, including personal data and informed consent.

The scope covers all datasets collected, processed, or generated during SustainML: software artefacts, HW libraries, training/evaluation datasets, experimental results, scientific publications and associated metadata. It also covers both internal (consortium-only) and public datasets, with the DMP as a living document updated annually.

The methods / approach follow a “Big Data Life Cycle” with phases for planning, acquiring, preparing, analyzing, preserving, and discovering data, supported by continuous quality assurance. Data are summarized per dataset using a structured template (identifier, description, type, formats, generation mechanism, expected size, quality control, access type, ethical issues). The plan sets naming and versioning conventions (e.g., SustainML_YYYYMMDD_DatasetName_Type_vX.Y) and mandates rich metadata including DOIs, funding references, and persistent identifiers.

To make data findable and accessible, SustainML will:

• use standardized metadata and keywords;
• store internal data in consortium repositories (institutional storage, shared drives, private website areas);
• publish open data via platforms such as Zenodo, arXiv, GitHub, YouTube, and the SustainML website;
• adopt common file formats (PDF, PNG/SVG, WAV/MP3, MP4/WEBM, IDL, source-code text, CSV/JSON).

Interoperability is supported by using open, widely supported formats and a standardized vocabulary informed by the DataCite metadata schema; this is particularly important because the project spans AI, HCI, hardware design and sustainable computing.

To increase re-use, the project plans to release datasets under Creative Commons licences (e.g., Attribution-ShareAlike or Attribution-NonCommercial-ShareAlike), limit embargoes to a maximum of six months, and ensure data remain accessible for at least one year after project completion, in line with grant conditions. Data owners are responsible for quality assurance, regular updates, and ensuring that datasets contain no erroneous or obsolete records.

The DMP also sets out resource allocation (publication costs, hosting, repository maintenance) and tackles data security and ethics, including anonymisation procedures, treatment of sensitive personal data, and informed consent templates.

This deliverable defines how SustainML presents itself to the outside world, builds its brand, reaches target audiences, and prepares the ground for exploitation. It sets the initial communication, dissemination, and exploitation strategy and reports activities from the first six months.

The report establishes the overall strategy for communicating SustainML’s vision: an application-aware, lifecycle-oriented framework and toolkit that improves the energy efficiency of ML systems from design and exploration through training and deployment. It clarifies that early-phase work focuses more on awareness and scientific dissemination than on end-user adoption, since the framework is still under development.

The main objectives are to define clear messages about SustainML’s mission and results, support collaborations with related initiatives, make results understandable to technical and non-technical audiences, attract researchers and early adopters, and prepare the basis for future market exploitation. Communication objectives (CO1–CO6) and dissemination objectives (DO1–DO6) translate these ambitions into concrete goals such as publishing research, building communities, and segmenting markets.

The scope covers brand positioning, target audiences, channels, content, and KPIs. SustainML positions itself as an open-source, Green-AI oriented framework that combines application task understanding, multi-stack optimization methods, and efficient hardware layers. The brand messaging framework sets vision, mission, core promise, and brand pillars (open source, affordable, transparent, flexible, user-friendly, explainable, green, plug-and-play, secure).

The methods and approach include a marketing funnel (Awareness → Interest → Consideration → Conversion → Adoption → Retention → Advocacy) and the PESO model (Paid, Earned, Shared, Owned media) to structure activities. The report segments target audiences into AI researchers, ML developers from novice to expert, SMEs and large enterprises, and key communities such as ROS, AI4EU, and Human-AI networks.

The key results for months M1–M6 are the creation of a visual identity and logo variants; definition of colors, typography, and templates; promotional posters and materials for events; internal support tools (report templates, email and contact templates); presence at selected events and seminars; initial scientific publications; and baseline analytics for social media, website, and community posts. A roadmap lays out phased activities: initial awareness and brand cognition; then community building and consideration; and later, conversion and adoption once tools are mature.

The exploitation plan introduces SustainML’s mission and vision, target customers, competition and SWOT analysis, unique selling propositions, and individual exploitation strategies per partner (eProsima, DFKI, RPTU, KU, Inria, IBM, UPMEM). It stresses open-source licensing, long-term maintainability and community building, and leveraging existing ecosystems such as ROS.

The conclusions and impact are that D7.2 gives the consortium a shared playbook for how to talk about SustainML, where to do it, and how to measure impact. By aligning brand, messaging, channels, and exploitation paths early, it aims to maximise visibility, support future adoption of the SustainML framework, and ensure that research outputs are positioned for real-world uptake and commercial follow-on.

Efficient implementation of Deep Neural Networks (DNNs) in hardware requires rigorous exploration of the design space on different layers of abstraction, including algorithmic, architectural, and platform layers. At the highest level of the design hierarchy is the algorithm, which is the most abstract description of the data and control flow in the form of a DNN topology. The architecture layer maps the topology to a hardware design, which is implemented on the platform. At the lowest level is the platform, which describes the hardware and its physical properties.

As the DNN topology is on the highest level of the design hierarchy, the changes applied on the topology have potentially the highest impact on the properties of the final implementation. DNNs have to be designed to achieve the required accuracy and fulfill hardware criteria, especially on edge devices because they have small memory, constrained computing capabilities, memory bandwidth, power, and energy budgets. The analysis shows that arithmetic operations are “cheap” while memory accesses are “expensive” and have a cost that is a function of the size of the memory being accessed. The relative energy cost should be used as the main guidance for designing DNNs. The hardware-aware DNNs have to be small to fit into on-chip memory ideally, or to mitigate any communication with the external memory, they have to use fewer operations and leverage low-precision data types.

There is a lot of ongoing research on developing networks with lower computation costs and storage consumption without impairing classification accuracy. The efficient models became possible due to multiple macro- and micro-architectural improvements of the models. The types of layers and their arrangement are referred to as macro-architecture. The efficient micro-architectural approaches are:

1. very deep models are replaced with fewer layers but with more channels,
2. activation feature maps are kept smaller,
3. models are enhanced with skip and residual connections that have been proven to improve accuracy,
4. standard convolutions are replaced with depth-wise separable ones.

The micro-architecture also defines methods applied to individual layers, like replacing big convolutional kernels with smaller ones and fusing different layers. Further techniques have been proposed to alleviate the computing and storage challenges. Among the most common ones are distillation [6], pruning [7, 8, 9], and quantization or a combination thereof [10]. The resulting model after knowledge distillation does not require special treatment during inference, unlike after pruning and quantization. In summary, pruning and quantization are the primary model compression techniques, but they require special treatment which raises the question of selecting an implementation platform that can fully benefit from them.

Comparing different hardware platforms — Central Processing Unit (CPU), Graphics Processing Unit (GPU), Field-Programmable Gate Array (FPGA) and Application-Specific Integrated Circuits (ASICs) — one can see that these options present a trade-off between flexibility and efficiency. While CPUs and GPUs are highly flexible, they cannot fully benefit from the major optimization techniques, like pruning and quantization. Only ASICs and FPGAs can fully benefit from them. The main reasons why FPGAs and ASICs are highly efficient in comparison to general-purpose computing platforms are:

1. optimized memory,
2. data specialization,
3. massive parallelism,
4. reduced overhead,
5. algorithm-architecture co-design.

Using these advantages, one can use FPGAs to provide ad hoc solutions to facilitate computationally intensive, time-critical tasks at low-power consumption in a reprogrammable manner, unlike ASICs. In summary, FPGA is a computing platform with a unique combination of programmability, development cost, and efficiency that can fully benefit from various compression techniques. How can one use these features to implement optimized DNNs efficiently?

The design of DNN topologies, custom hardware architectures for DNNs, and their implementation on FPGA is a time-consuming process. To efficiently implement DNNs on a specific FPGA platform and to meet certain requirements, e.g., power consumption and latency, we have to consider an enormous amount of design parameters starting from neural topology down to hardware architecture and physical implementation. Importantly, interdependencies between the different design layers have to be considered, making it impossible to find optimal solutions manually. Fast and efficient implementation of DNNs on FPGAs can be achieved by combining recent advances

1. co-design, meaning DNN topologies and hardware architectures have to be co-designed by joint optimization of performance and efficiency while maintaining accuracy, which can be achieved using Neural Architecture Search (NAS);
2. parameterizable hardware templates to build libraries of hardware components that support a wide range of parameters of various layers;
3. facilitated hardware design using, e.g., High-level Synthesis (HLS) to accelerate the development process.

There are many techniques for automatic exploration of the vast design space of DNNs. Among the most successful approaches are:

1. Reinforcement Learning (RL),
2. gradient-based methods,
3. evolutionary algorithms.

Evolutionary algorithms do not require training an agent, like RL-based methods, nor a supergraph, like gradient-based methods. It was shown that the evolutionary algorithm can outrun RL-based methods and outperforms state-of-the-art hand-designed models. The proposed methodology is based on the evolutionary algorithm presented in [30]. The main techniques responsible for the high efficiency of the NAS implementation are:

1. multi-objective Pareto optimization, which enables multi-criteria optimization;
2. Bayesian sampling method to improve the candidate selection process;
3. evaluation of the candidates using “cheap” and “expensive” objectives to accelerate the evaluation process.

Therefore, the NAS finds fully hardware-compatible solutions that are optimal with respect to the hardware optimization objectives and fulfill the application requirements.

To enable cross-layer optimizations, we present a flexible HLS hardware library of highly customizable hardware architectures, which can facilitate various DNN topologies. The hardware library is written as a collection of C/C++ template functions with HLS annotations and modularity in mind to make it easily expandable by new layers. The hardware architecture is designed to be low power and ultra-low latency. Primarily, this is achieved by:

1. keeping all weights and intermediate results in on-chip memory since off-chip transfers consume more energy and introduce extra latency;
2. external memory is only used to read input data and write results, therefore reducing memory access to the absolute minimum;
3. separate hardware modules dedicated to each layer are connected using streaming interfaces to facilitate fast design, debugging, interoperability, and ease of integration;
4. the architecture is fully pipelined, allowing all layers to operate concurrently and start the computation as soon as the inputs are ready to reduce latency and energy consumption.

One of the recent emerging trends to increase energy efficiency is to adapt low-precision data formats for both inference and training of DNNs. Typically, training is done using floating-point formats as they are able to provide wide dynamic range and high precision. The main challenge of using low-bit-width data formats is the limited dynamic range compared to floating-point 32-bit (FP32). One possible solution to compensate for the range limitation of low-bit-width data formats is to shift the data format’s representable range to the desired location. This can be done by varying the bias value of the floating-point equation instead of using a common bias value for all values. Our methodology uses statistical analysis to find the optimum bias value for a given DNN model. By this method, we find the bias value which shifts the low-bit-width data format dynamic range to the location with the highest coverage for a given DNN model. We demonstrate the efficiency of the approach facilitating floating-point 8-bit (FP8) format for several datasets and state-of-the-art DNN topologies. This enables a linear reduction of the total number of Dynamic Random Access Memory (DRAM) accesses, which increases energy efficiency while keeping the accuracy on par with FP32 format.

To facilitate fast implementation of topologies found by the NAS and mapped onto custom hardware architectures, we implemented the Holistic Auto machine Learning for FPGAs (HALF) framework that consists of two main components, which are the hardware-aware NAS and the FPGA implementation framework. The framework automatically produces a hardware implementation for the selected FPGA platform that fulfills the requirements. The HALF framework accelerates the design cycle significantly; it reduces the deployment time from months to days.

This report corresponds to Deliverable D4.1 - Exploring new ML models of the SustainML project and is structured into four parts: First, we report qualitative studies to better understand the current  awareness of ML and HCI experts on their impact on sustainability. We then present a framework that structures the different intersections between sustainability with ML and HCI and describe the resulting research areas based on recent work. In the third part, we focus on HCI for sustainable ML and how human-centered design practices can be used to support more sustainable ML development throughout the entire life-cycle of systems. We define various impacts of ML systems and propose strategies to mitigate these impacts, including data collection, model training, and recycling. To anticipate later stages and impacts in the life-cycle, however, requires ML experts to express realistic project requirements before starting a ML project, which can be challenging. Following a more human-centered design approach, we examine how current decisions in the early stages are made based on interviews with eight ML experts, which we describe in the last part of this deliverable.

In order to promote energy efficiency and avoid AI-waste throughout the entire life-cycle of ML applications, however, developers need the ability to make environmentally conscious decisions from the start. Our analysis of current sustainability awareness and existing work at the intersection of HCI and ML shows the potential for using HCI methods to redefine the life-cycle of ML models by supporting sustainable actions through human-computer collaboration. Our final objective is to create tools that help ML practitioners understand how a certain decision affects the environment and guide them in identifying suitable, more environmentally friendly models for their projects. This deliverable provides the background for developing such tools, which will be presented in later deliverables.

The SustainML project aims to develop a design framework and associated toolkit to improve energy efficiency throughout the entire life cycle of machine learning applications: from the design and exploration phase to initial iterations of training, testing and optimizing various system versions until final training in the production systems (often requiring large amounts of data, computations, and epochs) and (where appropriate) continuous online retraining of the inference process during deployment.

The resulting product SustainML Design Framework, aimed to be developed in this project, will provide an energy optimized hardware solution with the corresponding Machine Learning model for solving a desired machine learning problem formulated by the user.

Given the distributed nature of the SustainML ecosystem, the use of DDS (Data Distribution Service) is proposed to enable real-time communication between the different SustainML software modules (nodes). The Data Distribution Service is a data-centric communication protocol used for distributed software application communications. It describes the communications Application Programming Interfaces (APIs) and Communication Semantics that enable communication between data providers and data consumers.

The proposed software design will remain essential to ensure that the final SustainML Design Framework meets user needs, is adaptable to future changes, and exhibits high performance by carefully considering factors such as modularity, scalability, security, and user experience. UML diagrams and SOLID design principles will be used to assure the code quality.

Deliverable D5.1 Framework architecture, back-end, front-end design and release establishes the software requirements and covers the detailed architecture and design for the different components within the SustainML Design Framework, as well as the release procedure. This document will be updated every year to reflect the most recent state of the framework as the project progresses.

This report corresponds to Deliverable D7.7 - Annual report on Collaboration - Y1 of the SustainML project. This deliverable covers the main collaborations between SustainML consortium members and organizations, EU projects and communities external to the project.

The report provides a comprehensive account of research collaborations during the specified timeframe. It highlights partnerships with various research groups and the ROS community, shedding light on their objectives, activities, and contributions to the AI applied to robotics field.

Core WP3 deliverable advancing “resource-aware training cycle modelling”. It contributes concrete methods to reduce data, label and memory requirements during model training, directly supporting SustainML’s goal of lower energy and carbon footprints.

This deliverable tackles the resource costs of the training phase in the deep-learning lifecycle by proposing three complementary techniques that treat data, labels and memory as critical resources. The aim is to enable similar or improved performance with fewer training examples, fewer labels, and lower memory consumption, thereby reducing compute and energy usage.

The first method introduces dataset condensation with robustness guarantees. Using Minimum Finite Coverings (MFC), it constructs small synthetic or coreset datasets that cover the “robust frontier” of the original data, balancing performance, efficiency and adversarial robustness. Experiments on MNIST, CIFAR-10/100, SVHN and TinyImageNet compare MFC-based coresets with distribution-matching and gradient-based condensation methods. Results show that the proposed coresets achieve superior robustness–accuracy trade-offs under adversarial training, revealing that many existing condensation methods overfit to standard accuracy and degrade robustness.

The second method proposes an efficient self-supervised framework for image segmentation. A small fully convolutional network (≈10.8k parameters) is trained contrastively on image patches, using confidence maps and an entropy-based patch sampler to derive positive and negative pairs without labels. Applied to nuclei segmentation in histopathology images, this approach achieves performance comparable to supervised and more complex self-supervised baselines, while reducing training time by a factor of about 144 and converging in minutes on high-resolution datasets. This demonstrates substantial savings in compute and energy relative to conventional pretext-task-based methods.

The third contribution is an activation-compression scheme for graph neural networks using block-wise quantisation and improved variance minimisation. By quantising activation maps during training, the method reduces memory usage by around 15% across several graph ML tasks while maintaining competitive accuracy. This has direct implications for running graph models on memory-constrained hardware and for lowering the energy cost of training.

In addition, the report outlines synergies with other SustainML partners, including extending activation compression to vision transformers, applying minimal covering sets to civil-inspection data, and building surrogate models for FPGA energy consumption. Collectively, these methods form a toolbox of resource-aware training techniques that can be integrated into the SustainML framework, enabling developers to choose greener training pipelines without sacrificing robustness or task performance.

Initial WP4 deliverable that translates earlier qualitative studies into design implications and a first prototype concept for interactive, explainable tools helping ML practitioners make sustainable, energy-aware modelling decisions within the SustainML platform.

The deliverable addresses how interactive systems can support ML practitioners in understanding sustainability trade-offs and choosing greener models without treating AI as a black box. It focuses on two challenges: (1) helping users articulate objectives and constraints for a given ML project, and (2) supporting them in evaluating and selecting environmentally friendly models that meet those objectives.

The scope spans the full ML lifecycle, from problem understanding through data preparation, modelling, evaluation and deployment, viewed through a sustainability lens. Building on extended interview studies with 13 ML experts across domains (biology, medicine, CPUs, etc.), the report extracts design implications for each lifecycle stage, such as enabling users to define project needs and constraints, explore and profile data with sustainability indicators, compare model performance and energy usage, configure efficient computing environments, learn from others’ experiences, and iteratively adjust models to reflect performance–sustainability trade-offs.

The methods comprise qualitative analysis, formulation of design implications, and conceptual design of a prototype decision-support tool. The prototype structure revolves around three sustainability-critical aspects: data (uploading or reusing datasets, assessing their storage and processing impact), problem (capturing task type and project-specific constraints via knowledge-graph-driven questions), and environment (hardware, deployment scenario, and operational constraints). Design rationales emphasise providing rather than prescribing information, visualising interconnections between data, goals and environment, decomposing sustainability impact (training vs inference vs hardware), and supporting human decision-making instead of delivering a single “optimal” answer.

The prototype is conceptually embedded in the wider SustainML platform: it takes inputs from the task-modelling and benchmark components (e.g. domain-specific baselines) and outputs guidance and explanations that help users refine project definitions and understand energy consequences of different choices. Paper prototypes illustrate the interaction flows and layouts for expressing requirements, exploring options and inspecting trade-offs.

Key results are: a refined research objective (three concrete research questions about baselines, visibility of interconnections, and impact of transparency on behaviour), a structured set of design implications aligned with the ML lifecycle, and a coherent prototype concept ready for implementation and user evaluation in subsequent deliverables. The impact lies in operationalising “explainable sustainability” for ML by designing human-computer collaboration mechanisms that can later be integrated into the SustainML framework and shared through tutorials and tooling for the broader AI community.

Updated Data Management Plan for WP7, defining how all research data from SustainML is collected, stored, shared and preserved according to FAIR principles and legal/ethical requirements at month 18.

This deliverable describes the lifecycle of data in SustainML and the policies governing their management, focusing on making research outputs Findable, Accessible, Interoperable and Reusable (FAIR) while respecting GDPR, IPR and ethical constraints. It updates the initial M6 DMP with refined FAIR data policy, ethical considerations and an expanded inventory of project datasets.

The scope covers all data produced or reused in SustainML: software and hardware libraries, task models, ML benchmarks and catalogues, framework source code, experimental results, and documentation. A Big Data life-cycle model (plan → acquire → prepare → analyse → preserve → discover/describe → assure quality) is adopted to manage the significant volume and heterogeneity of data.

The methods/approach are structured around:

• Open access policy: In line with Article 17 of the Grant Agreement, all deliverables and datasets are intended to be openly accessible unless restricted for confidentiality, security, IPR or commercialisation reasons. Conflicts between openness and protection are evaluated per dataset, with options such as delayed release or patenting.
• Dataset description template: Each dataset is documented with a rich schema (ID, lead partner, WP/task, description, data types/formats, generation mechanisms, size, quality control, access type, ethics, metadata, storage locations, reuse plans, licensing and costs). This template unifies how partners describe and assess datasets.
• FAIR implementation: The plan details naming conventions (including versioning rules), use of persistent identifiers (DOIs), metadata standards, and keyword strategies to ensure findability. It specifies open formats (e.g. PDF, PNG/SVG, CSV/JSON, IDL, standard source-code formats) and preferred repositories such as Zenodo, arXiv, GitHub and institutional repositories for public access, while Google Drive and institutional data centres are used for internal sharing.

Data security measures address controlled access to internal repositories, secure storage media and responsibilities for data stewardship. Ethical aspects cover self-assessment, anonymisation procedures, handling of any personal or sensitive data, and informed-consent requirements, with an annexed consent-form template to ensure compliance with GDPR.

The impact of the DMP is to provide a unified, project-wide framework for responsible data handling that both supports open science and protects legitimate interests. It facilitates reuse of SustainML datasets by the consortium and external researchers, enhances transparency of results, and underpins long-term preservation of key artefacts and benchmarks beyond the project’s lifetime.

Mid-term report for WP7 summarising how SustainML has been communicated, disseminated, and prepared for exploitation during the first 18 months. It refines the Dissemination and Communication Plan (DCP), reports on activities and KPIs, and outlines joint and individual exploitation plans.

The deliverable defines a comprehensive communication and dissemination strategy to maximise the impact of SustainML across its target audiences: AI researchers, developers at all expertise levels, SMEs and large enterprises adopting AI, and key communities such as ROS and Humane-AI-Net. The overarching objective is to promote the SustainML framework as a long-term, open-source, Green-AI platform rather than a short-lived EU project output.

The scope covers branding and positioning, communication objectives and key messages, channels and tools, concrete outreach activities, KPI monitoring, and exploitation planning. The strategy is structured using a marketing funnel (awareness → interest → consideration → conversion → adoption → retention → advocacy) and the PESO model (Paid, Earned, Shared, Owned media) to ensure a 360° approach.

Methods include defining brand pillars (open source, affordable, transparent, flexible, user-friendly, explainable, Green AI, plug-and-play, secure) and a messaging framework centred on SustainML as a life-cycle-oriented, energy-efficient ML co-design framework. The plan segments audiences, tailors key messages to the project phase (from research-heavy early stages to tool releases), and selects channels such as scientific publications, events, social media, the project website, ROS communities, newsletters and webinars. Visual identity elements (logo, typography, templates, promotional material) and internal tools (templates, guidelines) are also standardised.

Key results by M18 include: establishment of a clear brand and online presence; participation in relevant conferences and workshops; production of promotional materials; initial scientific publications highlighting early SustainML research; and social-media and website analytics that evidence growing awareness. The report documents events attended, publications, online metrics, and other posts, assessing them against predefined KPIs to identify strengths and gaps.

The exploitation section articulates a shared vision and mission, identifies target customers and stakeholders, and analyses competitors and SWOT aspects. It defines unique selling points (e.g., full life-cycle energy focus, human-centric interaction, open-source ecosystem) and describes how each partner plans to exploit results in its own context (e.g., commercial offerings, integration into existing communities or research agendas).

Overall, the deliverable consolidates SustainML’s narrative, tools and metrics for communication and dissemination, while laying out realistic paths to exploit project outcomes beyond the funding period. It positions SustainML as a credible, sustainable Green-AI initiative with a strong community-building component.

The Final WP1 report for tasks T1.1–T1.3 consolidates the task-modelling concepts, knowledge structures, and interactive tools that connect user-level problem descriptions with sustainable, hardware-aware ML design choices. This deliverable refines SustainML’s vision of task-centric, knowledge-driven ML design that explicitly considers energy and hardware constraints from the earliest stages of the ML lifecycle. It argues that existing ML taxonomies are either too abstract or too narrow and do not adequately support non-expert users in formulating problems and choosing efficient solutions.

The main objectives of this work are to build a scalable knowledge base for ML task modelling from user problem descriptions; design interactive tools that expose and manipulate latent knowledge in models; integrate hardware and resource considerations into early design; and explore new ways to use latent representations and LLM-based Retrieval-Augmented Generation (RAG) for more sustainable ML pipelines.

The scope spans the first phases of the ML lifecycle (conceptualisation and model design), while maintaining links to training and deployment constraints. A lifecycle diagram on page 9 shows this end-to-end view, mapping data requirements, computing environments, user groups, and sustainability concerns across stages.

• RDF-based Knowledge Graph: Encodes ML paradigms, data types, algorithms, and goals as subject–predicate–object triples. SPARQL queries support the exploration of alternative solution paths (e.g., temporal data → STFT → spectrogram → image-like representation → CNN). A graph visualisation is on page 10.
• Interactive Latent Space Explorer: A modular front-end/back-end tool for experts to inspect and manipulate latent representations of trained networks. Proposed as a host for integrating the knowledge base, allowing user interactions to trigger code execution and knowledge-graph queries.
• Hardware-aware ML Design: Demonstrated through the FieldHAR RTL framework for human-activity recognition with heterogeneous sensors. A block-diagram on page 13 shows hardware–software co-optimisation for efficient FPGA implementations.
• Embedding Space Exploration and Human-in-the-Loop (HITL) Training: Users inspect low-dimensional projections of latent spaces during training, adjust cluster structure, and feed changes back via knowledge-distillation.
• Latent Boost: Augments conventional loss functions with distance-based terms derived from latent space structure, improving convergence speed and cluster quality.
• RAG-based Web Agents (Talk2AIoD): Store web content as vector embeddings to reduce token usage and CO₂ emissions, providing a blueprint for later graph-RAG integration with the SustainML knowledge base.

Key results include a refined ML taxonomy embodied in a usable knowledge graph, a working prototype of the latent-space explorer, a hardware co-design case study (FieldHAR), and several associated publications and submissions. Collectively, these results show that explicit handling of knowledge—semantic, latent, and hardware-related—can reduce wasted experimentation and training.

In conclusion and impact, D1.2 demonstrates how knowledge-centric task modelling and interactive tools can guide users from informal problem descriptions to implementable, hardware-aware ML pipelines. By surfacing latent and structural knowledge early, the approach aims to shorten development cycles, lower energy consumption, and provide a solid basis for the implementation-oriented work in D1.3 and later Work Packages (WPs).

Initial implementation deliverable for tasks T1.4–T1.5 in WP1. It turns the conceptual task-modelling work of D1.1/D1.2 into a concrete knowledge base, integrating Hugging Face models, abstract task “bins” and Green-AI best practices into a queryable registry that SustainML tools can use.

The deliverables’ main objectives are to:

• define a model-centric RDF knowledge graph that captures models, problems, libraries and tags;
• implement knowledge sorting and abstract-level binning so that heterogeneous ML problems map onto reusable task categories;
• gather and encode efficient practices and sustainability-oriented methods as part of the problem-modelling space;
• integrate external model repositories (notably Hugging Face) and enable both structured querying and LLM-assisted Graph-RAG interaction with the graph.

The scope covers the base hierarchy of the knowledge graph, the technical pipeline for importing and structuring Hugging Face model metadata, traditional SPARQL-based querying functions, and a first Graph-RAG prototype. It also documents public code and graph artefacts via the Knowledge2Model GitHub repository, making the WP1 assets reusable by other WPs and external users.

In terms of methods / approach, the graph is organised around Model nodes as central entities, with relationships such as hasProblem, hasCoverTag, usesLibrary and hasTag pointing to Problem, CoverTag, Library and Tag nodes respectively. Attributes like downloads, likes or lastModified are stored as literals, enabling popularity- or recency-aware queries. This structure mimics Hugging Face’s categorisation while remaining flexible enough to host SustainML-specific concepts and efficient practices.

Hugging Face integration proceeds in two stages:

• Metadata ingestion – API calls obtain pipeline tags and model lists, which are written into an SQL table (model ID, problem, tags, coverTag, library, downloads, likes, lastModified). Frequent tags are filtered to avoid graph bloat.
• Graph population – SQL records are converted into RDF triples that connect each Model node to its problem, cover tags, libraries and selected tags.

A set of traditional querying functions (e.g., “get cover tags”, “get problems for cover tag”, “get models for problem”, “get model details”) is defined using rdflib and SPARQL, forming the backbone for tool integration and future UI work in WP4.

The deliverable also introduces a Graph-RAG prototype: an LLM-driven layer that turns natural-language questions into SPARQL queries while respecting the graph schema. Built with LangChain and OpenAI models, it aims to let users search for models and practices conversationally, with future work focused on assessing its sustainability compared to deterministic querying.

Key results are: a functioning knowledge graph and dataset, a pipeline for integrating Hugging Face models, a first set of efficient-practice encodings, implemented querying utilities, and initial Graph-RAG experimentation. Planned extensions include adding custom (non-HF) models, integrating WP3 energy-consumption data into the graph, and broadening the efficient-practice registry via systematic literature review.

D1.3 operationalises SustainML’s knowledge-centric vision by providing the actual graph, code and integration paths that later tools will depend on. It lays the groundwork for sustainable model search and selection, where users can compare models not only by task and architecture, but ultimately also by energy behaviour and recommended Green-AI practices.

WP2 deliverable that introduces an open-source toolchain to automatically generate FPGA-based hardware implementations for DNN models. It underpins SustainML’s ability to estimate latency, resource usage, power and ultimately carbon emissions of ML inference on FPGAs without requiring costly full implementation flows for every candidate design.

The deliverable addresses the challenge that accurate energy estimation for DNN inference on FPGAs typically requires full hardware synthesis, place-and-route and bitstream generation, steps that are too slow and compute-intensive to repeat for many architectures.

The main objectives are to:

• design an automated Hardware Exploration Toolchain that can generate, test and deploy FPGA implementations from high-level DNN descriptions;
• collect resource-utilisation and power-consumption data for many instances, enabling training of an “agent” that predicts hardware cost for new models;
• develop a hardware library suitable for low-power, low-latency implementations while supporting a broader range of DNN layers and data formats.

The scope covers DNNs described in ONNX format, hardware generation for AMD-Xilinx FPGAs, and exploration of different memory-mapping strategies (all weights and feature maps on-chip versus mixed on/off-chip storage). It focuses particularly on U-Net-style segmentation networks as a representative, memory-intensive use case.

The methods / approach are organised into four major components:

• Hardware Generator – parses ONNX models and application-level requirements (e.g., throughput) to produce C/C++ HLS code using a templated hardware library. Parallelism is chosen to meet performance constraints.
• Software testing – compiles the HLS code into a library wrapped for Python so that the original DNN testbench can be reused. The library supports FP32 for functional equivalence and multiple fixed-point and integer formats to explore quantisation strategies, including post-training quantisation pipelines.
• Implementation – uses Vivado HLS and Vivado (or an alternative Vitis-based flow under investigation) to generate RTL, perform synthesis and place-and-route, and produce a bitstream. Vitis additionally generates software for memory-mapped communication, at the cost of some kernel-launch overhead.
• Hardware testing – configures the FPGA and evaluates implementations on board using Pynq Overlay or Vitis-generated OpenCL software, measuring runtime and validating correctness.

The hardware library is based on a low-power, ultra-low-latency architecture that keeps weights and intermediate results in on-chip memory wherever possible. Diagrams on page 8 compare a fully on-chip design with an alternative architecture where selected buffers (e.g., U-Net skip-connection FIFOs) are moved to off-chip memory, freeing scarce on-chip resources for deeper models.

Key results are:

• a functioning, open-source hardware exploration flow from an ONNX model to a tested FPGA implementation;
• a reusable C/C++ HLS library for common DNN layers with flexible precision;
• demonstration that mixed on-/off-chip architectures can extend the range of deployable models on FPGAs;
• a clear roadmap for training a resource- and power-prediction agent using data generated by the toolchain.

D2.2 provides the hardware side of SustainML’s model–hardware co-design story. Automating FPGA implementation and preparing data for predictive agents enables rapid exploration of energy/performance trade-offs across many DNN variants. This capability will be crucial for later stages where SustainML recommends or automatically selects architectures and hardware mappings that minimise energy and carbon footprint while meeting application constraints.

D3.3, under WP3, develops practical “knowledge recycling” techniques that turn large vision and activity-recognition models into reusable functional cores. These methods make it cheaper and greener to adapt foundation models to new tasks and domains within the SustainML framework.

The deliverable addresses the high computational and energy cost of retraining ML models whenever tasks or domains change. It proposes to reuse functional knowledge cores—pre-trained components extracted from foundation models—and to adapt them efficiently using three complementary techniques: semantic-aware knowledge distillation, probabilistic feature matching for visual prompting, and parameter-efficient fine-tuning.

First, the TSAK (Two-Stage Semantic-Aware Knowledge Distillation) framework targets wearable Human Activity Recognition (HAR) in smart factories. A multi-modal teacher model combines accelerometer and capacitive sensor data and uses attention and causal encoders to capture rich temporal–spatial structure. A second stage distils this knowledge into a lightweight student model using five strategies, including attention, causal and combined representations and, most effectively, a semantic-classifier-based logit distillation (SC-Logit). TSAK reduces model size by 79 %, speeds inference by a factor of 8.88 and cuts computational power by 96.6 %, while improving F1 scores by up to 10 percentage points on both an in-house smart-factory dataset and the OpenPack benchmark.

Second, the Softmatcher framework introduces a training-free, probabilistic feature-matching method for visual prompting in segmentation tasks. Using frozen vision foundation models and Segment Anything (SAM), users provide simple scribbles on a reference image; SAM turns these into a reference mask, which is then probabilistically matched to target images. Optimal-transport-style soft matching with cosine similarity and Random Fourier Feature approximations yields scalable, low-latency segmentation, enabling interactive, real-time refinement through a web interface.

Third, the deliverable studies ensembles of Parameter-Efficient Fine-Tuning (PEFT) methods—LoRA, IA3, prompt tuning, and adapters—applied mainly to the SAM decoder head. By exploiting their orthogonality, combinations of these methods achieve better adaptation than any single PEFT technique, while tuning only a small subset of parameters. This preserves generalisation, reduces training cost, and offers strong performance even in low-data regimes.

Overall, D3.3 provides concrete, experimentally validated tools for reusing large models as functional cores, substantially reducing compute, energy and data needs for new tasks, in line with SustainML’s Green-AI objectives.

This WP4 deliverable presents the first working prototypes of SustainML’s ML-exploration interfaces and the study protocols to evaluate them. It focuses on how interaction design can help ML practitioners define and select models in a way that foregrounds sustainability (energy, CO₂ impact) alongside performance and usability.

The deliverable starts from the observation that today’s ML practice, especially with large, general-purpose models, has a rapidly growing environmental footprint. Inference costs of widely deployed models can surpass training costs, and the widespread use of multi-purpose LLMs for narrow tasks can be significantly more energy-intensive than task-specific models. The authors argue that a holistic, life-cycle view is needed, where sustainability considerations are integrated into early phases of problem definition and model selection, rather than treated as an afterthought.

Main objectives are to:

• design interactive tools that support developers in articulating ML problem requirements and constraints, including sustainability;
• prototype interfaces that help compare candidate models in terms of performance, interpretability and ecological impact;
• define empirical studies to assess whether these tools improve sustainability awareness, decision quality and efficiency compared to current practice (e.g., browsing Hugging Face directly or using a generic LLM interface).

The scope of the deliverable is twofold. First, it reviews related work on model selection practice, sustainable ML and sustainable HCI, highlighting gaps in how human decision-making and trade-offs (accuracy vs interpretability vs energy) are currently supported. Second, it describes two concrete interface prototypes integrated conceptually into the SustainML pipeline and the experimental protocols for their evaluation.

The methods / approach follow a user-centred design. The platform is described from a user’s perspective through persona-based walkthroughs:

• Prototype for model definition – An input-oriented interface guides the user through:
  • Data: upload a sample dataset, inspect columns, handle missing values and outliers, and confirm metadata.
  • Goal and domain: specify problem type (e.g., classification), data modality, task, and application domain (e.g., finance) to align with domain norms and regulations.
  • Environment: set geographical location, deployment constraints (e.g., local inference only), and hardware availability.
    The interface estimates CO₂ footprints for different decision dimensions and lets the user explore alternatives before requesting model suggestions, which are returned with environment parameters and predicted emissions and linked to Hugging Face.
• Prototype for model selection – A comparison-oriented interface visualises candidate models across multiple dimensions, such as ecological impact, requests per day, explainability and interpretability (as shown in Figure 3 on page 14). Users adjust priority sliders (e.g., accuracy, latency, training time, environment level) and interact with scatterplots to focus on models that fit both technical and sustainability preferences, then inspect detailed descriptions and CO₂ estimates before downloading.

For each prototype, the deliverable defines evaluation studies using within-participant designs with experienced ML developers. The first study compares three conditions: baseline Hugging Face website, an LLM-based interface from DFKI, and the guided interface. The second compares baseline Hugging Face with the adaptive comparison interface. Both use think-aloud protocols, log quantitative metrics (task time, energy used during exploration, final model consumption) and gather questionnaires and interviews, with a follow-up to probe longer-term effects on practice.

D4.4 delivers concrete, working prototypes and rigorous study plans rather than final user-study results. It advances SustainML by operationalising its human-centric, sustainable-ML vision into UI concepts that can be empirically tested and later integrated with the full pipeline. The expected impact is to show whether carefully designed interactive tools can steer developers toward more sustainable, better-justified model choices without sacrificing effectiveness.

D5.2, led by eProsima in WP5, defines the SustainML Design Framework software architecture. It specifies the back-end node-based system over DDS, the front-end web GUI, the orchestrator and data models, and describes how all WP1–WP4 components are integrated and released at month 24.

The deliverable refines and updates the initial architecture (D5.1) to reflect new requirements and integration progress. It first summarises major changes: new module nodes for application-level requirements and hardware constraints, updated data models, a Python-based orchestrator node that can run as a regular module, and extended front-end fields for richer user input and comparative result views. Several module nodes now embed partners’ real implementations instead of mocks.

The background section motivates the use of Data Distribution Service (DDS) and eProsima Fast DDS as the communication backbone. DDS provides a data-centric publish/subscribe model, QoS policies (durability, reliability, history) and multiple transports (UDP, TCP, shared memory), allowing SustainML nodes to be decoupled yet interoperable in both local and distributed deployments. Fast DDS Python bindings enable partners to implement modules in Python while reusing a C++ communication layer.

D5.2 then states general and WP-specific software requirements. All WP modules must be delivered as Python executable components, communicate over DDS, and comply with shared APIs and lifecycle status reporting. WP-specific requirements define, for example, the Application-level Requirements Node and Model Provider Node in WP1, the Hardware Constraints and Hardware Provider Nodes in WP2, the Carbontracker Node in WP3, the multi-view front-end in WP4, and the Orchestrator Node and communication library in WP5.

The core of the report details the framework architecture: a web front-end where users describe ML problems and constraints, and a back-end of DDS nodes that perform task encoding, model selection, hardware exploration, carbon footprint estimation and result aggregation. Diagrams describe the SustainML pipeline from user input through task encoding, model generation, hardware selection, footprint prediction and iterative feedback.

The software design section presents SOLID-compliant UML diagrams for the Node abstraction, dispatcher, listeners, and sample queues, demonstrating how inputs from multiple topics are coordinated per task ID before invoking node-specific callbacks. The orchestrator node handles overall pipeline control, task tracking and interaction with the front-end. Finally, the deliverable explains versioning, release procedures, demonstrators for module nodes and the full framework, and outlines future work towards full integration of all WPs.

D6.1, under WP6, provides the first validation of SustainML concepts in a real application: automated visual inspection for civil engineering, focusing on bridge maintenance. It demonstrates how foundation models and knowledge-recycling techniques can enhance defect detection and operational workflows.

The deliverable addresses the challenge of inspecting aging critical infrastructure such as bridges and runways. Manual visual inspection is slow, costly and prone to missed defects, and high-quality, fine-grained annotated datasets for structural defects are scarce. To overcome this, the work leverages self-supervised foundation models and data-efficient fine-tuning to support automated defect detection while integrating seamlessly into engineering workflows.

Section 2 reviews foundation models for visual inspection. It emphasises self-supervised learning (SSL) as a way to exploit large unlabeled image corpora collected, for example, by drones. Techniques such as MAE, PIRL, BYOL, MoCo and SimCLR on Vision Transformer-based backbones (ViT, DINO, MSN) are discussed as pre-training strategies that improve downstream performance, particularly when labelled data is scarce.

For bridge defect detection, the team trains a foundation model with a four-stage curriculum. The model is first pre-trained in SSL mode on generic object datasets, then further pre-trained on a large “source” dataset of unlabeled concrete-structure images. A third SSL stage focuses exclusively on the Storebælt Bridge, and a final supervised stage fine-tunes a cascade Mask R-CNN head using around 1,000 annotated images with ~8,000 defects. This hierarchy yields improved recall and cleaner segmentation compared to conventional supervised models, as illustrated in qualitative comparisons on Storebælt pillars.

Section 3 introduces IBM Inspecto, a research–industry platform for enterprise-scale visual inspection. The system supports long-term data management, stitched high-resolution views of structures, 3D models and maps, hierarchical asset navigation, and AI-driven defect detection. Components include an image viewer, prediction viewer with pixel-accurate masks and attributes (e.g., crack length), overview and merged-prediction views, statistical dashboards and a configurable report generator (HTML/PDF).

Inspecto is validated on the Storebælt Bridge, Stenungsund Bridge and Dübendorf Airbase runway. For Storebælt, stitched mosaics and fine-tuned foundation models allow engineers to explore large concrete surfaces, with defects such as cracks, net-cracks, spalling, rust and algae highlighted in distinct colours. For Stenungsund, fine-tuning on ~500 images yields accurate detection of severe spalling and corroded rebars. For Dübendorf, the model can detect small runway cracks, including growth around previously repaired areas, even without additional fine-tuning.

The conclusions highlight IBM Inspecto as a concrete example of how foundation models and SustainML-aligned techniques can digitise inspection, reduce time and cost, and improve safety. Future work includes multimodal foundation-model assistants for large-scale defect review and temporal comparison of stitched surfaces across inspection campaigns.

D7.8, in WP7, documents SustainML’s collaborations during project year 2 with other EU and national projects, research communities and the ROS ecosystem. It shows how SustainML is embedded in the wider trustworthy and sustainable AI landscape and how external synergies amplify project impact.

The deliverable first summarises collaborations with other EU projects and initiatives. It highlights links to Humane-AI-Net, the ALMA project on Algebraic Machine Learning, and several hardware- and embedded-AI-focused projects such as TALON, BioPIM, STRATUM, TIRESYAS and ARCHYTAS. These collaborations support cross-fertilisation on topics like sustainable embedded AI, processing-in-memory architectures, and energy-efficient accelerators that can feed into SustainML’s hardware and model co-design work.

A major part of the report is devoted to the “Trustworthy AI Cluster”, a group of nine Horizon Europe projects funded under the same call (AutoFair, ENEXA, EVENFLOW, REXASI-PRO, SAFEXPLAIN, SustainML, TALON, TUPLES, ULTIMATE). The cluster coordinates communication and dissemination activities to raise the visibility of trustworthy, human-centric AI research across Europe. Joint actions include a shared communication and dissemination plan, cross-promotion on social media, common visual identity (cluster banner) and coordinated participation in events, workshops and webinars.

The deliverable details, for example, the joint webinar “Trustworthy AI: Landscaping verifiable robustness and transparency”, co-organised with ADRA-e and flagship projects AI4Europe and DeployAI. The webinar showcased cluster projects’ contributions to robustness, transparency and verification, with SustainML presenting its perspective on energy-efficient, lifecycle-aware ML as a dimension of trustworthiness. Recordings and slides are made publicly available through the ADRA-e website, extending outreach beyond the immediate research community.

D7.8 also reports on research-oriented collaborations (e.g., contributions at the ASAP conference and joint work with Diehl Aviation), and extensive engagement with the ROS and ROS 2 communities, including ROS Discourse, the ROS 2 Technical Steering Committee and the AI Integration Working Group. These activities aim to bring SustainML concepts closer to robotics practitioners and to influence standards and best practices for AI integration in real-time systems.

Finally, the report lists national-level collaborations of individual partners (e.g., GenoPIM, ODYSSAI and chip-design-efficiency initiatives), which provide additional channels to test SustainML ideas in specialised domains. Overall, D7.8 demonstrates that SustainML is not an isolated effort but part of a rich network of projects jointly advancing trustworthy, efficient AI in Europe.