Sustainable software by construction

Sustainable-by-construction is my proposal for a model-based engineering approach in which explicit sustainability requirements constrain the design, generation, deployment, and evolution of software systems to generate better software faster. Now with “better” also meaning “as frugal as possible”.

This was my contribution to the upcoming LiveWRITE 2026 event. The structure of the proposal follows the BASE structure described in the Call for One-Page Abstracts and, given the own nature of the event is short but I believe includes some key concepts worth discussing (let me know what you think!).

Bet (Motivation)

Environmental sustainability is an increasing societal and regulatory concern, including for the digital sector. For instance, the European Commission estimates that digital technologies already represent a meaningful share of energy consumption and greenhouse-gas emissions, while the growing adoption of AI is further increasing electricity demand. Among many other initiatives, the EU Energy Efficiency Directive represents a core legal framework aimed at reducing final energy consumption in the European Union.

Existing fields such as Life Cycle Assessment and ecodesign aim at integrating environmental aspects into product and material design and development to minimize adverse environmental impact throughout the product’s life cycle. While originally proposed for physical products (e.g. chemical products), these concepts have also been applied to digital systems, e.g. sustainable computing [12], and more importantly to the actual software running in those systems [5] including its more and more prominent AI components [14].

I believe we need to go beyond current sustainable-by-design discussions and propose Sustainable-by-construction, a model-based engineering approach in which explicit sustainability requirements constrain the design, generation, deployment, and evolution of software systems, while preserving traceability from stakeholder requirements to architectural decisions, implementation elements, and runtime observability.

Indeed, sustainable by construction is an end-to-end process that covers:

• The generation and deployment of a software system that respects the initial set of requirements, including sustainability requirements
• A fully auditable and traceable explanation of the decisions and trade-offs (energy vs accuracy vs latency vs …) that influenced the system and the choice of its specific embedded APIs, LLMs, etc.
• A runtime observability mechanism to make sure the system keeps respecting the requirements and the proper adaptability mechanisms to react when deviations are detected

Approach (Methods)

Putting in place this approach requires innovations in all the involved areas (see Section 4).

To begin with, we need to suggest new domain-specific languages (DSLs) to formalize the sustainability requirements, the sustainability costs (in terms of energy but also, for instance, in terms of water consumption or CO₂ emissions) of different architectures and (AI) components. And a way to define the best match among requirements and potential implementation strategies. This match should not look only at the sustainability requirements but at all of them (both functional and non-functional) and propose different trade-offs and alternatives, also based on an interactive dialogue with the stakeholders.

Then, we need generation and deployment strategies able to transform these requirements into the running software mixing agentic and rule-based techniques. This will also need to make sure the code itself is properly instrumented for observability purposes. Self-adaptation methods for AI-enhanced systems would then be required to leverage this data.

Note also that uncertainty management now becomes a key element of the whole process. There are many uncertain aspects (on the measurements themselves, on the quality of the AI components, on their monetary cost and availability, etc, see also Section

Finally, human aspects are also critical to maximize the adoption of this method.

Sources (Data)

To make informed decisions, this work requires numerous data points. Examples are metrics on the energy cost of running any type of software [9, 10]; the cost of training ML models on different architectures [3, 11]; and the inference cost of different types of tasks on LLMs [1, 8]. We need to extend these approaches to cover, for instance, the energy cost of proprietary models, the cost of agents (that may trigger many requests to LLMs or other agents), the added cost of multilinguism, etc.

We also need to go beyond purely quantitative metrics and collect qualitative ones. For instance, regarding the sensitivity of some of the energy-driven trade-offs, i.e. how much performance, accuracy,… users are willing to sacrifice to meet their own sustainability requirements?

Experience (People)

This proposal requires a mix of backgrounds and expertise, comprising interdisciplinary efforts on sustainable AI [14]; requirements engineering for sustainability (see [4] and the series of RE4SuSy workshops); DSLs, model-driven and low-code approaches [7] to formalize and compare the different specifications and design alternatives and generate the corresponding systems from them; and self-adaptive [15] and models at runtime [6] to monitor and adapt the running systems.

I bring expertise in the software modeling domain, leading the BESSER project [2], an extensible open-source low-code platform that can be used to specify, validate and generate AI-enhanced software systems. I have also worked at the intersection of software modeling and AI, for instance, proposing an extension to ModelCards to express sustainabilty properties [13].

References

1. AI Energy Score. 2025. AI Energy Score: Initiative to Establish Comparable Energy Efficiency Ratings for AI Models. https://huggingface.github.io/AIEnergyScore/. Accessed July 20, 2026.
2. Iván Alfonso, Aaron Conrardy, Armen Sulejmani, Atefeh Nirumand, Fitash Ul Haq, Marcos Gomez-Vazquez, Jean-Sébastien Sottet, and Jordi Cabot. 2024. Building besser: an open-source low-code platform. In International Conference on Business Process Modeling, Development and Support. Springer, 203–212.
3. Lasse F. Wolff Anthony, Benjamin Kanding, and Raghavendra Selvan. 2020. Carbontracker: Tracking and Predicting the Carbon Footprint of Training Deep Learning Models. arXiv:2007.03051 [cs.CY] https://arxiv.org/abs/2007.03051
4. Peter Bambazek, Iris Groher, and Norbert Seyff. 2023. Requirements engineering for sustainable software systems: a systematic mapping study. Requirements Engineering 28, 3 (2023), 481–505.
5. Christoph Becker, Ruzanna Chitchyan, Leticia Duboc, Steve Easterbrook, Birgit Penzenstadler, Norbert Seyff, and Colin C. Venters. 2015. Sustainability Design and Software: The Karlskrona Manifesto. In 2015 IEEE/ACM 37th IEEE International Conference on Software Engineering, Vol. 2. 467–476. doi:10.1109/ICSE.2015.179
6. Nelly Bencomo, Robert B France, Betty HC Cheng, and Uwe Aßmann. 2014. Models@ run. time: foundations, applications, and roadmaps. Vol. 8378. Springer.
7. Jordi Cabot. 2024. The Low-Code Handbook: Learn How to Unlock Faster and Better Software Development with Low-Code Solutions. Independent. https://www.amazon.com/dp/B0DK2ZPTQX
8. Joel Castaño, Jaime Bustillo, Xavier Franch, and Silverio Martínez-Fernández. 2026. Data-driven multi-objective optimization of ML inference hardware configurations for energy, performance and cost. Sustainable Computing: Informatics and Systems 50 (2026), 101347. doi:10.1016/j.suscom.2026.101347
9. Benoit Courty, Victor Schmidt, Sasha Luccioni, Goyal-Kamal, MarionCoutarel, Boris Feld, Jérémy Lecourt, LiamConnell, Amine Saboni, Inimaz, supatomic, Mathilde Léval, Luis Blanche, Alexis Cruveiller, ouminasara, Franklin Zhao, Aditya Joshi, Alexis Bogroff, Hugues de Lavoreille, Niko Laskaris, Edoardo Abati, Douglas Blank, Ziyao Wang, Armin Catovic, Marc Alencon, Michał Stęchły, Christian Bauer, Lucas Otávio N. de Araújo, JPW, and MinervaBooks. 2024. mlco2/codecarbon: v2.4.1. doi:10.5281/zenodo.11171501
10. Ornela Danushi, Stefano Forti, and Jacopo Soldani. 2025. Carbon-efficient software design and development: a systematic literature review. ACM computing surveys 57, 10 (2025), 1–35.
11. Santiago del Rey, Luís Cruz, Xavier Franch, and Silverio Martínez-Fernández. 2027. Estimating Deep Learning energy consumption based on model architecture and training environment. Computer Standards & Interfaces 99 (2027), 104170. doi:10.1016/j.csi.2026.104170
12. Udit Gupta, Mariam Elgamal, Gage Hills, Gu-Yeon Wei, Hsien-Hsin S Lee, David Brooks, and Carole-Jean Wu. 2022. ACT: designing sustainable computer systems with an architectural carbon modeling tool. In Proceedings of the 49th Annual International Symposium on Computer Architecture. 784–799.
13. Gwendal Jouneaux and Jordi Cabot. 2025. Towards Sustainability Model Cards. In Proceedings of the 2nd Workshop on Green-Aware Artificial Intelligence (Green-Aware AI@ECAI 2025) co-located with the 28th European Conference on Artificial Intelligence (ECAI 2025), Bologna, Italy, 25-30 October 2025 (CEUR Workshop Proceedings, Vol. 4165), Riccardo Cantini, Luca Ferragina, Davide Mario Longo, Anastasija Nikiforova, Simona Nisticò, Reza Shahbazian, Dipanwita Thakur, Irina Trubitsyna, and Giovanna Varricchio (Eds.). CEUR-WS.org, 20–28. https://ceur-ws.org/Vol-4165/short10.pdf
14. Raghavendra Selvan. 2025. Sustainable AI. O’Reilly Media, Inc. https://www.oreilly.com/library/view/sustainable-ai/9781098155506/
15. Meghana Tedla, Shubham Kulkarni, and Karthik Vaidhyanathan. 2024. Ecomls: A self-adaptation approach for architecting green ml-enabled systems. In 2024 IEEE 21st International Conference on Software Architecture Companion (ICSA-C). IEEE, 230–237.