Software Engineering for AI-driven Building Operation
This paper identifies the unique software engineering challenges in deploying AI-driven building operations—where failures have irreversible physical consequences rather than just digital ones—and proposes a new foundation and best practices to bridge the gap between existing AI software practices and the safety-critical nature of cyber-physical building systems.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Buildings are among the largest consumers of energy on the planet, responsible for a significant share of global greenhouse gas emissions. For decades, the systems that manage heating, cooling, and lighting have relied on fixed, rule-based instructions. These systems are set up once and rarely changed, struggling to adapt to shifting weather patterns, changing numbers of people inside, or the natural wear and tear of the machinery itself. In recent years, engineers have looked to artificial intelligence to solve this, hoping that smart algorithms could learn from data to run buildings more efficiently. However, a critical gap has emerged between the promise of these smart systems and the reality of putting them to work. While artificial intelligence thrives in the digital world, where a mistake can be undone with a simple click, buildings exist in the physical world. There, a bad decision does not just delete a file; it wastes energy that cannot be recovered, makes people uncomfortable, or damages expensive equipment.
A team of researchers from Austria, working across civil engineering and computer science, has set out to bridge this gap. They argue that the standard methods used to build and test artificial intelligence software are insufficient for the physical world of building operations. Their work, drawn from two ongoing projects involving real university buildings, identifies why current approaches fail and proposes a new way forward. They found that treating a building like a standard computer program ignores the slow, messy, and irreversible nature of physical systems. Instead of relying on the massive amounts of clean data that digital systems enjoy, they suggest designing systems that can function with limited, noisy information. They also emphasize that because building managers are legally responsible for safety and comfort, they must be able to understand and override the artificial intelligence, making transparency a requirement rather than an optional feature.
The researchers began by observing that the standard rules for software engineering do not apply when the software controls a physical environment. In the digital realm, if a program makes a wrong prediction, the consequences are often limited to a poor user experience or a lost sale, and the error can be fixed quickly. In a building, a faulty control decision has immediate physical costs. If an artificial intelligence system mistakenly turns on both heating and cooling at the same time, that energy is wasted forever. If it causes the temperature to swing wildly, occupants lose trust in the system, and the building manager may simply turn it off, rendering the technology useless. Even worse, aggressive changes to equipment settings can cause mechanical stress that shortens the lifespan of the machinery. While true disasters like fires are rare because buildings have built-in safety buffers, the cumulative effect of small, irreversible errors changes everything about how the software must be built.
To address these unique challenges, the researchers identified several key problems that arise when moving from theory to practice. One major issue is the nature of the data. In many digital applications, data is abundant and clean. In buildings, data is often sparse, noisy, and scattered across different systems that do not talk to each other. A single building might have heating systems from the 1990s, lighting controls from 2010, and modern sensors, all using different languages and formats. Collecting this data is difficult, and sensors often drift out of calibration over time. The researchers found that trying to build complex models that require perfect data leads to fragile systems that fail in the real world. Instead, they advocate for a principle of data minimalism. This means designing systems that can work with the limited, imperfect data that actually exists, rather than waiting for perfect conditions that may never arrive.
Another critical finding is that testing these systems directly on real buildings is too risky. Unlike digital software, where developers can run two versions side-by-side to see which works better, you cannot test two conflicting heating strategies in the same building without risking discomfort for the people inside. The researchers observed that relying solely on computer simulations is also insufficient because real buildings never behave exactly like their digital models. Thermal mass, weather, and human behavior create delays and variations that are hard to predict. To solve this, they propose a "simulation-first" approach where new strategies are tested extensively in a virtual replica of the building before they are ever deployed. This virtual replica, known as a digital twin, is not just a static model but a living copy that updates with real-time data from the actual building. It allows the artificial intelligence to explore different actions and learn from mistakes without causing any real-world harm.
The human element also plays a central role in their findings. In digital applications, if a user does not trust a recommendation, they simply ignore it. In a building, if a facility manager does not trust the system, they will disable it entirely. Because these managers are legally responsible for the safety and comfort of the occupants, they need to understand why the system is making a decision. The researchers argue that explainability is not just a nice-to-have feature but a fundamental requirement for adoption. If an artificial intelligence cannot explain its reasoning in plain language, or if it cannot be manually overridden when necessary, it will not be used. This means the software must be designed to be transparent from the start, providing clear reasons for its actions and allowing humans to step in when needed.
The team suggests that the path forward involves a shift in how engineers approach these problems. They propose that the digital twin should be the central tool for development, serving as a safe space for testing and a way to generate data for rare situations. They also emphasize that the workflow from design to deployment needs to be integrated, connecting the building's physical model with the artificial intelligence and the control systems. This approach acknowledges that buildings are complex physical systems governed by thermodynamics, not just data patterns. The researchers are careful to note that their work is a starting point for a new research agenda rather than a finished solution. They have identified the missing pieces and proposed a set of practices, but they acknowledge that these ideas need to be validated in the field. The goal is to create a foundation for building artificial intelligence that is robust, safe, and trusted by the people who must live and work with it.
Ultimately, the work highlights that the challenge of bringing artificial intelligence into buildings is not just a technical problem of better algorithms, but a software engineering problem of managing physical consequences. The researchers found that success depends on respecting the laws of physics, the limitations of available data, and the need for human trust. By treating the building as a physical system first and a data source second, they offer a blueprint for making smart buildings that actually work in the real world. Their findings suggest that the future of building operations lies not in replacing human judgment with machines, but in creating systems that support human oversight while navigating the complex, slow, and irreversible nature of the physical world.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.