What if predicting building energy demand did not require long simulation runs, heavy workflows, or massive datasets?
Xinyue Wang recently defended her PhD, “Leveraging Machine Learning to Improve Early-stage Building Energy Optimization.”
Originally trained as an architect, her research sits at the intersection of building energy simulation and machine learning. Her work focuses on making data-driven methods usable in early design stages, prioritising simplicity, interpretability, and practical relevance over complexity.
Below, Xinyue reflects on her research, her design perspective, and what she hopes others will take from her work.
You began in architecture and now work at the intersection of building energy simulation and machine learning. How do you usually describe what you do to people?
I normally say, ‘I used machine learning to predict the energy demand based on how the building looks, so I can help architects to design more energy-efficient buildings.’
Do you feel like your architectural training shaped the kinds of research questions you care about, compared to someone who might come from a purely technical background?
Definitely yes! As an architect (or architecture student), I understand that architectural problems cannot be approached from a single technical metric alone, like energy consumption. I also recognize that architects cannot simply give up considerations of design intent and aesthetics. Therefore, I am more inclined to approach energy reduction from an architect’s perspective, by designing tools that architects can actually use—tools that help reduce building energy consumption without compromising much design quality.
A central contribution of your work is identifying which architectural design variables actually matter. Why is this step more important than simply building more powerful prediction models?
Identifying which architectural design variables actually matter is the very first step in building my machine learning models, and it forms the foundation of all subsequent research. Moreover, this step involves the most interaction with stakeholders throughout my entire research process, which is why I devoted a significant amount of time and effort to it.
You deliberately chose to work with simpler, more interpretable machine learning models. What does that choice say about your philosophy toward tools in design and research?
From the outset of defining my research scope, I made a conscious decision to prioritize model simplicity and interpretability over maximizing predictive accuracy. This preference reflects the practical constraints of architectural research, as lighter-weight models can be developed with smaller datasets and require significantly less computational and training time. Moreover, my research focuses on the early stages of architectural design, where time efficiency is more critical than marginal gains in accuracy. Therefore, simpler and more interpretable machine learning models are more appropriate for this context.
More broadly, in machine learning research, I do not believe it is always necessary to pursue increasingly complex or sophisticated models. If a simpler model is sufficient to accomplish the task, there is nothing inherently inferior about that choice.
What did you learn about the relationship between model accuracy and usefulness in practice through your experiments?
My research results show that in early-stage architectural applications, accuracy does not need to be overly emphasized. Part of the reasons is that there is inherently a high level of uncertainty in the early phases of architectural design, and at this stage, time efficiency is far more important than accuracy.
Your work with synthetic datasets highlights both opportunities and limitations. What should researchers and designers be most careful about when using synthetic data to inform decisions?
I believe the most important thing is to clearly understand the limitations of synthetic datasets when applying them, and to combine the results derived from synthetic data with real-world conditions in practice.
Given the context of your work, how do you see the relationship between machine-generated outputs and an architect’s judgement evolving?
I believe the two will never be fully aligned. In the context of architectural design, subjective aesthetic judgment by the architect will always play a dominant role. Machine-generated outputs are more likely to function as intermediate results or benchmarks, providing guidance rather than replacing the architect’s decision-making.
How do you hope your research might influence the way data science is taught to architecture and engineering students?
I hope it can help students realize that data-driven approaches are not something overly grand, highly technological, or difficult to achieve. It is not necessary to rely on extremely complex algorithms or massive datasets to build meaningful models and apply them within the field of architecture.
What kinds of questions, projects, or collaborations are you most curious to explore next?
There are many! To give a few examples, I am interested in exploring how different models can be combined with optimization algorithms to examine what level of accuracy is sufficient in practice. I would also like to apply my research to renovation projects or urban planning contexts. Additionally, I am curious about shifting the objective from energy performance to other goals, such as life cycle assessment (LCA) or daylighting analysis.
Outside of research, what do you enjoy spending time on when you are not thinking about energy models and machine learning?
Climbing. But mainly just yap with my friends in the climbing gym instead of actually climbing.
Finally, if someone reads your thesis but only takes away one idea, what do you hope that idea is?
That meaningful models for early-stage applications can be built using a relatively small amount of data and simple machine learning algorithms.
If you are interested in collaboration, applications in renovation or urban contexts, or data-driven methods in architectural design, feel free to reach out.
Once again, congratulations to Dr. Xinyue Wang on this important milestone.
