Looking Ahead - Machine Learning and AI
Module 8 Questions:
Part 1 – Study One Talk
I chose Using Generative Design and Machine Learning for Faster Analysis Feedback. The main problem discussed in the talk is that building performance analysis can take a long time, especially when many design alternatives need to be evaluated. In the early stages of a project, designers often want quick feedback on the impact of different design decisions, but running detailed simulations for every option can become impractical. The approach presented in the talk is to generate many building variations, simulate them, and then use the results to train a machine learning model. Once trained, the model can estimate performance metrics for new designs much faster than running a full simulation every time.
The thing that surprised me most was the idea of using generative design mainly to create training data rather than to find an optimal solution. Before watching the talk, I mostly thought of generative design as a way of exploring alternatives or performing optimisation. Here, it was being used to generate enough examples for a machine learning model to learn from. I found that interesting because the focus shifts from finding a good design to building a dataset that can later be used to evaluate designs much more quickly.
One thing I would question is how well this approach translates to real projects. Building performance depends on much more than geometry. Factors such as occupancy patterns, façade design, HVAC systems, control strategies, and site conditions can all have a significant impact on the final result. Because of that, I think these models are probably more useful for identifying trends and comparing options than for predicting actual building performance with a high degree of accuracy.
I would be interested in trying a similar approach for energy, daylight, solar radiation, and HVAC studies. One thing I noticed during this quarter is that it is relatively easy to create many design alternatives, but obtaining useful performance information from them takes much longer. A machine learning model that could provide reasonably accurate feedback in seconds would make exploring different ideas much more practical. Rather than replacing simulation, I see it as a way to reduce the search space by quickly filtering many alternatives down to a smaller set of promising candidates. Detailed simulations and optimisation could then be focused on those candidates, making the overall workflow much more efficient.
Part 2 – Reflect on the Quarter
One of the main challenges I encountered during the quarter was not creating parametric models but evaluating and comparing the results. Generating alternatives was usually straightforward once the workflow was set up. The difficult part was deciding how to compare solutions when several objectives were involved. In the optimisation assignments, I had to work with metrics that had different units and scales, which required determining appropriate minimum and maximum values and normalising the results before they could be compared. I also found it challenging to understand how different optimisation approaches handled competing objectives. In a brute-force workflow, every alternative could be evaluated and ranked directly once a scoring method was defined. Multi-objective optimisation was less intuitive because it produced a set of promising solutions rather than a single best answer, interpreting trade-offs a significant part of the process.
An AI- or ML-augmented workflow could help with the evaluation stage rather than simply generating more alternatives. For example, it could suggest ways to normalise metrics, explain trade-offs between objectives, and help interpret optimisation results. It could also provide guidance on why certain solutions are worth investigating further, instead of leaving all the interpretation to the user. I would want that type of augmentation because it would allow me to spend more time thinking about the design implications of the results rather than processing and comparing data. At the same time, some of that friction was valuable because it forced me to think about what “better” actually means when several competing objectives are involved.
Part 3 – Scout the frontier
Tacit
Tacit is an AI platform focused on building operations and performance. It connects to existing building systems and helps facility teams identify inefficiencies, operational issues, and opportunities for energy savings by analysing data from multiple sources. What I find interesting is that it focuses on the operational phase of a building rather than the design phase. Many discussions about AI in AEC focus on generating designs, while Tacit focuses on understanding how buildings perform once they are occupied. This would not have directly changed my work during the quarter, but it is highly relevant to my professional interest in HVAC systems, controls, and building performance.
Raven
Raven is an AI assistant designed specifically for Grasshopper. It can help users understand components, troubleshoot definitions, and build workflows through natural language interaction. What makes it interesting is that it addresses one of the biggest barriers to computational design: the learning curve. During this quarter, I spent a significant amount of time learning new tools and optimisation workflows, often having to understand both the design problem and the software at the same time. A tool like Raven could have reduced that friction by providing guidance directly within the modelling environment and helping me focus more on the design questions themselves.
Augmenta
Augmenta is an AI-driven design platform focused on engineering systems, particularly MEP design. Instead of generating architectural concepts, it automates tasks such as system layout and design development based on engineering requirements and project constraints. I find this particularly interesting because engineering design often contains a large amount of repetitive work that follows well-defined rules. Unlike many AI tools that focus on renderings or floor plans, Augmenta applies automation to the technical aspects of building design. While it would not have directly changed the assignments in this course, it represents a possible future where engineers spend less time on repetitive modelling tasks and more time evaluating design decisions and project requirements.