Part 1: Study one talk
Selection of Option A - The technical pipeline: Using Generative Design and Machine Learning for Faster Analysis Feedback (AU 2020)
The talk primarily discusses a key bottleneck for early-stage architectural design; slow, complex simulation workflows result in inefficient evaluation of building performance. Traditional simulation tools are considered both too time-consuming and computationally expensive. Due to the limitations of these traditional simulation tools to keep up with the rapid iteration necessary during early conceptual design stages, this hinders the ability to achieve a wider exploration of the design space. The time to set up and perform the analysis of a partially automated process can be disruptive.
The speakers in the talk propose a workflow that combines generative design with machine learning. Dynamo and Generative Design are used in Revit to generate a large synthetic dataset of the building geometries and corresponding performance metrics (e.g., energy use). A machine learning model is trained to act as the surrogate for simulation; this allows for rapid predictions of the building performance and shifts the workflow from slow, simulation-based feedback to faster data-driven evaluations. This is particularly important for maximizing performance and achieving lower cost during the early-stage design analysis. Examples of model types include linear regression, gaussian process, random forest, and deep learning. In regard to training the model, the training set allows for fitting of the model to data, the validation set allows for both model selection and tuning, and then test set is used for evaluation of the final model.
Advancements in machine learning can enable expedited analysis feedback which provides directional guidance to evaluate design choices and does not necessitate domain-specific expertise for the setup. However, some challenges include that machine learning requires extensive data collection. For instance, the AEC industry’s training data is less developed, there are limited open-source architectural datasets (often limited to 2D images or vector drawings of building floorplans, lack of data for 3D architectural models), as well as a lack of diverse, high quality, and well-formatted data. Still, there are benefits for using an automated workflow with a machine learning model and synthetic dataset generated with generative design features that are available from Revit and Dynamo. The surrogate models can advantageously assist in the conceptual design stage by allowing for instant evaluation and estimates of the building performance. This facilitates a quicker turnaround time for design concept assessment and exploration of the design space. A key takeaway is that having different objectives will produce different optimal geometries since the design objectives will redefine the evaluation criteria that is used by the solver. Furthermore, having target taxonomies allows for a structured, diverse, and strategically explored design space which assists with the management of large datasets to ensure both diversity and accuracy in design exploration.
The general workflow for the creation of a generative design study in Revit involves the establishment of design variables and design objectives which are reduced to inputs (e.g., height, width, etc.) and ranks/outputs (e.g., cost, energy, etc.) respectively. Prior to generating the dataset, an appropriate solver is selected to ensure that the algorithm can evaluate the design space in the context of achieving the specified dataset objective and having identified the nature of the dataset. The algorithmic solver automates the sampling of the input design variables. Following the generation process, design options are created and exported or used to tweak the study further. The process can be repeated iteratively to evaluate additional cases. These steps in the workflow contribute toward defining and guiding the exploration of the design space as well as filtering and optimizing the desired solution that satisfies the design objective. By defining the objectives as what success should be and taxonomies as how the data is structured, using solvers for exploring the design space, and conducting iteration for refinement of designs and datasets, this framework allows for data-driven parametric design and optimization for generative design studies.
One thing in the talk that surprised me was the reliance on synthetic datasets in comparison to real-world data. It is understandable that there is a lack of available architectural datasets which raises challenges for having sufficient real-world data for the design study. However, I am curious about how well the proposed approach will generalize in real design conditions. For instance, if the dataset is not sufficiently diverse or contains hidden biases, this could result in poor models with misleading predictions. I did find it interesting that generative design can be used as a tool to build datasets, using parametric modeling not only for design exploration but also for data creation and the training of machine learning models. Moreover, I was interested to learn about the four different solvers used for generation of building masses in the Generative Design tool (randomize, cross-product, like-this, optimize solvers). By gaining a better understanding of the different solvers available, this can be leveraged in a practical parametric design project using Dynamo by using the appropriate solver for the desired application. For instance, the randomize solver achieves the highest diversity and allows for scattered sampling across a broader design space. The cross-product solver can be used for systematic coverage that explores the parameter combinations in the design space. The like-this solver can be used to focus on clustered solutions for localized evaluation of the variations around a base design. The optimize solver can be used when it is desirable to have converged solutions that cluster around an optimal region. Therefore, in Dynamo, a general recommendation would be to use the randomize solver for early-stage exploration, the cross-product solver for parameter sensitivity studies, the like-this solver to refine the narrowed down design space that is most promising, and then the optimize solver can be used to finalize the optimal, best-performing solution. This sequential use of multiple solvers will allow for a strategic approach that provides broad exploration followed by targeted refinement and optimization.
I would be interested in trying to implement this proposed automated workflow to achieve faster feedback on metrics such as envelope efficiency, which is an area that I previously explored in past assignments with parametric models. However, I would also want to gain greater confidence that the surrogate model in the proposed workflow would be sufficiently accurate to provide informed design decisions. To implement this workflow in a project, I would require a well-validated dataset and have a clear understanding of the limitations of the model. It would also be favorable to have a method to cross-check the results against those produced by simulation. If these criteria and conditions are met, the proposed workflow could provide a significant boost in expediting the early-stage iteration and enhanced decision-making process of generative design studies.
Part 2: Reflect on the quarter
One of the larger friction points experienced this quarter was the management of data and calculations when handling large parametric outputs, especially when aggregating metrics or normalizing results across different design options. While Grasshopper and Dynamo are both powerful tools, there are various limitations; they require manual wiring through multiple steps, troubleshooting data structures issues, and repeatedly recalculating values when something in the flow did not work. Substantial time was also allocated toward tracing the sources of errors in the node logic and checking whether the calculations were meaningful in the context of the design objective. Therefore, extensive effort is dedicated toward handling the data before the design insights can be interpreted. Another challenge was the process of efficiently exploring design tradeoffs in the study. While there were tools like Octopus that helped with this process in Grasshopper, there is still quite extensive manual work needed to interpret the results and identify meaningful patterns. Moreover, another friction point was the repetitive nature of making small changes to inputs, evaluators, or data structures that would require me to revisit parts of the node logic or workflow, rerun the studies, and manually verify and interpret the results. At times, it also took a while for the rerun in the generative design studies to complete and produce the new outputs for review. The current tools were useful for generating design options, but they are quite manual and limited for managing the complexity and extracting the insight from the produced results.
One proposal would be to use an AI-augmented workflow that streamlines this process through the automation of data processing and insight generation. Embedding the intelligent AI assistant in the process could be helpful to automatically detect issues in the node logic. It would also be beneficial to have an AI assistant capable of automatically computing and normalizing performance metrics, flagging errors in the data structures, and determining which evaluators may be the most meaningful based on the design objectives. Furthermore, AI assistants could aid with the interpretation of the tradeoff plots, such as identification of correlations and suggestions of promising regions of the design space that would warrant further exploration. This would also help with highlighting which variables are driving the performance differences through the identified trends and evaluated solutions. At the same time, I feel that encountering some of these frictions throughout the course was valuable, as it drove me to further research on how to understand the underlying data and logic at a deeper level. Through first-hand learning how to better structure the model and node logic, it helped me to remain engaged in troubleshooting various issues encountered while conducting my design studies. I also gained a greater understanding of the relationships between the building geometry, metrics, and optimization process. I feel that AI should assist with repetitive or error-prone tasks that are traditionally manually handled; this would remove the extensive time spent on repetitive debugging and time-intensive processing. However, I would still prefer having control over defining metrics and interpreting the results rather than fully automating this workflow to be an AI only-based decision process.
Part 3: Scout the Frontier
Hypar (AI Platform for space planning and intelligent layout suggestions)
Hypar is a self-contained, cloud-based generative design platform that generates building layouts and systems through the combination of parametric design, automation, and rising AI capabilities. This tool equips users with the capability of creating functions that encode design logic and automate building generation. Hypar maps the natural language input to Hypar parametric functions. It is interesting how the tool not only focuses on modular, reusable workflows, which is similar to what I have learned with Dynamo this quarter, but also offers a more scalable and collaborative approach. Unlike the traditional parametric workflows that use one-off scripts, Hypar allows for design logic to become a shareable building block that can be implemented across multiple projects. This tool could have impacted my workflow this quarter by allowing for the reuse of parametric logic across different design studies rather than having to manually rebuild scripts. For instance, rather than re-wiring Grasshopper or Dynamo definitions repeatedly, I could have reused or adapted the existing functions which would allow me to shift my primary focus toward refining the design intent and performance metrics rather than manually reimplementing the same logic. Hypar would have assisted with automating parts of the design generation process, allowing for a more scalable ecosystem in parametric design.
Veras (EvolveLAB’s AI powered visualization add-in for Revit)
Veras is an AI visualization tool that allows for the generation of photorealistic renderings based on Building Information Modeling (BIM) or modeling environments through prompts. It’s interesting because it lowers the barrier between concept and representation and allows the designer to render the same seed which can aid one’s experimentation with various design iterations or fine-tuning of results for the targeted design objective. For instance, instead of manually modeling the detailed geometry, this tool allows designers to rapidly explore varied design aesthetics and forms. Multiple design variations or lighting schemes can be explored while maintaining the same perspective and structure. There is also an innovative Geometry Slider feature that allows for exploration of many possibilities as well as providing flexibility to fine-tune designs. This could also lead to a pivot from how a traditional workflow uses modeling prior to rendering; Veras could allow for simultaneous design and visualization. If used during this quarter, this tool would have helped with early-stage form exploration by assisting with rapid visualization of design variations without the requirement to fully parameterize or model each individual option. This tool would have also aided me in rapidly generating visual interpretations of varied design options to facilitate the process of comparing and refining my ideas for the design study.
TestFit (Real estate feasibility platform with generative design AI)
TestFit is a generative AI-driven site planning and feasibility platform that allows for the early-state building layout generation process to be automated. It uses inputs such as parking requirements, site geometry, and zoning constraints to then quickly produce buildable site plans in addition to quantitative metrics (e.g., yield, cost). It provides the ability to map the parcels or draw out the custom boundaries for the proposed typology, build the site with a customizable layout, and have 3D visualization of the site plans. This tool also enables real-time iteration capability and comparison of the proposed schemes while simultaneously maintaining connections to downstream tools such as Revit. TestFit is interesting because it addresses the challenge faced with early feasibility design studies. Rather than manually testing the layouts which can be time-intensive, this tool enables the automation of repetitive tasks such as placing the parking, adjusting setbacks, and calculating area yield; this rapid feedback and reduced turnaround time allows for the designers to evaluate the proposed options in a shorter timeframe. Since this tool contextualizes design with financial and regulatory considerations, this can be powerful for processes involving both design and decision-making. TestFit would have changed how I worked this quarter by expediting the early-stage design exploration process and significantly reducing the manual, time-intensive layout adjustments. Rather than rebuilding each option manually, when evaluating different massing or layout strategies, TestFit would have allowed me to generate and compare multiple different scenarios within a relatively quick timeframe; this would allow me to dedicate my focus primarily toward the evaluation of performance and tradeoffs rather than the construction of the models themselves.