Peisen Zhao

1. How does defining the design objectives influence the outcomes in a parametric design study using Generative Design in Revit?

Theoretically speaking, defining design objectives in a parametric study directly shapes the range and nature of the generated outcomes. In my previous design assignment, the objectives—Product Quality Score, Sustainability Score, and Construction Time—each favor different geometric configurations with diverse foucs to different interests for a construction project. For example, maximizing Product Quality Score encourages taller geometries with an optimal base-to-height ratio, while minimizing material usage (for sustainability) may favor more compact forms. If the objectives were instead cost minimization, the Generative Design Study algorithm would produce very different options. These varying objectives would instantly steer the design alternatives toward distinct directions for different foucs and interests while finding an optimal solution, proving that well-defined goals are crucial for meaningful generative outcomes.

2. Discuss the importance of identifying target taxonomies when generating synthetic datasets in architectural design.

Identifying target taxonomies, where clear categories are defined such as geometry type, building function, construction cost, or sustainability level, helps maintain structure and clarity when generating large synthetic datasets. In my study, varying product geometries according to height and base radius generates a dataset of different configurations. By constraining these outputs with performance categories (e.g., “high sustainability” or “low construction time”), I can better manage the dataset, filter for targeted outcomes, and ensure that diverse scenarios are represented. That is why it is essential not only just for design exploration but also for training machine learning models or supporting data-driven decision-making in AEC industry as a whole.

3. What are the potential benefits and challenges of using automated workflows for generating diverse synthetic datasets in parametric design?

Generally, automated workflows would bring scalability, consistency, and speed to the generation of design options compared to traditional workflow to generate design options, which is more time-consuming and labor-intensive. Specifically, in the context of my geometry-based building design from last week, automation allows the generation of multiples of variations with slight changes in radius and height, each evaluated for construction quality, sustainability, and construction time.

From this perspetive based on my design, the benefits of this automation include rapid iteration, objective comparison. Additionally, the system is modular where each evaluator is a separate function to enable easy substitution (e.g., switching from volume-based sustainability to carbon-based metrics). Therefore, scalability is achieved by to handle even more compliex design requirements by adjusting or adding new input parameters.

However, challenges arise in validating results, managing combinatorial outcomes (since too many options), and ensuring that generated options are both meaningful and constructible. This is because the model may create extreme geometries that are theoretically optimal but practically unfeasible. These must be filtered or constrained using architectural judgment or additional evaluators or human evaluations. Additionally, another issue is computational bottlenecks where geometry generation and evaluation can be resource-intensive and time-consuming as complexity grows.

4. Explore the role of iterative processes in optimizing design options within the generative design framework.

Generally, iterative processes are essential in refining generative outputs to better align with project goals. In my design, the initial study was trying to reveal configurations that rank high in one metric (like quality) but perform poorly in others (like construction time). By tweaking the input ranges and slightly adjusting the formulas of objectives, I can guide the next iteration toward better tradeoffs with a more optimal solution. This loop of generation, evaluation, and refinement ensures that the design evolves meaningfully rather than simply generate outcomes that are constructional infeasible. That is being said, without iteration, the process risks producing suboptimal or unbalanced solutions that don’t fully satisfy real-world constraints or user priorities.

5. Describe the general workflow of creating a generative design study in Revit. What are the key steps involved, and how do they contribute to the generation of optimized design options?

The workflow for a generative design study in Revit, as applied in design project last week using Generative Design, involves six structured phases:

1. Define the Objective: the objective was to optimize a building mass for Construction Time, Product Quality Score, and Sustainability Score. These were translated into measurable output metrics in Dynamo.
2. Set Design Variables: I parametrized key aspects for input such as the height of the building as well as the base radius to be flexibly adjusted in Generative Design Study. Other parameters could be involved as well: such as the number of volumes, the relative heights with surrounding builidings, footprint shapes, and spatial arrangement.
3. Establish Goals and Constraints: Using formulas defined, I am enabled to relate geometry to three different evaluators. For instance:

Construction Time – calculated as Construction Time = 90 + (Volume * 0.02) + (Height * 1.5)

Product Quality Score – calculated as Product Quality Score = 100 × (base_area / height)

Sustainability Score – calculated as Sustainability Score = 10000 / (base_area × height)

Based on these formulas, goals and constraints are defined with a purpose either to maximize or minimize each evaluator, and corresponded range of constraints.

4. Use Solvers: I started with the Randomize Solver to broadly explore combinations, then transitioned to Optimize Solver for refining promising candidates since optimization is what I ultimately want for my design.
5. Evaluate Results: Each design iteration would be analyzed with Parallel Coordinates Graph where I am enabled to select ranges of each evaluator to ensure I could dynamically adjust and select the best combinations.
6. Refine and Repeat: Based on result clustering, I am able to identify trade-offs and insights to report and may rerun the study with adjusted parameters based on analysis.

6. Given the properties of a good synthetic dataset—large, labeled, parametric, expandable, diverse, and balanced—how would you apply these principles to create a dataset for a building morphology study using Dynamo?

To create a high-quality synthetic dataset for our building morphology study, it would be necessary to follow a structured taxonomy and parameter-driven approach:

• Taxonomy Definition: it is essential to classify masses such as linear blocks, stacked volumes, courtyard forms, etc. Each type was tied to a Dynamo subgraph, promoting parametric modularity.
• Parametric Input: for each type, parameters like Building Height, Base Radius, or Footprint Area are controlled via sliders. These would form a multi-dimensional input space, denoted: X={x1,x2,…,xn}, where each xi is a design variable
• Labeled Outputs: evaluation metrics were computed using consistent formulas. For example, the Construction Time label was derived from: Construction Time = 90 + (Volume * 0.02) + (Height * 1.5)
• Automated Expansion: it would be good to use the Cross Product Solver to generate comprehensive permutations of inputs, ensuring dataset diversity and feasible consistency.
• Balancing: samples will be grouped by morphology type and normalized using z-scores for each label to prevent bias in downstream ML models.

7. Identify and discuss the four different solvers mentioned in the lecture that can be used to generate building masses in the Generative Design tool. How do these solvers impact the sample space and variety of design options produced?

The four solvers, Randomize, Cross Product, Like-this, and Optimize, each would play distinct roles and functionalities in shaping the generative design output space:

• Randomize:
• Samples randomly from the parameter space
• Encourages variety but lacks convergence
• Useful in initial stages when exploring morphologies with undefined constraints
• Cross Product:
• Computes the full Cartesian product of parameter values
• Best for systematic sampling and ensuring maximum design diversity
• Like-this:
• Conducts local sampling around a “base” design where one design option would be similar to its neighbor design options
• Ideal for sensitivity analysis or design refinement stages
• Optimize:
• Uses genetic algorithms or other heuristics to converge toward best performers
• Useful when targeting optimal solutions involving trade-offs, such as the design variables of my design last week: Construction Time, Product Quality and Sustainability.

8. Reflect on the examples of building masses generated with different solvers from the class handout. What insights can you gain about the relationship between solver choice and design diversity? How would you leverage this understanding in a practical parametric design project using Dynamo?

The building mass examples from the class demonstrate how solvers shape the different design approaches and emphasises:

• Randomize would deliver scattered but diverse outputs. In our study, it is able to reveal outlier designs with exceptional Product Quality Scores but poor sustainability due to irregular geometry.
• Cross Product would map the entire design space and help identify clusters of solutions with balanced trade-offs. This was crucial in understanding the pareto front for example between Construction Time and Sustainability Score.
• Like-this would be effective in narrowing configurations where a single tweak could improve all metrics slightly—a design local optimum.
• Optimize would reveal solutions with high performance but unconventional forms, valuable for performance-centric projects and comparing different tradeoffs.

In a real-world Dynamo workflow, I would sequence these solvers as below:

1. Begin with Randomize to brainstorm diverse ideas,
2. Use Cross Product for dataset expansion and ML training,
3. Apply Like-this for iterative tuning with a specific focus of local optimal solutions,
4. Finish with Optimize to select high-performance solutions.