In partnership with

Receive Honest News Today

Join over 4 million Americans who start their day with 1440 – your daily digest for unbiased, fact-centric news. From politics to sports, we cover it all by analyzing over 100 sources. Our concise, 5-minute read lands in your inbox each morning at no cost. Experience news without the noise; let 1440 help you make up your own mind. Sign up now and invite your friends and family to be part of the informed.

Sign up today!

AI for Generating Reasonable, Not "Right" Outputs

One of the fundamental challenges in architectural design and engineering is that many problems do not have a single "correct" solution—only a range of reasonable solutions, each balancing different constraints, aesthetics, and functional goals. Unlike traditional computational methods that enforce strict optimization toward a singular outcome, AI systems can be designed to produce diverse, contextually appropriate outputs that allow for exploration and flexibility.

To achieve this, AI models must be designed to:

1. Learn to generate reasonable solutions rather than correct answers.

2. Co-optimize between competing requirements rather than maximizing a single objective.

3. Introduce controlled randomness to produce varying outputs from the same inputs.

Let’s explore how these principles can be implemented using Generative Adversarial Networks (GANs), multi-objective optimization, and stochastic modeling.

Using GANs to Learn Reasonable Design Patterns

Traditional rule-based computation requires explicitly defined rules for generating outputs. However, GANs (Generative Adversarial Networks) offer a more flexible approach—learning patterns from existing data and using them to generate new, yet realistic, outputs.

How GANs Work in Design Contexts:

• GANs consist of two neural networks—the Generator, which creates new solutions, and the Discriminator, which evaluates whether those solutions are plausible based on training data.

• Through continuous feedback, the Generator improves its ability to produce realistic outputs that resemble successful past examples.

• Over time, the system learns the underlying structure of good design without being explicitly programmed with rigid rules.

Example: GANs for Generative Façade Design

• A GAN trained on thousands of existing building façades can generate new façade concepts that maintain proportion, rhythm, and material logic.

• The system does not copy existing designs but learns reasonable variations based on real-world patterns.

• Instead of producing one "perfect" façade, the model can generate multiple options, allowing architects to explore different aesthetic directions.

Key Takeaway: GANs enable AI systems to learn the essence of good design rather than forcing predefined outputs.

Co-Optimizing Between Competing Requirements

Many design problems involve conflicting priorities—such as maximizing daylight while minimizing heat gain, or increasing density while preserving open space. A traditional computational approach might optimize for one metric at a time, leading to solutions that neglect other factors.

How AI Can Handle Trade-Offs:

• AI models can be built with multi-objective optimization frameworks, allowing them to balance multiple criteria at once.

• Instead of searching for a single best solution, the AI can generate a set of viable solutions, each representing different trade-offs.

• These solutions can be ranked probabilistically, giving designers insight into which options best align with project goals.

Example: Multi-Objective Optimization in Urban Planning

• An AI system designing a mixed-use development must balance:

  • Building height and density for financial viability.

  • Public green space for livability.

  • Circulation and access for efficiency.

• Instead of providing a single "optimal" master plan, the AI generates several options, each prioritizing different trade-offs.

• Planners can then evaluate the different scenarios, rather than being forced into a single predefined approach.

Key Takeaway: AI can simultaneously optimize for multiple goals, producing a range of solutions rather than one fixed answer.

Introducing Randomness to Enable Design Variation

In many design workflows, variation and iteration are key. However, traditional computation often produces one output per input, meaning that a given set of constraints always leads to the same result. AI can introduce controlled randomness, allowing for variation in outputs while still maintaining feasibility.

How AI Uses Randomness in Design Generation:

• AI models can incorporate stochastic elements—meaning that each time the system generates a result, small variations are introduced.

• This allows one set of inputs to produce multiple valid outputs, encouraging design exploration.

• Instead of requiring users to manually tweak inputs, AI can automatically generate diverse alternatives.

Key Takeaway: AI’s ability to introduce randomness and variation allows for design exploration, rather than funneling users into one rigid answer.

A Creative Partner

In the realm of architectural design and engineering, AI’s greatest strength lies in its ability to generate a spectrum of reasonable solutions rather than chasing a single "right" answer. By leveraging techniques such as GANs, multi-objective optimization, and controlled randomness, AI can support creative exploration, balance competing priorities, and enhance iterative design workflows. Instead of replacing human judgment, these systems serve as collaborative tools—offering designers a diverse set of possibilities that can be refined and adapted to project needs. As we continue to develop and refine AI models, the focus should not be on dictating outcomes but on expanding the range of viable solutions, empowering designers to navigate complexity with greater flexibility and insight.

Thanks for tuning in for this weeks Not Magic, Just Math - if you made it this far and are not subscribed, please consider doing so! It means so much to me when I see someone new has joined the community!

Till next week!