So what is generative design? Is it modeling, reality capture, scripting, artificial intelligence, or something else?
When it comes to how designs have been traditionally created, there have been a variety of ways:
- Modeling by hand where a human expresses a design to a computer by drawing geometry such as lines, arcs, circles, etc.
- Scripting where a human expresses a set of functions with parameters and computer software produces a model according to those functions and parameters.
- Capturing Reality using scanners or photographs and having computer software convert that data into a model.
Now that generative design is part of design practice, to which of these traditional ways does it belong?
As its name suggests, the idea behind generative design is that a designer specifies requirements, and the computer generates the design. Some examples include:
- For an object, the requirements could be the forces that act on the object. Generative design would then compute its shape based on the forces.
- For a building, a designer could specify which rooms should be adjacent, and generative design would compute the floor plan.
- For a community, a planner could seek to maximize the amount of backyard space, and generative design would layout the lot boundaries.
Actually, generative design does not generate one design but multiple designs that satisfy the requirements and evaluates/scores the designs against those requirements. The designer then selects one based on the scoring or some other factor of his choosing. So the act of designing is to specify requirements, explore the options, and make a selection. The result is the best design instead of the first design that meets the requirements.
So from this definition, how does generative design relate to traditional design methods?
Generative design is not modeling because it is done by computer instead of by hand.
Generative design is much like scripting, but instead of specifying formulas and parameters, the designer specifies requirements.
A large part of generative design is following the same algorithms found in nature. It's often a form of biomimicry. It's like a cousin of reality capture but instead of capturing an object and reproducing it, the process from nature has been captured and is able to be applied.
As a combination of scripting and biomimicry, one might ask, "Is this a form of artificial intelligence?" The answer is "No" and Yes."
No, biomimicry is not a form of artificial intelligence. The computer simply executes a preprogrammed algorithm to convert the input data into the output.
Artificial Intelligence refers to systems that perform tasks that normally require human intelligence. Hence the name artificial. Machine learning is a subset of artificial intelligence that includes systems that improve their performance of a given task with more and more experience or data. The machine is said to learn from the data. Deep learning is a cascading network of multiple layers of machine learning algorithms. Machines attain greater knowledge of their subject matter by diving deeper into the relationships found in the data.
Generative design was first applied to the product design and manufacturing industry for designing parts. It was primarily associated with topology optimization algorithms, but evolutionary found-in-nature algorithms were also used. Generative design has since been applied to other industries such as architecture, engineering, construction (floor planning) or civil engineering (community layout). Given the generic nature of generative design (i.e., a multivariable problem solver), Autodesk's longer-term approach will leverage artificial intelligence. Rather than computing successive generations of designs from scratch, technology will look at past simulations and optimizations and get 95% of the way there via machine learning and only compute the remaining bits as needed. This approach will still use the same scripting-like algorithmic approaches, just only computing "the last mile" via artificial intelligence. This hybrid approach, in theory, should be more cost and time effective. As designers add new goals and constraints such as multiphysics, thermal, fluid flow, etc., the scripting-like computation required goes up exponentially, so reductions based on machine learning can help mitigate those increases.
Yes, generative design based on deep learning would definitely be considered Artificial Intelligence.
Given this, some of the projects that Autodesk Research has underway include:
Project Discover is a workflow for architecture that involves the integration of a rule-based geometric system, a series of measurable goals, and a system for automatically generating, evaluating, and evolving a very large number of design options. The result is a tool to explore a wide design space, and get closer and closer to achieving all of the goals simultaneously.
The Bionic Partition Project was developed in collaboration with Airbus and APWorks to make the world's largest metal 3D printed airplane component. The new Bionic Partition is created through a pioneering combination of generative design, 3D printing, and advanced materials..
Project Dreamcatcher addresses the question: "What if a CAD system could generate thousands of design options that all meet your specified goals?" Dreamcatcher enables designers to craft a definition of their design problem through goals and constraints, and this information is used to synthesize alternative design solutions that meet the objectives. Designers are able to explore trade-offs between many alternative approaches and select design solutions for manufacture.
Design Variations and Optimization leverages advances in design and simulation software and cloud computing to make it possible for designers to offload complex computational tasks to the cloud so they can work productively with their desktop PC, while compute intensive services continue working in the background. One such complex computational task that could be offloaded to the cloud is the generation of multiple design alternatives. Designers are able to input design objectives including specific materials, performance, and cost criteria and let the power of the cloud offer possible variations on the design that meet the design criteria.
Complex Constraint Authoring explores interactive and programmatic means for authoring constraints to control the behavior of systems that must adhere to those constraints.
Generative design has a bright future to make machines a collaborative partner in the design process, and Autodesk Research is working on that. Collaboration is the key ingredient as it's not artificial intelligence to replace designers. In practice, the designer works with the software to arrive at the best design instead of the first design that works.
Design is alive in the lab.