Transforming Concepts into Infinite Possibilities with deepSPACE Design Tool

According to Jordan Smart, “We aimed to replicate what large AI language models have achieved with text for engineering and design. Currently, when you launch engineering design software, you’re met with a blank screen. With deepSPACE, you outline your needs, and it produces between 100 and 1,000 feasible concepts in the time it would take a person to examine just one or two, providing a comprehensive view of the entire design landscape.”

Furthermore, Smart assures that deepSPACE is not restricted to physics questions alone. “It’s trained on a combination of historical data and simulations, but it can also utilize standard cost estimation tools to provide at least a basic level of feedback for cost analysis.”

To showcase its versatility, Smart and his research colleague Emilio Botero employed deepSPACE to create designs for physical systems involving beams, wheels, and aircraft, as well as for operational logistics networks. They formed partnerships with major aerospace and automotive companies to ensure that deepSPACE is genuinely beneficial for both researchers and industry experts.

Smart noted, “While users might desire deepSPACE to come preloaded with features, companies prefer to establish custom models linked to their own data and insights. We can develop our own models for research or design, but the platform can also function starting without any data. It’s adaptable.”

Smart claims that deepSPACE outperforms traditional optimization algorithms in efficiency. “Where others reported needing 20,000 simulations to begin analyzing their design space, we managed to achieve similar results using only about 250 samples. That means with roughly 100 times fewer data points, you can grasp the design space’s trade-offs.”

When it comes to designing an aircraft, understanding how altering the wings, adding an engine, or increasing the payload affects the overall design can get quite intricate. Conventional methods might require thousands of design points to interpolate effectively between them. Thanks to its ability to create a complete generative model, deepSPACE can make accurate interpolations using significantly fewer data points. We can achieve the same predictions with equal accuracy but in a quicker and more cost-effective manner.

The cost-effectiveness of deepSPACE renders it particularly advantageous for aerospace industries. “We depend on simulations because constructing aircraft is costly. However, we are exploring its applications in other sectors too.”

Additionally, deepSPACE’s generation of a 3D CAD file is a significant feature. Smart mentions that the outputs from many other image generation tools cannot be seamlessly integrated into other design software while preserving all layers and effects.

With deepSPACE, you receive a raw file that is identical to what a human engineer would produce, allowing for straightforward edits or modifications. It integrates effortlessly into your workflow, as if the work were outsourced to another company, making it one of their deliverables.

According to Smart, deepSPACE fosters a unique design dialogue with the engineers who train it.

He recalled one instance when deepSPACE generated a design that appeared absurd at first. “We initially thought it didn’t make sense. It was designed according to specific requirements, yet it was unlike anything in our training data. Upon examining the outcomes, we found that the actual simulation results were logical and met the requirements.”

The aircraft in question featured shorter wings with control surfaces positioned rearward for balance and stability. Smart mentioned that it wasn’t unrealistic or unbuildable. As they investigated further, they realized it resembled an existing airplane designed by a top aircraft manufacturer.

He explained, “I was responsible for setting up the training data, simulations, and learning algorithms. We provided deepSPACE with training data from three conventional tube and wing aircraft, the Concorde, and one blended wing body concept. From that, it began to create its own concepts while validating them against simulations and learning from the results. Occasionally, it would produce non-physical results, but that helped it to understand the limits.”

DeepSPACE conducted its own experiments, akin to brainstorming, and discovered unexpected results. “My inclination would have been to discard it,” Smart admitted.

However, deepSPACE demonstrated to him how the simulation results aligned with the set requirements, successfully identifying a viable solution to the design challenge.

“We supplied it with a structured dataset of historical data, which enhanced its understanding, enabling it to explore and experiment. While I can create a base model to obtain results, I can also treat it like a creative environment or sandbox. This allows me to run new simulations that aren’t included in the historical data and observe how that expands my knowledge base.”

Over the years, Smart felt that although they possessed remarkable analytical skills, they became the bottleneck in the process. “We have simulations, but a human can’t consistently run thousands of them to ignore the subpar results and identify the best outcomes while cultivating intuition. deepSPACE marks the first generation of systems designed to function as an assistant engineer. You can establish the problem and, upon returning, find a variety of solutions, paving the way for deeper insights based on your existing capabilities.”

Although the primary design was tailored for professional academics and industry experts, Smart envisions broader uses.

“I aspire to have middle school students utilize something like deepSPACE. They might not grasp the physics or possess all the skills necessary for advanced CAD design, but if they have an idea for a car, train, spaceship, or something else, they can communicate it to deepSPACE and test it out. They can then modify it and observe the subsequent outcomes.”