AI Enables Data-Driven Product Design

Building a Digital Expert

While an AI copilot can help answer questions or simple automation, this does not address the core issue of complexity. A new user will not know what questions to ask or have a workflow to automate, and even for an expert user, navigating a User Interface (UI) with dozens of menus and thousands of commands is a daunting task. By embedding AI directly into design tools, it can not only learn from expert users as they use the software – capturing valuable domain expertise that is often lost due to retirement or transfer but also use that knowledge to simplify the use of the tool.

By understanding expert workflows, best practices, and the context of the part or product being worked on, an integrated AI system can intelligently put the following tool the user might need just one click away, even if a new user might not know what tool they need next. This process can be extended further, allowing the AI to recommend tools and design practices. It could offer these suggestions conversationally when integrated with a copilot system, as an experienced colleague might. An AI system with a deep understanding of best practices and expert workflows could suggest everything from the best spot to place components on a PCB to the optimal way to reinforce a wing, all learned from a company’s proprietary design data and easily referenceable by the AI system in a way that a human wouldn’t be able to. After all, the best way to learn is often to have someone experienced to act as a guide, demonstrating and showing best practices and answering questions.

In the future, overarching generative AI copilot systems will be able to leverage their connections with design and simulation tools, their training on and access to vast libraries of historical design and simulation data as well as their ability to intuitively communicate with humans to reshape the early stages of the design and prototyping process. Drawing on past design data, a generative AI model could be used to interpret a brief description or back of the napkin sketch and turn it into a complete, functional 3D model in a fraction of the time it would take using traditional methods. At the same time, it could leverage the power of machine-learning-based (ML-based) Reduced Order Models (ROMs) to assess the viability of proposed designs nearly instantly. While simply asking a computer to design something may seem like a far-off dream, the basis for this technology already exists today and is poised to revolutionize the design process one day.

An AI support system integrated directly into design and simulation software can flatten these tools’ traditionally long learning curve, providing novice users expert insight. By using AI to take on the burden of interacting with complex software, turning it into a new form of human-machine interface, users of every skill level cannot only contribute but also focus their efforts on creativity and innovation, not learning software. This cannot only increase job satisfaction but also allow for more complex products to be designed faster and more efficiently.

Capturing Knowledge through AI

As products and processes grow increasingly complex, they also generate large amounts of data during their design, testing, and utilization. While this type of big data is essential, extracting value from it using conventional methods can be difficult. AI, however, is well suited to make accurate data-driven correlations in these design spaces despite their sparse, non-linear, and vast nature. This key data is then used to train a reduced-order model capable of making inferences within that domain, effectively building an AI that understands the relationship between inputs and outputs and can ‘simulate’ the results of various combinations of input parameters without actually having to run a simulation. With this approach, a ROM can encapsulate the design knowledge inherent within the training data and offer a new way to utilize it.

An example would be designing an internal combustion engine for a car. This involves, among other things, a large number of complex and data-rich simulations that will need to be run throughout the design process. By training an AI model using early simulation data, extracting the pattern that exists relating various input variables such as material choice or fuel ratio to outputs such as torque or heat, it can then be used to accurately infer the outcome of a complete simulation in real-time instead of in hours or days.

AI isn’t just limited to capturing data for ROMs, either. Similar methods can be applied to help train copilot and support systems, encoding hard-to-capture institutional knowledge within the AI model itself. These models, presented to users as digital assistants and chatbots, can help train new users and show experienced users insights that might be otherwise difficult to extract from records or lost entirely if they were never written down.

ROMs Accelerate Simulation

Once an ML-based ROM is created, it can be utilized in several ways to speed up the design process and expand the design space. A ROM based on earlier, detailed simulations can run in real or near real-time. After a sufficient number of simulations have been run to train the model, it can replace the need to run a full simulation every time the user wants to inquire about a new design point. It is important to note, as well, that these models go beyond simple interpolation by using the patterns in their training data to infer the answer to similar but distinct cases that fall outside the range of the simulations they were trained on. That not only creates the equivalent of a high-fidelity simulation being completed in a fraction of the time, but in the case of needing to make a significant change in a parameter, the model can still provide guiding inferences without having to rerun long simulations completely. AI and ML-enhanced ROMs can continue to infer results well beyond their initial training data range, albeit with reduced accuracy, offering the ability to provide basic validation of significant changes quickly. Even though the model gets less accurate the further from its training data, an inference is created: having a rough idea of whether a concept is even worth pursuing before investing substantial resources into it is of great value to any company. Suppose a design concept is shown to be viable using existing ROMs. In that case, additional high-fidelity simulations are run, and the data can be used to refine the model further in active learning.

Beyond validating concepts as a whole, ROMs can also be deployed to help teams working in different domains design individual elements more closely in sync. Many modern products, such as smartphones or cars, are far too complex for a single person or team to quickly say whether a design concept is viable with many factors, such as operating temperatures and pressures, power consumption, or physical stresses needing to be accounted for before a design can be considered feasible with parts designed by different teams each contributing to the characteristics of the whole (Figure 1).

Figure 1

Fully simulating a car takes too long to quickly validate a design concept. With ROMs, it’s possible to validate requirements quickly (Image credit: Siemens Digital Industries Software)

While some of these variables can be checked easily, some require mock-ups and low-fidelity simulations to validate concepts. Even a low-fidelity simulation requires a great deal of time to complete compared to doing a few simple calculations or querying an AI model, reducing the available time for exploring the design space as each design takes longer to validate. However, leveraging a surrogate ROM instead makes it possible to receive nearly instant validation on whether, for example, a proposed design for an engine is thermally sound or if a smartphone design could withstand a drop from a particular height.

AI Redefines Design

Generative design is a pillar of the modern design process, bringing together design tools, simulation, and product constraints in a single, cohesive design space exploration process, which is already starting to benefit from the addition of AI and ML, including powerful tools such as copilots and ROMs. But what if things could go a step further? AI offers the potential to enable a genuinely autonomous design process, accepting a natural language input describing a part or product before using a custom AI-driven generative design model trained on a company’s own proprietary design and simulation data to perform every step of the design process (Figure 2). This includes everything from the initial design concept to validation and refinement before presenting a finished design for human inspection.

Figure 2

AI-driven generative design represents the next leap forward in the design innovation process, connecting a complete ecosystem of software with powerful AI automation and natural language processing (Image credit: Siemens)

To enable this new paradigm, companies will need to leverage custom AI models trained in-house, which has the twofold benefit of keeping proprietary IP data within the company that owns it and training a model that is fluent in a company’s design language. Keeping valuable design data in-house alleviates one of the major trust issues when bringing AI into the design process since no third party would ever need to interact with the data. At the same time, a custom AI model would better maintain consistency between past and present designs, maintaining a strong brand identity consistent with the years of hard-won knowledge and experience encoded within a company’s completed projects.

AI-driven generative design represents the next leap forward in the design process as new types of models are created that can interface directly with the complete ecosystem of design software from CAD to simulation and design space exploration, all while presenting and receiving information from users through the intuitive conversational interface enabled by large language models. In the future, generative AI will also be critical in early prototyping. It bridges the gap between simple descriptions or back-ofthe-napkin sketches and functional digital models, living up to its generative name in bringing ideas to life as working digital prototypes. Once this initial creation step is complete, traditional generative design processes take over, using custom AI-powered generative design models to refine further and optimize the design based on a wealth of past data and expertise to finally present a finished, highly polished design to the end user.

AI Realizes Next Generation Innovation

Artificial intelligence is finding its place across nearly every industry and product segment. While it won’t take root in every sector, what AI can offer to the design process is a boon too great to be ignored. In an industry that is both replete with rich data sources and one that must constantly strive to stay one step ahead of increasingly high demands for new, complex products, AI’s ability to capture and share knowledge, accelerate slow processes, and, most of all, bridge the gap between humans and machines cannot be ignored. Embracing AI on such a broad scale won’t happen instantly. Still, it offers the potential to enable and enhance creativity to address today’s problems and create tomorrow’s inventions.