Recent AI tools have proved to be so helpful in both creative and technical disciplines that knowledge workers dealing primarily with text and speech – in particular in sales, marketing, support, consulting, or legal – adopted them very rapidly. A recent survey by McKinsey found that the number of companies using AI in at least one business function jumped from 33% to 71% in the span of just 18 months.
This growth has also been fueled by an equally rapid expansion of model capabilities. The first steps toward multi-modality came quickly and introduced the same text-to-output inference to other content types. We already almost take for granted the ability to generate high-quality images, video, and source code through such tools.
Can AI Generate Engineering Output?
Mechanical engineering teams have adopted these tools as well to accelerate all sorts of work processes. For example, to analyze and summarize RFQs faster or to search faster for technical information. But these use cases are mostly adjacent to the core engineering work and mechanical design. So why is it that we can ask AI to generate very useful text, images, video, and code but not a useful engineering design?
Let’s consider how these types of AI models are trained. Generative AI models have been trained on trillions of tokens, primarily from the internet. Transformer models on huge datasets of public text/code and diffusion models on equally large datasets of text-image pairs. Not only is this training data available in vast quantities, but the data formats are also very straightforward to read and use for model training.
Things look rather different in the engineering realm, the most obvious challenge being that, unlike text or source code, there is little to no public product design engineering data available. Then there is also the question of data quality, in the sense of whether or not a given design is fit-for-purpose, meeting the requirements that it was designed for. Added to that is the fact that the most widely used data formats storing mechanical design information are proprietary, requiring commercial licenses even to read it, let alone manipulate it. In summary, the idea of obtaining and processing millions of engineering designs to train a generative model still looks like a very challenging ask today, but technical progress in this field is happening fast.
Does That Mean That Core Engineering Work Will Remain AI-Free for Now?
Absolutely not. In due course, novel AI approaches might rise to the challenge of handling big chunks of typically manual engineering workflows, possibly including the transformation of a text prompt into a meaningful design, but it is going to take time to get there.
Meanwhile, there are AI engineering workflows that are easier to attain while still very helpful. We can get a long way by using AI to speed up the cycle time for a single design iteration to such an extent that it appears to be instantaneous. We will do this by accelerating all of the steps in the workflow, including CAD generation, model preprocessing and setup, simulation workflows, and the analysis of results.
Once we have all that proven out, an AI agent can then drive the (accelerated) machine, taking design decisions along the way and looping around to discover optimal solutions.
Replacing a human-in-the-loop with a machine-in-the-loop in this way has the advantage of leaving the workflow and toolchain fundamentally unchanged, with the AI system ‘driving’ the tools in the same way that a human does. This means the human can easily understand what is being done and intervene at any point. Most importantly, the human can provide input to direct the AI, for example, where a design needs to balance competing objectives – decisions that require careful consideration and mutual understanding.
Not Just a Case of “Prompt Engineering”
Let’s dig into how we deploy AI to accelerate and augment engineering workflows. Let’s start by looking at how these processes work today. They tend to be centered around the manual engineering work where humans make decisions to advance the iterative design by designing and evaluating the design’s performance, depicted in green below. The CAD system involved can be conceptualized as a computational process going from parameterization to geometry (yellow) and the CAE system going from the simulation setup to the results (blue).
This is a very simplified conceptual view of the engineering process, but helpful as it differentiates between the unstructured, human workflows in the middle and the purely computational ones left and right. All three can be automated already, to search through a prescribed design space for example. But this automation is very much rate-limited when using so-called traditional physics solvers to evaluate each design. What’s more is that AI can transform this process into something not only automatic, but autonomous.
Introducing Physics AI & Engineering AI
Let’s tackle that first bottleneck of simulation run time (the right-hand block in the diagram above). Depending on the physics and fidelity needed, a computing time of hours to days is not unusual. A growing set of AI methodologies to speed up this solve process is available, from deep learning surrogate models that replace full physics solvers to tools that speed up those ‘traditional’ solvers. Given the availability of a suitable, pre-trained, method, you can reduce the solve time almost to zero. We call these ‘Physics AI’ methods to indicate that, at the core, it’s about predicting physics with AI, and with the big benefit of being able to do that very fast.
The second, more dispersed bottleneck visible in the process is the human interaction needed to go from a given design to a well defined simulation setup, then to consider the results of that simulation, and lastly to determine which point in the design space to look at next (the middle block in the diagram). These are all steps where an AI agent can assist, facilitate, accelerate, as well as act autonomously – performing complete workflows by operating on the existing tool stack just as a human would. As such, it is performing a series of discrete and logical steps that can be justified or even debated, as you might with a colleague. Since this agent is performing the core engineering work for you, we refer to it as ‘Engineering AI’.
Lastly, let’s turn our attention to the left-hand block – the CAD definition of a design. Once a model has been created and parameterized, generating a new variant based on a new set of parameters is already near-instantaneous. What is very much slower, though, is the process of creating that CAD model in the first place.
There are several exciting technologies emerging in the CAD space that could make the process of CAD generation far faster and more robust. Latent space parameterization, implicit representations, and cloud-native BREP are just three such technologies that could enable vastly faster design iterations, and we are actively working on integrating them into SimScale.
We Are Placing AI Tools in the Hands of Every Engineer
Thanks to its cloud-native architecture with built-in AI infrastructure, SimScale is uniquely able to provide AI features to help you navigate engineering workflows and accelerate performance predictions by leveraging your simulation data in the cloud. As we have explored so far in this blog, unlocking value from AI means touching almost every aspect of the simulation workflow. It requires a deep and immediate connection to models and data which is only practical to do in a cloud-native stack.
Engineering AI and Physics AI are built into SimScale in such a way that it can become second nature to use these tools to supercharge your productivity. SimScale users do not need to deal with any of the typical headaches experienced when attempting to deploy AI tools such as data cleaning/organizing/relocation, model versioning and management, or provisioning of suitable GPU resources for model training and execution. All of these are taken care of by the vertically integrated tool stack and intuitive user experience.
At NVIDIA GTC 25, we announced that we are making it even easier and faster to adopt Physics AI for certain applications by building a set of pre-trained foundation models. The unique aspect of these models is that they are pre-trained on a broad set of designs, providing users with a Physics AI model that they can use out-of-the-box or that they can augment with a small amount of their own proprietary training data. To learn more about foundation models in SimScale, check out this blog.
Unlock AI Value by Selecting an Impactful Application to Start With
Once you have test-driven the capability, the next step is to test-drive the value unlock. Each engineering team we work with has unique legacy data stored, sometimes from decades of engineering work. We frequently see teams expecting to start there, trying to find value in it. The reality is that finding and processing legacy data can be an immensely difficult task, and one that may take a very long time to yield results, even if useful data exists.
We recommend a different approach: Select an engineering process in your organization that – if collapsed to seconds – would create hard value for your organization (revenue or costs) and try tackling that with an AI-powered workflow.
Remember: The best time to start leveraging AI systems in your engineering team was yesterday. The second best is today – give us a ring!
