How to automate virtual prototyping with AI?

Machine learning is all about using data to learn from it and automate the workflow. Many companies I know whether they are automotive, robotics or component manufacturers have ambitious goals to reach ultimate automation by 2025.

What do they mean by ultimate automation? They mean they want to click a button and a new product gets virtually created so I can start manufacturing it at scale in a robot factory. Can I see that work for a new hydrogen aircraft? Forget it. Can I see this work for designing new stainless steel screws? Definitely. However, designing screws is so simple that you can easily automate that without AI, too. So the question is as always: where is the sweet spot based on what is currently possible?

The answer to the question: how much automation can I get for my product development process depends on how many products you manufacture and how iterative your process is.

Let's say you build 10 cars per year, and you need about 20 iterations to get the metal components stamped correctly, that gives you 200 data points per year - not bad.

Let's say you build and customise 500 sealing solutions per year and each of them goes through 10 tests and design iterations each, which gives you 5000 test data points to learn from. If you run about 50 simulations for design optimisation for each that give you 25000 simulation data points.

Now let's look at some scenarios in detail:

1. 10 designs, simulations or test results:?All you want to do, is to make sure that the next engineer you hire has access to the previous design files and information you gathered to get up to speed faster. This can be work with 3 data points or designs. To give a simple example: let's say an aerodynamicist just joined a racing company and needs to design a spoiler for a new car. The first step they will do is to have a look at the last 3 cars the company to learn from them. You can make this search and learn process a lot easier if you save the data in interactive 3D + functional data dashboards than by going through old folders. You can also solve it by creating 3 very neat folders and making sure the data in them is structured neatly so that people can compare future cars, so no need to use AI here.
2. 50 designs, simulations or test results:?You want to use insights from product testing or development to help your engineers make better decisions faster. You have noticed that there is considerable repetition, and from the last 50 projects you have done you can learn quite a few things using algorithmic methods. You can detect correlations, you look at failure scenarios, you can build simple models to make recommendations of what to test next. At this data level, algorithms can be a nice extension of the engineering expertise you already have.
3. 150?designs, simulations or test results:?You can make build an AI model to predict the result of tests or physical simulations. For example, you could predict the performance of a rim in a wind tunnel test, predict the maximum stress in a suspension system, etc. The prediction results will be not great if the problem is hard, and mostly we see people use AI at this data level for faster decision-making at the pre-design stage and they build models based on their own test data that is not biased by simplifying physical assumptions.
4. 250?designs, simulations or test results:?You can build recommender systems that can deliver really useful insights into what other solutions you could try. You can run targetted optimisation codes that tell you how to design things differently. This is the typical size of 'design of experiment' for optimisation studies based on CAE models, so we tend to get a lot of those at this data level.
5. 500?designs, simulations or test results:?You can build AI models that can predict the outcome of repetitive processes with good accuracy - sometimes as good as or even better than simplified physical simulations. This tends to be really beneficial for companies as they end up saving a lot of time and money on performing tests or running simulations if they can prove the use of an AI model for this scenario.
6. 1000?designs, simulations or test results:?You can build fully automated workflows. This is every engineering CIO or CTOs dream. Imagine this: a customer provides your team with their requirements, and you enter them into an online form where an AI algorithm will go into your PDM system and create a new product or component for these requirements fully automatically. I have seen this work for repetitive components like sealing solutions, pumps, bearings etc. so in general for suppliers who create many 1000 versions of the same component every year.

Here is an example of a dashboard where a user created a fully automated design workflow based on 500 designs and simulations.

When you enter your Performance Targets in the left corner Monolith returns:

1. A CAD file that was generated automatically from previous CAD files without the need to manually create a parameterization
2. Fluid flow simulation results (CFD)
3. Structural design simulation results (FEA)