Mythbuster: If we use AI in engineering we will no longer understand how things work.

One of the main myths around machine learning in engineering is that some engineers like to say: "Using AI to model systems will take away the knowledge of how things work and you end up with a system that nobody will understand anymore. Do you want to sit in an aircraft that nobody understands?"

Ironically, aircraft are a terrible example, because autopilots have been built for half a century using dynamic system identification methods. This is nothing other than a mathematical model that is automatically identified from real data - for which the fashionable term is supervised machine learning or AI nowadays.

Developing highly accurate mathematical models based on experimental aerodynamic flight data is what building good controllers and autopilots is all about, see for example here. Once you dive deeper into the topic, you will quickly find that the entire field is basically machine learning where engineers are trying to optimise the performance of complicated systems within a complex non-linear operating environment and that neural networks, Gaussian Processes etc. have all been used here for decades including for embedded systems that require certification. Here is a good introduction to non-linear system identification that also covers the typical machine learning basics.

In the field of flight vehicle system identification, data-driven, machine learning or AI models are not being condemned for not being purely physical. On the contrary, they are used complementary to physical models to

• gain a better understanding of a complex system
• simulate specific scenarios engineers more realistically based on real data
• predict the performance of a modified design
• diagnose and find faults
• run in real-time as part of an operating system

Now here is the part that is surprising: while this is really old news when building autopilots, it is still a rather unusual approach in the majority of other engineering applications even though the world of engineering is ripe with use cases where this can be applied.

The reason for that is that the majority of engineering problems are not revolutionary. And the approach above works in any discipline where current physical models are hard to set up or are biased. For example, modelling cardboard boxes is really hard but there is plenty of test data. Modelling ketchup is surprisingly hard but there are lots of data as well. Modelling the read/write head in a hard drive is hard but there is lots of data.

By now you have probably spotted the main drawback and the one thing all these examples have in common: machine learning only works if you already have real data. So you can use this only for cases where

• you have built a prototype and are trying to model it better e.g. you are trying to understand how your tires are affecting vehicle dynamics
• you have built products in the past and you are trying to learn more from their complex behaviour (e.g. a crash test database, Airbus A320, A321, A321neo)

Every sector of engineering in my experience has applications where physical simulation is hard, however, companies have already collected data to better understand the issues. These are the applications where machine learning can really help engineers to better understand the system in less time - in particular, if you embed the model in a really interactive user interface so engineers can model and understand much faster.

So in conclusion:

AI in engineering is not about avoiding to understand a system, it is about using algorithmic methods to better understand a system.

Machine learning models and interactive dashboards make it a lot easier to understand, predict and optimise complex system performance by basing the model on real-life data (and that's why this is literally Monolith's mission statement)