Breakthrough or Hype? Myth-Busting Common Claims in Engineering Data Science
Altair EmployeeI actually appreciate meeting engineers who are skeptical of machine learning. We should ignore the hype about AI and instead, for each task in our own jobs, focus on weighing the benefits of a machine learning approach against the existing ones. And while there are certainly cases when ML is forced onto the wrong problem (let's call this "wishful ML"), I'm a firm believer that machine learning has become another important tool for engineers of all kinds.
Today I'd like to myth-bust some of the most common claims that I hear in the industry. Note that the following are only my opinions, and if you disagree I'd love to hear about it. Ok, let's get into it!
"There's no point in using machine learning to predict physics because we already know the governing equations."
False.
Just because we know the governing equations for fluids or structures doesn’t mean we have a quick way to solve them. Depending on the complexity of the physics and geometry, traditional simulation methods like the finite element method can take anywhere from seconds to days to arrive at a solution. Machine learning offers a shortcut. By running a few simulations and using them to train an ML model, we can produce a surrogate model that is potentially hundreds of times faster than the simulation but limited to a specific application.
"Machine learning is really just curve fitting."
False.
Fitting a curve to points is a form of supervised machine learning called regression, but that’s just the tip of the iceberg. ML can also be used to predict whether you’ll find something to be beautiful, and help identify parts based on their geometry. It can organize data into groups (see how BMW clusters simulation results automatically), extract features from complex data formats like images and geometry, or detect when something is strange or unusual (see Digital Twin Anomaly Detection and A Visual Introduction to Novelty Detection). Machine learning can even be used to explore new part designs. What these examples have in common with curve fitting is that an algorithm has found useful patterns in the data.
"Machine learning models are a black box so they are not trustworthy."
True and False.
One aspect of trustworthiness is accuracy. The question is never whether the model is right or wrong, but rather whether it is accurate enough to be useful to you. That depends on your application. The way we determine the accuracy of the model is to test it on some data that’s similar to the data that we will encounter in real life.
Besides accuracy, sometimes it’s also important to know why a model has made a particular decision in order to trust it. Some algorithms like decision trees provide very clear explanations for why the did what they did. Other algorithms like neural networks are a bit more opaque, but it’s still possible to gleam some insights. For example, you could change one input feature and see if the output changes.
"Machine learning will never take the job of an engineer."
True.
I don’t believe machine learning will ever replace engineers, but it is certainly already changing what engineers do. Many low value-added engineering tasks can be automated by machine learning. For example, machine learning models can identify parts of an assembly to automate the creation of simulation models. They can make fast physics predictions when simulations are slow. And they can automatically sort through simulation results to find natural groups of behaviors that it would take a human a long time to spot. These efficiency gains give us, the engineers, more time to apply critical thinking and solve the real challenges in product design. Machine learning is simply another tool.
Do you agree with these takes? What myths would you like addressed in EDS? I'd love to hear about them