Applying machine learning to engineering challenges has both exciting implications and unique challenges. Here is how to get started in this emerging domain.
What is Engineering Data Science?
Long since before I can remember, I have been an engineer more into math that most, especially applied mathematics. I preferred classes on numerical methods and finite element analysis over hands on lab work. Since graduating, I’ve spent my career using, teaching, and creating CAE software that lets others design and improve their products. Over the years, I’ve observed that a lot of an engineers’ time gets spent on non-value added tasks unrelated to core engineering concepts. Meshing models and setting up optimization problems are time consuming, but ultimately they are only prerequisite steps to the final results. Improving the efficiency of these steps has a been a person goal of mine. Consequently, I’ve written many custom automations and taken part in many discussions on software design. But that early passion for numerical methods still interests me today and has led me to an interest in engineering data science.
Artificial intelligence and machine learning have catapulted numerical methods into the public consciousness under the broad term data science. The basic premises are not new, of course. Linear regression has its roots in the early 19th century and is still a viable technique in any modern data scientist’s toolbox. Neural network research began in the middle of the 20th century, although the minds who created those rudimentary models certainly couldn’t imagine the complexity of modern deep learning architectures. Regardless, at heart, machine learning is nothing more than a tool to solve problems. We don’t think of physics simulation software such as OptiStruct in terms the linear algebra that powers it; instead, we recognize how it enables efficient product design. In a similar vein, machine learning can be used to enable efficient engineering.
Ultimately, the goal can be imagined in a Venn diagram overlap that requires a familiarity with both engineering problems and an understanding the broad and rapidly developing area of data science techniques such as machine learning. Merging these two worlds together effectively requires integration of these numerical methods into engineering software to save time and offer new insights to make better products.
Realistically, data science has some unique challenges when applied to engineering. One place is data sparsity. Virtually crashing a car into a wall is certainly cheaper than the real thing, yet having enough data to satisfy notoriously data hungry neural networks can be an obstacle. Another complication can be the data itself. Many “text book” applications of data science are focused on tabular data, that is data that lends itself easily to spreadsheets. In contrast, the data in many engineering applications is more abstract such as a 3d solid geometry.
Eyeing an example, let’s consider the idea of computer vision. This is a very mature field for artificial intelligence, so the question is how to apply it? Motivated by the imagery we see in popular science about self-driving cars, one proposal is to find similar parts within a 3D assembly using image recognition techniques like convolutional neural nets. This would undoubtedly work, but in order to learn, the neural network will require many images and extensive training time. In contrast, the matching workflow in HyperWorks combines alternative data representations and machine learning methods to solve the same problem more efficiently with less data and no training. This example illustrates the application of machine learning enabled by specific domain expertise.
Working together toward engineering applications of data science
Nearly similar to how the wheels of slot machine must align to hit the jackpot, the art of data science is really about aligning three concepts for success: data representation, a machine learning method, and the application problem. Having two of three aligned won’t be sufficient. Stated another way: the right techniques can still be applied to the wrong problem. The job of an engineering data scientist is to match data and methods to engineering problems. To begin working in this area, I suggest working backwards in this metaphor though the following questions.
- What is an engineering challenge or pain point for your organization?
- Can you solve the problem using a machine learning technique?
- What data would you need to collect, and how can you get it?
I’d like to hear your thoughts about how data science can be applied in engineering. Even if all you have is an answer to the first question above, it is beginning of the journey and once you step onto the road, there is no knowing where you might be swept off to.
