Applying artificial intelligence (AI) in engineering has recently become a standard practice in the industrial sectors to boost productivity and efficiency. As AI progresses across various fields and new capabilities are quickly shared on social media, it can be very tempting to think about an engineering problem and instantly consider how AI might be used to solve it. Despite this thinking, for many years, engineers have solved problems using established engineering calculations, statistical methods, physical models, experiments, inspections, and most importantly, their own experience and technical judgment. These approaches are still widely used today and, in many cases, continue to work very well. Well, “Just because AI can be used for an engineering problem, does that mean it should be?”

This is where engineering judgment becomes very important. Before deciding to use AI, it is worth asking what the actual problem is, what data are available, and whether AI can provide a meaningful improvement over the existing approach. AI is certainly powerful. Machine learning models can analyze large amounts of data, discover relationships that may not be obvious to humans and generate predictions within a short period of time. However, having access to AI does not automatically mean that every engineering problem should become an AI problem.

Imagine a manufacturing company that has collected several years of machine data. The information may include temperature, vibration, pressure, operating hours, maintenance records and equipment failures.

At first glance, this appears to be an ideal opportunity for AI. A machine learning model could be trained using historical information and later used to predict whether a machine is likely to experience a problem. Engineers could then schedule maintenance before a failure occurs instead of waiting for equipment to break down. This is one of the commonly discussed applications of predictive maintenance.

But before building the model, several questions need to be answered. What exactly are we trying to predict? Do we have enough examples of equipment failures? Are the sensor measurements reliable? Will predicting the problem earlier actually help the company make a better maintenance decision? These questions are important because having a large amount of data is not the same as having useful data.

AI models are based on training data. That means that if the data the AI was trained on doesn’t resemble the true nature of the engineering problem you want it to solve, then your predictions won’t be very good. This is an example of something called Garbage-In-Garbage-Out, or simply GIGO, which is a common concept in all forms of data analysis.

There’s another problem here too. Sometimes there’s a much simpler way to fix the problem. For instance, if an engineering relationship is fairly well-understood and if you could calculate it accurately with known physics, statistics, or industry standards, then you may not need the extra power of an AI. The goal should never be “use the best tool” just because it exists. The goal should always be “pick the right tool.”

With that said, we know that the strength of AI lies in the ability to take advantage of very complex relationships within engineering systems. Take a piece of industrial equipment like a pump with multiple sensors that run continuously through varying loads and environments. One single measurement like vibration may not tell you anything bad. But a specific combination of vibrations, temperatures, pressures, and histories may tell you a long time before actual failure.

AI can find those combinations over massive amounts of data. Other examples include manufacturing defect detection via computer vision; energy consumption forecasting using machine learning; anomaly detection for finding odd behavior in machines; optimization for searching huge spaces of potential designs or operating parameters. But even after creating an accurate model, this is only half of the solution.

Let’s say an AI model gives us an 80% chance our machine is going to fail. Do we stop production now? Inspect the machine? Schedule maintenance as soon as the next shutdown?
No one wants to have someone else making critical engineering decisions without their input. Even though the decision isn’t made by the AI model itself, engineers still have to know the equipment, the risk associated with failure, how expensive it is to shut down for maintenance, what kind of working environment they’re in, etc.

That’s why expertise related to physical systems is so valuable. They have to determine if what came out of the AI model is realistic given how things work in reality. The value in having such high levels of processing capability may not lie in completely replacing engineering with AI. Rather, it will likely come from leveraging the advantages of each area. As AI becomes increasingly easy to obtain and utilize, I expect to see many new types of engineering applications emerge. While some will be transformative, others will create unneeded complexity. Therefore, regardless of whether an engineer is developing or utilizing an AI product, researchers and organizations should begin with two key questions:

Firstly, fully understand the engineering problem you’re attempting to solve.

Second, be sure you understand the data you have available.

Finally, and only after understanding both items above, ask yourself if using AI is going to give you a better solution than you currently have.

In a nutshell, when answering the question, “Do we really need AI for every problem?” in engineering, perhaps the starting point shouldn’t be, “What can AI do?” but rather, “What am I actually trying to solve?”

Written by: Dr. Mas Omar Mas Rosemal Hakim, Senior Lecturer, Faculty of Artificial Intelligence (FAI), Universiti Teknologi Malaysia (UTM)