Time to Rethink Risk Itself As AI Hits Industrial Defect Detection

It's bigger than you think

Consider: why does a modern bomber fly?

In the most direct sense -- yes, because of its engines. And fuel. And wing shape. And many innovations in aerospace.

However, suppose a severe fault developed on that bomber, unknown to all. It would crash at an unexpected point in its mission. Given its potential payload and the type of its mission, it would spell multiple kinds of disaster. Subsequently, any further flight of this kind would be exceedingly hard to justify.

For all complex, mission-critical technology, advanced defect detection (often called predictive maintenance) is what enables the actual use of the system. The bomber flies because we know, to a more-than-sufficiently-high accuracy, that its components contain no dangerous defects.

Hence the interest in new innovations in defect detection. For example, attention has turned to using neutrons to detect defects, rather than traditional X-ray-based and other detection techniques.

Turns out, you can shoot neutrons at complex machine components to get useful information, such as aircraft turbine blades, spacecraft parts, and artillery shells. Much like an X-ray would. Except neutrons are more sensitive to heavier materials, making them ideal to detect many phenomena for which X-rays are unsuitable. Where X-rays penetrate, neutrons often scatter, leaving detectable traces, which can now be analyzed digitally by machine learning. Called neutron radiography -- an imaging technique used for non-destructive testing of materials, which has been around since the 1960s.

The challenge thus far? Until now, to make the neutrons, you needed a nuclear reactor, making it impractical for most commercial applications.

Companies are now building on emerging industries to make neutron radiography viable. Greg Piefer, CEO of SHINE Technologies, uses technology developed for its eventual application in commercial-grade fusion power -- namely, the ability to fire a beam of deuterium atoms at tritium in a gas -- to produce a suitable neutron source. No need for nuclear reactors anymore.

Piefer suggests AI can play a significant role in enabling new capabilities in neutron radiography, as a way to improve the accuracy and speed of the imaging process. He states:

"[Neutron radiography] is an application where I think a machine learning algorithm can very easily find the defects. It's mostly a speed and cost improvement. So what it would do is give you much, much better speed and certainly better cost -- because we wouldn't need to pay people to read all these images."

Speed and cost improvements are valuable, but typical for applications of machine learning. More tantalizing however: we discussed the space after defect detection, where new kinds of value can materialize. For example, using neutrons to scan a cargo container opens a new possibility: a machine learning model could be trained to detect problematic content of cargo containers, such as by using modern object recognition models, and thus tease out statistical insights which would otherwise require a team of human experts to synthesize manually. Piefer described:

"With cargo containers -- unlike with turbines where we know exactly what we're looking for -- who knows what's in that container? It could be anything. This is where such multimodal detection would be really useful."

In short, machine learning can plug into neutron radiography to automate key technical processes from the get-go, as well as discern complex patterns after defect detection. This is possible because of another trend in the industry: new digital cameras instead of film to detect the neutrons.

Hence, it turns a traditionally analog capability into digital. In my interview with Will Flanagan, Director of Research and Development at Cerium Laboratories and Affiliate Professor of Physics at the University of Dallas, he described:

"Currently, industry standards in neutron radiography are analogue. There are long volumes you can buy about exactly what you need to do to ensure that the radiograph was a high enough quality and exactly how you need to inspect the image. But once this field goes digital, it's the wild west."

"The second you are digital, you can of course apply any handful of multivariate techniques that are not machine learning. But currently, neutron radiography is the central application we're looking at. We're very excited about exactly this topic of going towards machine learning, or the fact that as soon as they're using cameras, instead of film, machine learning is possible -- because you're digital instead of analogue. You can't really apply machine learning to analog."

The emergence of machine learning capabilities marks a significant, all-industry inflexion point -- for every future digital transformation. From now on, any industry switching any technology from analog to digital can decide to make an immediate jump to machine learning. As long as it adds value, is technically viable, and suited for the specific regulations of an industry, AI can now immediately replace traditional software algorithms.

Naturally, this requires a pipeline of data science talent. In my discussion with John Sturdivant, Managing Principal at Affirm Intelligence Partners, he described how they're seeing emerging demand for software engineers who are also capable data scientists. For a technology like machine learning which can plug into almost any industry, keeping team size small helps to reduce many kinds of risk. Hence, demand for cross-functional talent will only increase.

The global market for industrial defect detection is projected to reach $5.1B by 2027, currently growing at a rate of approximately 6%. And AI-based anomaly detection after defect detection will produce currently unknown capabilities.

Modern bombers are only allowed to fly because modern defect detection assures us that within a suitably tiny error margin, the plane will not crash. Likewise, by minimizing risk even further in many areas, leveraging ongoing industry-wide technology trends towards digital, AI will enable new mission-critical technology, which at present would present too great a danger.

Time to rethink risk itself.