Reflections on AI in Industry

Is AI powerful? If you still have doubts about this, read “Superintelligence: Paths, Dangers, Strategies” by Nick Bostrom, and/or “Homo Deus” by Yuval Noah Harari.

Yes, there is no doubt that the power of the AI is big enough to scare us both on the good and the bad sides.

In a recent article published by MIT Technology Review we see the side we want to develop as quickly as possible, AI applications that help the human being when the use of mathematical or physical models has limitation. It is in this context that AI fits into industrial applications, which is the focus of this article.

When reading the article published by MIT I learned three lessons:

AI projects need to be done with four hands

The first point is clearest of all, especially when the focus is on the industrial world, particularly process industries. In short, we are saying that a good AI project will not be built by a data science wizard who will knock on your door with a ready-made solution. Before the solution comes the problem and who knows the problem is the end user. Therefore, whatever the proposed solution, it will be the better the better the problem is described. In other words, there is no AI project without the active participation of everyone involved. I often say that AI brings with it the need for increasingly stronger process engineering departments, reversing a trend from the late 90s of reducing this type of professional and the outsourcing of responsibility to large technology-holder equipment manufacturers process. A subject I’ve covered before in an article that focused on a single sector but fits in the same context. In summary, although the point is very clear, this does not mean that the industrial sector has fully understood this new environment. There is certainly room for improvement and a lot.

AI can help us a lot when faced with complex problems.

The second point is very interesting, there are many people wanting to use AI for industrial processes that already have their mathematical models defined. There’s not much sense in that. Using AI to do mass balance or energy balance, or to model the speed of a chemical reaction that already has its kinetic behavior determined, should not, in principle, be the best types of applications for this technology. Of course, there are exceptions, the MIT article shows us that there are situations in which the number of parameters can be so large that even if the phenomenon has already been described by mathematical equations, the resolution can become so complex that even the most powerful supercomputer does not be enough to complete the task in a timely manner. And in this case AI can be useful.

In the industrial world it is not very different, but we should not think of extreme situations like the one in the MIT article. We must think of AI for applications for which the models do not exist or are bad due to the degree of uncertainty of the variables involved. Situations where the amount of data to be manipulated is large in volume and variation, where interactions between different unit operations make modeling complex. But we don’t stop there, we also should look beyond the production process, why not look at the administrative processes of a plant? For example, optimization of resources for maintenance routines, why not?

Whatever the application, what we don’t have is still a “big-brother” AI; that seems to be the dream and some end users, something ubiquitous and omniscient of everything that goes on in a plant. In the current stage of AI, at least at the industrial level, we must focus on dedicated and constrained applications, aimed at solving a specific problem and not just any problem. We must not overestimate the power of AI or else it will lose credibility and fail to take advantage of its benefits. Many talks about autonomous factories, but that won’t just come from AI, it’s actually just part of the solution.

Extraordinary results in AI still require extraordinary resources.

Finally, the third point. If by chance you’ve read a few books or articles on AI cases you’ve certainly finished reading surprised at what algorithms can do. Obviously, this inflates anyone’s expectations. The MIT article is a clear example, you read and marvel at the potential of this technology. But take a moment to think about it, we are talking about a solution designed by DeepMind in conjunction with a highly respected institution in its field, the Met Office, that is, there was no lack of human and technological resources for such project.

These are not the types of resources that companies can count on in their daily lives. And what is the conclusion here? The conclusion is that one should start with problems with low complexity, preferably recurring, those for which a machine operator or a production manager almost intuitively says: “look if you have something that looks at all these variables here, I think will be able to infer what is to come.” Obviously, as emphasized in point 1, work together. If DeepMind needed to work together with the Met Office, imagine AI companies that are at the base of the pyramid, or plants that have modest process engineering resources.

This resource limitation, present at both ends, provider and client, impacts the execution time of an AI project, as well as the risk of not having the expected end. And this is not necessarily linked to a deficiency or incapacity of the AI ​​models, even though it can occur.

In conclusion, AI will help companies a lot in complex problems and has infinite potential for that, but companies must have clear strategies about what they expect from AI and how much risk they want to take, there must be awareness that they are projects, that are made with four hands in which duration and risk are inversely proportional to the available resources. Just to avoid excessive expectations.