How Will AI Shape the Future of Steel?
Jane Zavalishina
President and Co-founder, Mechanica AI
Mechanica AI team has recently participated in the Future Steel Forum in Warsaw. In this blog, our President and Co-Founder Jane Zavalishina shares her key takeaways from the event.
AI has taken its place in the agenda
Last year, not a single presentation at this same conference had the words "AI" in the title. This year, we had a whole session dedicated to the topic, with quite a few examples of real-world use cases that were carried out with the help of AI technologies. For example, Diego Diaz from ArcelorMittal shared a case study on using machine learning for defining optimal slab yard configuration, and Dr. Valentina Colla from Scuola Superiore Sant'Anna presented a number of projects, including prediction of flow stress and mechanical properties of flat steel products.

This is quite an interesting development, given a large number of technologies that jointly form the "Industry 4.0" initiative. Clearly, AI with the promise of new levels of automation and operational efficiency takes a special place in the list.
Still, it makes sense to reiterate the basics
At the same time, despite all the hype there is lack of shared understanding on some basic principles of technology. Given how different - and sometimes counterintuitive - it is if compared to traditional knowledge-based modeling, grasping the whole set of implications is not an easy task. Just a couple of examples of the questions that were raised during discussions:
Can we build an AI model for one blast furnace and apply it to another plant?
Correct answer: no, but the process of building the model will be quite similar so part of the effort can be reused.

How is AI different from predictive modelling?
Well, it is very advanced predictive modelling in a sense: we are just doing the same things, but better.

How we should apply deep learning in steelmaking?
Maybe you shouldn't! Despite its very attractive promise of truly self-learning models, it is often not the optimal technological choice for the limited industrial data.

Such knowledge exchange and recognition of both possibilities and limitations of the technology among the business leaders is essential to foster widespread adoption.

In case you need a refreshment on the meaning of the key terms related to the industrial AI topic, check out our previous blog post.
Should all managers start taking courses in data science?
While choosing the correct technologies, model architecture and optimal tools are indeed very important tasks for a data scientist, business leaders should focus their attention on obtaining a different set of skills. Those are less about technology and more about management: how to make AI bring monetary results? How to set tasks correctly? How to choose the metrics that actually connect model quality with business value?

For example, contrary to common sense, having the most advanced tools or the most "accurate" models is not always at the best interest of the business. Model with low precision but high recall can lead to some false positives (think: you will need to send a technician to do one unnecessary check) but at the same time reveal a potentially very costly defect on time. Such practical usefulness of the model should be more of importance than scoring high on some arbitrary metric - though "99% accuracy" of course makes a better slide title. Same goes for the most advanced techniques: marginal quality improvement may be accompanied by high opportunity costs and lengthy development cycles.

Keeping the focus on actual business value and choosing correct metrics and problem settings is just one of the aspects that needs to be learnt in practice to successfully deploy AI.
The learning curve is still ahead
In absence of universal best practices, multiple other challenges of real AI use are to be learnt through own hands-on projects. This includes defining restrictions and targets for AI to optimize, building the new culture tolerant to experiments and failures, resolving the matters of trust and responsibility for the automated decisions, learning to collect the right data, and so on.

Designing some high-level AI strategy from scratch is likely to become a costly mistake unless it is built on the foundation of real experience of working with the technology. Starting with local, isolated projects is the best, and seemingly the only way to prepare for the AI future. We encourage all companies to start on their AI journey as soon as possible, to not miss out on what is going to become one of the most disruptive changes for the whole industry.

Curious how to start your first industrial AI project? Reach out to us.
We'd like to thank Quartz Business Media for organizing and hosting this high-quality event. It is indeed a pleasure to see the whole industry gather and discuss the main challenges that lie ahead, and we are looking forward to participating in the next editions!
More posts you may like
Made on