In an interview with Metal Market Magazine, Stephen Pratt, the CEO of Noodle.ai a US company that is working with Big River Steel explained progress on smart plants to date and the shape of things to come to Richard Barrett.
The long-term future of steelmaking will depend on todays rapid progress in digitalisation and big data analysis to create ever smarter, learning, automated plants. In the past few years, very rapid advances in data processing and computing capabilities have started to enable smart steel production.
Big River Steel (BRS) CEO David Stickler has said that his enterprise, in Osceola, Arkansas, USA, is a technology company that just happens to make steel. The 1.65 million ton per year flat-rolled steelmaker has embraced the benefits of using artificial intelligence (AI) to analyse data, generated by both its plant and business, to progressively improve performance and profits.
Noodle.ai is an artificial intelligence software company that focuses on complex enterprises and specifically on making very complex business decisions around operations. We focus from demand planning, demand shaping, all the way through supply, which is inbound materials, inventory management, scheduling, production, maintenance, and distribution said Pratt.
The company uses technologies that have only recently become available to make such complex decisions. The real breakthrough allowing us to do this is high-performance computing supercomputer technology. Where previously, processing a massive amount of data was not really affordable and not even attainable, said Pratt. Supercomputing has passed the teraflop barrier and is into the realm of petaflop computing power thousands of times faster than the fastest supercomputers available in the year 2000, he said.
People who built their steel mills in the 1970s, 80s, 90s, or even from 2000-2010, necessarily did things that are now akin to spreadsheets on laptops, said Pratt. What youre able to do today is process massive amounts of data on supercomputers and model many scenarios, creating much better decisions. This is having a dramatic impact on business, he added.
AI provides an opportunity to create probabilistic projections an important improvement upon static assumptions. For instance, in a mini-mill, calculating the probability of a caster breakout or a longitudinal crack, you can begin to optimise production, Pratt explained. Were looking at electricity consumption, and what mixture of steel products is optimal, given demand. Predicting steel width, you can reduce the amount of excess and can improve tolerances. Were also integrating logistics and inbound materials, to reduce working capital, he added.
A synthetic hedge
With Big River Steel, Pratt says that Noodle.ai has already put together a synthetic hedge for scrap steel prices, which he said is working well. Weve crunched huge amounts of data and identified tradable commodities, securities, and other instruments that, put together, act like a scrap steel futures market, Pratt explained. The overall objective for BRS is to maximise profit per mill hour and also take into account environmental impact. You see numbers coming out of Big River Steel that are dramatically better than what was possible even a few years ago, he added.
Noodle.ai is working closely with plant solutions provider SMS group, supplier of Big River Steels mechanical equipment as well as electrical and automation systems and process knowledge for the steel mill complex.
Selected data are fed into Noodle.ais BEAST (Beast Enterprise AI Supercomputing Technology) platform, which has petaflop computing power and trains each clients applications to allow processing of massive numbers of planning scenarios.
Big River Steel has thousands of sensors, and so crunching all those data and even external data, which are important for production, housing starts, automobile manufacturing, and internet behaviours, allow us to do better demand prediction. Thats it in a nutshell, Pratt summarised.
Pratt explained that Noodle.ais approach to learning algorithms AI and machine learning is based on the scientific method. You come up with a hypothesis: the prediction for an event is hidden within these data, for example. So lets crunch these massive amounts of data, and the algorithm will search for the patterns that correlate to the event that youre trying to predict, he explained.
An initial data transfer from client to Noodle.ai can comprise many terabytes of historical data. Once you have that historical data loaded, you only need to synchronize updates, Pratt explained. Then you come up with hypotheses on what data are needed to, for instance, predict longitudinal cracks in the caster. You ask yourself what data are relevant to that: caster speed, scrap composition, carbon content, mould shape, and temperature, for example all the things that are around the process.
A learning algorithm tries to find patterns. Then once it finds one, it says these are the rules as of right now. And then you get more data that says Hey, I just learned something, Im changing the rules. It is a dynamic learning process that improves over time. The more data you feed it, the better it gets, said Pratt. This evolutionary approach is unlike the fixed computational models or static software programs of the past.
Noodle.ai trains its learning algorithms on the BEAST. The core of the computing power of the BEAST is a supercomputer, but it also has about 200 interconnected software technologies to take the raw data, go through all of the data synchronisation, feature extraction, model selection, training of the models, and then putting it into production. We host the applications in the cloud and then we charge on a monthly basis for those applications: that is AI-as-a-service. Its a hosted cloud-based application, said Pratt. Sometimes a portion of the fees is based on the outcomes.
Pratt said that BRS consumes as much electricity as some cities, and being able to precisely predict energy usage based on a simulated schedule, at an hourly and daily level, allows BRS to shape their energy profile and production actions in a number of ways that drive immense financial and environmental benefits. That application has been working very well and that is highly predictive, said Pratt.
Production scheduling and planning are underway, but inbound and outbound logistics is done. Weve been able to dramatically reduce the cost of inbound and outbound logistics, said Pratt. You want to get as close to just-in-time arrival of raw materials as you can. You dont want to have a yard thats piled to the sky with various types of scrap.
Noodle.ai is also looking at scrap-mixes and DRI as a predictor of plant maintenance, because certain kinds of scrap are harder on the mill. If you use scrap containing a lot of copper, it changes the mill maintenance schedule. You can actually take into account choice of scrap not just for production or demand, but also how much downtime it will cause to the mill, Pratt said.
Noodle.ai is working with other steelmakers, But were much more advanced with Big River Steel, said Pratt.
We are building the foundation for the learning steel mill. Were putting intelligent learning algorithms at specific points within the mill, Pratt explained. Right now, we have the inbound and outbound logistics, but theyre separate from the optimisation of production. The next step is to deploy the layer on top of that, which is the interconnected layer, to take into account everything across the mill to optimise across the entire operation for profit-per-mill-hour.
Pratt emphasises that, for BRS, profitable steel production is much more important than tonnes of steel produced, Because sometimes you can make more money with stuff that weighs less. Is it best to go for timely high-quality-product delivery rather than quantity? Well, go for profit. Go for optimised return on invested capital (ROIC),
rather than tonnes of steel produced. Thats whats coming in 2018, Pratt concluded.
By: Richard Barrett