Adopting a machine learning-based workflow reduced one Gerdau plant’s costs by $3/ton, as well as conserving over half a million pounds of alloys and reducing quality variation by 15%, according to results shared at last week’s Digital Transformation Forum in Pittsburgh.
During the event, organized by the Association for Iron and Steel Technology, Carter Almquist from Gerdau’s Petersburg mill joined Fero Labs Head of Operations Pamir Ozbay on stage to discuss the results and share best practices with other steel industry personnel interested in digital transformation.
According to Almquist, Gerdau's objective was to reduce raw material consumption while maintaining the company’s high quality standards. Achieving this goal, however, was not easy due to several sources of variation in the production process.
Scrap quality was one source of variation, causing each heat to have a different chemical composition. Similarly, at the rolling mill stage, furnace and mill temperatures and other metrics were constantly changing. In order to maintain quality standards, the team implemented conservative raw material limits. But while this kept quality high, it also resulted in high costs.
The mill had already conducted process capability analyses to understand this variation, but these had limitations. The statistical methods employed in such analyses are typically linear and assume a normal distribution of data; in real life, a production process is highly non-linear and complex. Additionally, the scope of the analysis is often restricted to a specific set of process parameters, so it may not take into account other key secondary factors that could be influencing quality metrics. Therefore, relying solely on process capability analyses didn’t provide the Gerdau team with a complete understanding of the factors affecting the quality of their steel.
Another limitation of capability analyses is that they can’t handle large amounts of production data or adapt to changing conditions. The Gerdau Petersburg mill produces over 4,000 heats per year and tracks hundreds of process readings, and production conditions are constantly changing. Yet the team was only able to perform capability analyses once or twice a year, and only on a small subset of production data—meaning significant value was being left on the table.
Applying machine learning
Fero is a no-code software solution for manufacturing teams. Metallurgists, quality managers, lab technicians, process engineers, improvement facilitators, melt shop operators, and others can use the software to solve the specific challenges of their roles—whether it’s quickly cleaning data or creating and deploying ML models.
Using Fero, Gerdau's engineers were able to create a digital twin of their unique process in minutes, without needing any coding or data science knowledge. They could then make real-time predictions and optimizations in a few clicks by utilizing the software's ML models. In contrast to traditional capability analyses, which only consider linear relationships, ML models take into account hundreds of process parameters and the non-linear connections between them. Typically these models are "black-box," meaning that users see only the end result (a prediction or recommendation) but have no idea how the model came up with it. Fero's models, however, are white-box--allowing the user to explore all the insights gleaned from the data, including root causes and relationships between factors. Thanks to Fero’s live connection to the factory floor, these models could continuously learn from the latest production data and remain highly accurate, no matter what changes occurred in production.
At the Pittsburgh forum, Almquist focused on alloy reduction to show how the Gerdau mill has realized the benefits of machine learning. The mill has about 200 products and 25 grades, each with rigid quality specifications. Before using Fero, the alloy addition range for each grade was set at a conservative level to ensure that those quality standards were consistently met, despite the complex, dynamic nature of the production process. Fero, however, told operators the minimum amount of alloy they needed to add to each heat in real time to meet quality specifications—saving $3 per ton of steel in raw material costs.
The diagram below demonstrates how the optimization works. The live ML model renders the optimal alloy recommendation within 30 seconds of a spectrometer reading; the ladle operator then adjusts the heat chemistry based on that recommendation, ideally after treating the slag. Each recommendation is unique to the chemical composition of the heat and final product specifications.
Crucially, the process also makes the job easier for operators. Below is the screen the operators are using, which tells them how much alloy to add in a quick and simple interface.
During the presentation, the speakers emphasized that gaining trust in a new machine learning workflow takes time, and requires a sustained focus on change management to achieve a flywheel effect. Over time, as the use of machine learning increased, so too did the savings realized by the company. This underscores the importance of patience and persistence when implementing new technologies, and highlights the need for effective communication and collaboration with plant operations to facilitate a smooth transition. Ultimately, the rewards of investing in machine learning can be substantial, but a measured approach and commitment to continuous improvement are critical to achieving long-term success.
Improved sustainability was another benefit of the new workflow. Reducing unnecessary alloy additions conserved over 500,000 pounds of manganese, carbon, niobium and vanadium. This has a direct impact on the global supply chain by reducing demand for mining, refining, and transporting. Both speakers noted that the main driver behind this reduction was increased day-to-day use of ML software by the mill operations, while the secondary driver was models learning over time from new data obtained through production.
Last but not least, adopting the machine learning workflow helped Gerdau build a leaner production process with less quality variation. The team observed a 15% reduction in quality variation in production that utilized the machine learning workflow. This finding highlights that a streamlined operation, equipped with the appropriate tools, can lead to improved product quality and significant cost savings. Since 2018, Gerdau has expanded the use of Fero to five other plants in North America and now leverages the tool to address other process challenges beyond raw material reduction, such as minimizing defects, improving energy efficiency, and enhancing production yield. This proactive approach ensures that Gerdau remains at the forefront of innovation and drives long-term success for the company and its customers.