AI-Driven Inclusion Analysis Saves Energy and Improves Yield in Steel Production
Production of iron and steel ranks fourth in US industrial energy consumption, at 6% of total usage. About 80% of this energy goes to making liquid steel. In the critical molten stage of the production process, the presence of nonmetallic inclusions and trapped particles may clog machinery and spoil 1–2% of the finished steel—a tremendous waste of energy and resources.
Inclusions and particles are typically evaluated via scanning electron microscopy (SEM) and energy-dispersive x-ray spectroscopy (EDS) to determine their composition and devise a control strategy. This requires physical samples and takes hours or days—far too slow for real-time intervention.
ArcelorMittal North America steelmakers (AM) and Lawrence Livermore National Laboratory (LLNL) collaborated to automate the analysis by applying artificial intelligence in a computational model that determines the composition of an inclusion from its image. The team developed a Python-based program, "AI-Driven Inclusion Analysis of Steel” (ADIAS), with a graphical user interface for uploading inclusion images and tools for data visualization. The software generates a 3D projection based on 10 million steel-inclusion images and uses an interactive ternary plotter for visualizing inclusion chemistry. ADIAS’s modeling tools predict inclusion class and chemistry. By exploring various deep-learning models via LLNL high-performance computing (HPC), the team found an optimal convolutional neural network with 60.8% accuracy in inclusion classification and a mean absolute error of 2.6% for element-percentage prediction.
The large number of inclusion classes within the dataset and close resemblances among images presented a challenge. The ability to cycle through many deep neural network morphologies using the High-Performance Computers provided great results and opportunities for follow-on work.
This achievement allows near-real-time analysis of liquid steel and will improve quality and yield while reducing production costs and CO2 emissions. The cost savings from a 1% reduction in yield loss for the whole U.S. steel industry is estimated at $450 million per year, owing to material, energy, and labor benefits. Energy savings alone will reach nearly 15 PJ per year—equivalent to about 110 million gallons of gas, or enough power to run 350,000 typical American homes for a year.
For excellence in the management and execution of this project, the AM/LLNL team garnered an LLNL Global Security silver award and took second place in an ArcelorMittal competition for best emerging technology.
