AI-Driven Accelerated Inclusion Analysis for Energy Efficient Steelmaking (Final CRADA Report)
Castillo, Vic, Hu, Yeping, Chukwulebe, Bernard et al. · 2025
Read AI-Driven Accelerated Inclusion Analysis for Energy Efficient Steelmaking (Final CRADA Report) by Castillo, Vic, Hu, Yeping, Chukwulebe, Bernard et al. (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/doe-osti-3000502-ai-driven-accelerated-inclusion-analysis.txt.
Introduction
Victor Castillo, Yeping Hu, Bernard Chukwulebe and partners close a cooperative research agreement, CRADA TC02406, between Livermore and ArcelorMittal USA Research, in LLNL-TR-2013112. Technical work ended 8 February 2024. Steel quality turns on inclusions, tiny non-metal particles trapped in the metal. Reading their microscope images by hand is slow. The team built AI tools to classify inclusions and guess their chemistry at speed.
The software, renamed ADIAS, bundles four tools. An embedding projector spreads thousands of inclusion images into a 3D map using features from a convolutional neural network. An interactive ternary plotter charts chemistry three ways at once. Two predictors then take a fresh image and return its inclusion class and its elemental make-up. All deliverables finished on time, with no property exchanged between the parties. Faster inclusion reading supports steelmaking that wastes less energy on rejected heats.
Catalog record
| CREATOR | Castillo, Vic, Hu, Yeping, Chukwulebe, Bernard et al. |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 12 KB |
| LICENSE | Public Domain |
| RIGHTS | PUBLIC DOMAIN |
| SOURCE | DOE OSTI |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Cooperative Research and Development Agreement CRADA TC02406 February 8, 2024 Dr. Victor Castillo, AI-Driven Accelerated Inclusion Analysis for Energy Efficient Steelmaking Innovations ÍIJîϙÍŘťIJôŘŜēĖŕŜϙiƯĖèô Prepared by LLNL under Contract DE-AC52-07NA27344. LLNL-TR-2013112 AI-Driven Accelerated Inclusion Analysis for Energy Efficient Steelmaking Cooperative Research and Development Agreement Final Report CRADA No. TC02406 Date Technical Work Ended: February 8, 2024 February 8, 2024 A. Parties This project was a relationship between Lawrence Livermore National Security, LLC and ArcelorMittal USA Research LLC. Lawrence Livermore National Security, LLC Lawrence Livermore National Laboratory 7000 East Avenue Livermore, CA 94550 Vic Castillo, L-153 Tel: (925) 424-5049 Email: [email protected] Yeping Hu, L-224 Email: [email protected] ArcelorMittal USA Research LLC 3001 E. Columbus Drive East Chicago, IN 46312 Bernard Chukwulebe Tel: (219) 256-7760 Email: [email protected] Dr. Tathagata Bhattacharya, Project Manager Tel: (219) 256-7687 Email: [email protected] This project is sponsored under the High-Performance Computing for Energy Innovation (“HPC4EI”) Program of the Department of Energy’s Advanced Materials & Manufacturing Technologies Office (“AMMTO”) within the Energy Efficiency and Renewable Energy (“EERE”) Office. CRADA TC02406 Page 2 Project funding table for CRADA: Year 1 Funding Type Funds-in *In-kind Totals ArcelorMittal $0 $75,000 $75,000 Dept. of Energy $300,000 $300,000 [EERE] Totals $375,000 $375,000 **Federal $0 $0 Administrative Charge (FAC) B. Project Scope This was a collaborative effort between Lawrence Livermore National Security, LLC (LLNS) as manager and operator of Lawrence Livermore National Laboratory (LLNL) and ArcelorMittal USA Research LLC (“ArcelorMittal” as the Participant), to use scanning electron microscopy (SEM) images, computer vision and machine learning methods, and high-performance computing to accelerate the inclusion analysis process of liquid steel so that new methods can be used for near-real time process control on the shop floor. The project was originally designated as a 12-month project, and consisted of four major tasks and the following four deliverables: Task Deliverable Responsible Party Due Date 1 Dataset of inclusion and EDS spectra ArcelorMittal End of Month 2 2 Final validated model package LLNL End of Month 10 3 Updated IntelliSEM code containing ArcelorMittal End of Month 12 validated model and memo documenting results of testing the updated IntelliSEM software 4 Final Report and Abstract LLNL, with input Completion or from ArcelorMittal termination date of this CRADA All of the deliverables for this project were successfully completed on time. There were no amendments or No-Cost Time Extensions associated with this project. No tangible property was exchanged. There are no outstanding liens. 2/8/24 CRADA TC02406 Page 3 C. Technical Accomplishments Tools Input Output Visualization Embedding Projector Image 3D Image Projection Interactive Ternary Plotter Element% Ternary Plot Modeling Inclusion Type Predictor Image Inclusion Class Inclusion Chemistry Predictor Image Element% A python-based software, originally known as "ArcelorMittal_Inclusion" but will be known going forward as “AI Driven Inclusion Analysis of Steel (ADIAS)” was successfully developed to achieve the goal of steel inclusion analysis. The software is equipped with multiple functionalities and the tools are summarized in the table above. The data visualization tools feature the embedding projector and the interactive ternary plotter, while the modeling tools include the inclusion type predictor and inclusion chemistry predictor. The following subsections contain detailed descriptions of each tool. C.1 Visualization Tools The embedding projector processes a large dataset of steel inclusion images, generating a 3D projection through the utilization of image features extracted from convolutional neural network (CNN) models. The results of this projection are illustrated in Figure 1. This projection can help visualize the intrinsic structure of the dataset, thereby enhancing the fundamental comprehension of the images and various inclusion classes. Various filters can be applied to facilitate the visualization of images sharing common inclusion class, project, sample, etc. In subsequent versions, additional filters can be included such as plant, specific process route, grade, and other key properties. Figure 1. 3D embedding projection of steel inclusion images. In addition to the embedding projector, the toolkit contains an interactive ternary plotter designed for visualizing the chemistry of steel inclusions across various inclusion types. As depicted in Figure 2, users can select targeting ternary elements and the plot is generated for the entire 2/8/24 CRADA TC02406 Page 4 dataset, with data points colored based on their inclusion class. In future applications, this tool holds the potential for comparing predicted elemental distribution against the actual elemental distribution. Figure 2. A ternary plot showing relative chemistry for steel inclusion images. Selected elements are Mg, Al and Ca. 2/8/24 CRADA TC02406 Page 5 C.2 Modeling Tools Figure 3. A Graphical User Interface for inclusion type predictor and inclusion chemistry predictor. The modeling tools we designed have the capability to predict both the inclusion class and inclusion chemistry. Through the utilization of High-Performance Computing (HPC) resources, we conducted a systematic exploration of various deep learning models. Among them, ResNet, a convolutional neural network, emerged as the optimal choice, demonstrating a 60.8% accuracy in inclusion classification and achieving a mean absolute error (MAE) of 2.6% for element percentage prediction. Our implementation incorporates a user-friendly Graphical User Interface, as depicted in Figure 3, allowing users to upload an inclusion image for predicting both the inclusion class and chemistry. Moreover, we introduced an innovative model based on graph neural networks (GNN), which yielded a 59.3% accuracy in inclusion classification. In future advancements, augmenting the dataset and incorporating advanced machine learning models has the potential to further improve model performance. Additionally, delving into uncertainties will offer users insights into the reliability of predictions. A notable challenge in our modeling approach is the significant diversity of classes within the dataset, where images of different inclusion types may closely resemble one another. Effectively addressing this challenge requires a thorough examination to devise an improved labeling 2/8/24 CRADA TC02406 Page 6 strategy, ultimately resulting in a more accurate and efficient tool. Another challenge lies in the imbalanced distribution of data across different classes, necessitating a detailed evaluation and careful selection of training data for the prediction models. D. Expected Economic Impact D.1 Specific Benefits Benefits to DOE This CRADA was co-sponsored under the High-Performance Computing for Manufacturing (HPC4Mfg) Program of the Department of Energy’s Advanced Materials & Manufacturing Technologies Office (AMMTO) within the Energy Efficiency and Renewable Energy (EERE) Office. LLNL was the lead Lab for the HPC4Mfg Program. This benefits the DOE by directly supporting the mission of the EERE AMMTO through its HPC4Mfg program. This CRADA was selected for this program by the DOE. This project advances the goals of the HPC4Mfg program by leading to reduced carbon emissions across the industrial steel manufacturing sector by improving efficiency and quality, saving energy, reducing waste, and ultimately reducing CO2. Also, this project allows LLNL to exercise their expertise in HPC, computer vision, and machine learning. Developing artificial intelligence (AI) methods to identify and control specific defects in metal processing is extremely important to the NNSA and AMMTO missions. Benefits to Industry This project will improve quality of steel, reduce cost for steel production, and reduce CO2 emissions. According to the Participant’s approved HPC4EI proposal, yield losses of 1-2% can be attributed to harmful inclusions/particle entrapment during slab solidification from liquid steel. These losses, which are due to defective final products, reprocessing of intermediate products and yield loss at various processing steps, serve to reduce energy efficiency and productivity by significant amounts across