With the vigorous development of computer technology, the application of AI and large models in the steel industry has become a key force in promoting the intelligent transformation of the industry. This research mainly focuses on the application of large-scale models in the steel industry. Firstly, it summarizes the construction methods and typical application areas of industrial large-scale models; Secondly, elaborate on the characteristics of the steel industry and summarize the relevant technologies of the steel industry's large-scale model; Finally, a discussion will be conducted on the application scenarios of the large model, highlighting typical application scenarios of the large model in the steel production process, such as breakthroughs in perception and cognitive tasks. In the future, the steel industry is expected to deeply apply big models to the development of new products and systems, as well as provide comprehensive decision support, realizing the application of energy, raw material scheduling, and full process monitoring of steel enterprises, promoting the high-end, intelligent, and green development of the steel industry, and providing new ideas for digital development.

With the rapid advancement of general-purpose language models such as DeepSeek, research in the industrial domain is increasingly shifting from foundational model exploration to their specialized application in vertical, domain-specific scenarios. Taking converter steelmaking as a representative application context and the technological evolution of industrial-scale foundation models was systematically reviewed, including language, vision, and temporal models. Based on this analysis, it proposes a preliminary design for a multimodal large-scale model architecture tailored to the converter steelmaking process. The proposed architecture leverages heterogeneous data sources—such as scientific literature, flame video streams, flue gas composition, and molten iron chemistry—to address key technological challenges, including complex furnace condition recognition, dynamic endpoint prediction, and real-time process control. The objective is to overcome the theoretical limitations of conventional approaches in modeling high-temperature multiphase reactions, multivariate coupling behavior, and real-time decision-making under complex operating conditions. Furthermore, this study analyzes the core challenges encountered in the deployment of large-scale models in metallurgical applications—such as data heterogeneity, domain adaptation, and model interpretability—and discusses corresponding mitigation strategies. Finally, future development directions are explored to provide a theoretical foundation and methodological reference for the intelligent transformation of the steelmaking industry.

In the current era when the wave of intelligence is sweeping the world, the steel industry is confronted with an urgent need for transformation and upgrading. As a core production factor in the new era, the extraction and utilization of data value are of great significance for enhancing production efficiency, controlling product quality, and promoting digital development in the steel industry. Although the steel industry is rich in data resources, it is currently facing a deep-seated predicament in the release of data value. Aiming at the common problems such as "insufficient data and poor-quality data" faced in the application and development of large model technology in the steel industry, this paper designs a set of data engines that meet the data requirements of large models, based on three technologies: data enhancement technology of time series large models, data spatio-temporal alignment technology of process mechanisms, and cross-process data association of knowledge graphs. Establish a twin data asset shell file with materials as the object, combine the semantic knowledge of metallurgical production processes with the time series data of the production process, provide corpus information of the entire product life cycle for industry large models, and meet the demand for semantic data in the training and application of large models. Meanwhile, the traditional offline data governance mode was transformed into an online mode by using the workflow mode. Taking the data of a continuous casting production line in a certain steel plant as a pilot, significant improvements were observed. The abnormal rate of data within batches was reduced to 0.02%-0.5%, facilitating the digital transformation of steel enterprises.

As materials science research enters the fourth paradigm , artificial intelligence technologies are reshaping research  in  this  field.   currently ,  large  language  models  ( LLMs) ,  trained  on  massive datasets and utilizing extensive parameter scales , have overcome the technical limitations of traditional machine learning in text processing and human-computer interaction through capabilities such as multi - task integration ,  context-aware  generation ,  and  decision-making .    These  advancements  have  opened new  avenues for the intelligent transformation of the steel industry .  The technical features , research di- rections and application of LLMs were systematically summarized.  considering the data characteristics of the steel industry , a detailed overview of frameworks for large language models tailored to steel ap-plications was provided , along with proposed evaluation criteria specific to the industry .  The potential applications of LLMs  in  steel were  explored  in  depth , including  data  extraction  and  protection , per- formance modeling and prediction , inverse design of materials , and intelligent steelmaking .  Addition- ally , a comprehensive analysis of the  current development of LLMs in the steel industry was presented , key challenges were identified , and corresponding strate gies were proposed.

As a vital pillar of the national economy, the iron and steel industry faces challenges such as high costs, low efficiency, and complex processes, which urgently require intelligent transformation to break through development bottlenecks. With its cross-modal understanding and multi-scenario generalization capabilities, large model technology provides a new path for deep data value mining and optimization of production processes in the steel industry. This research proposes an overall construction framework of the "Five-in-One" large model platform for the steel industry and elaborates on ten key technologies required to build a large model platform adapted to the industry needs. Through the deep coordination of platform, computing power, data, model, and scenario, and the integration of general models with steel industry knowledge, this framework forms an integrated "AI + Steel" solution, which significantly reduces the research and development threshold and cost of models in the steel industry and enables AI technology to better integrate into the industry. Finally, this research combines the business characteristics and practical exploration experience of the steel industry and forecasts the application scenarios, providing insights for the industry's intelligent transformation.

With the in-depth advancement of the digital transformation in the steel manufacturing industry, intelligent quality management has become one of the key links to enhance the core competitiveness of enterprises. Aiming at the product quality defect analysis process of steel enterprises, an architecture for a quality defect analysis system based on a retrieval-augmented large language model was proposed. The system adopts a modular design to build a case knowledge base, implements a two-stage knowledge retrieval enhancement architecture that combines semantic vector recall and confidence re-ranking, and designs a hierarchical information extraction and aggregation mechanism for local causal extraction and global causal aggregation. It realizes similar case matching and cross-case knowledge collaborative reasoning, while embedding an interpretability support mechanism and a dynamic knowledge update mechanism to ensure the reliability of the analysis results and the long-term effectiveness of the system. The system has been deployed and run in a steel enterprise, effectively improving the efficiency of the product quality defect handling process and providing a successful case for the application of large models in the field of industrial product quality management.

As a key index to reflect the stability of furnace condition and operation smoothness, the accurate prediction of blast furnace gas permeability index is of great significance to recognize and prevent abnormal furnace condition in time. A method for predicting the gas permeability index of blast furnace based on generative time series model was proposed. Firstly, the box plot method was used to identify and correct the anomalies in the data, and secondly, an algorithm based on the fusion of Boruta and SHAP (shapley additive explanations) of random forest was proposed to select the features of the permeability index. Finally, a time series generation (TSG-GPT) model was designed to predict the gas permeability index of the blast furnace. The model was trained, validated and tested using actual production data, and the results show that the proposed model can accurately predict the gas permeability index.

In the context of new industrialization, digital and intelligent transformation has become a core pathway for the steel industry to break through efficiency bottlenecks and achieve high-quality development. Meanwhile, large language models, as a rapidly advancing artificial intelligence technology in recent years, have demonstrated remarkable results across various sectors. Therefore, how to leverage general-purpose large models to accelerate the intelligent transformation of the steel industry has emerged as a significant research and application direction. This paper focuses on exploring three mainstream methods for constructing large models in the steel domain: Domain-continual pretraining, which enhances the model’s understanding of industry-specific knowledge through continuous pretraining on steel-related corpora; supervised fine-tuning on scenario-specific tasks, which improves model performance on domain-specific tasks using labeled data from real-world industrial applications; and knowledge distillation, which leverages a powerful teacher model to extract knowledge from steel industry corpora and transfer it to a smaller student model during training, thereby enhancing the model’s retention of domain-specific knowledge. Finally, the advantages and disadvantages of these three methods were evaluated using a constructed steel knowledge assessment dataset. Through systematic analysis and comprehensive comparison, this research provides practical guidance and methodological references for building specialized large language models in the steel industry and other vertical domains, thereby facilitating industrial intelligent upgrading.