Digital and intelligent technologies are driving the intelligent transformation of blast furnace ironmaking in China, emerging as a kind of new quality productive forces. Nowadays, using the universal large language models （U-LLMs） as a base framework to build industrial vertical large language models （V-LLMs） for guiding industrial production via secondary training with domain-specific corpora has become a new trend. Despite the emergence of V-LLMs for the entire steelmaking process, research on V-LLMs specifically for blast-furnace operations is still in its infancy. By reviewing the recent evolution of blast furnace ironmaking intelligence, a new idea of its paradigm reconstruction and fusion driven by V-LLMs was proposed. Classifying the blast furnace V-LLM tasks scenario into scheduling and decision making, the "Data-Application-Perception" penetration and application path is presented for the first time. Also, 5-dimensional evaluation system for future performance assessment and optimization is puts forward, including process understanding, safety and reliability, knowledge transfer, real-time performance, and continuous learning. Then, 3 kinds of new paradigms for intelligent upgrade driven by blast furnace V-LLMs are discussed, including blast furnace condition characterization, blast furnace condition metaverse, and multi-scene fusion. A 3-dimensional collaborative deep representation architecture of "Physical↔Virtual↔Perception" with blast furnace V-LLMs as the core and a new concept of “blast furnace portrait” are proposed. The construction route of blast furnace condition metaverse and the policy of multi-scene fusion are sorted out and discussed. Finally, the key issues and possible solutions in the future development of blast furnace V-LLMs are analyzed. Focused on the feasibility of blast furnace V-LLMs in construction, application, and evaluation, the paradigm reconstruction of intelligence update driven by V-LLMs in the context of industry development is discussed. It aims to offer theoretical guidance for the deep application of V-LLMs in China's blast furnace ironmaking and further promote its intelligent transformation and development.

Guided by national industrial policies, Hesteel Group actively promotes industrial restructuring and transformation and upgrading, and is committed to building Tangsteel New District into a green, intelligent and efficient new generation process steel mill. Since the planning, design, construction and operation of the project, we have earnestly practiced the theory of "metallurgical process engineering", integrated a series of new processes and technologies, and effectively improved the process technology and equipment level of each process. Especially in the iron front system, in the process of raw material storage and transportation, the integrated design of port and storage yard is adopted to build an intelligent unmanned enclosed environmental protection material yard, which realizes the green storage and transportation of raw fuel and efficient deployment; The coking process adopts large coke oven with wide carbonization chamber and low carbon environmental protection coking technology, and the process technology equipment and energy efficiency level reach the domestic leading level; The process technology and equipment of domestic large-scale belt roasting machine have been developed, and the production technology of medium and high silicon flux pellets has been developed under the condition of complex fine powder resources. Sintering process adopts thick material layer, high proportionof flue gas circulation, annular cooling waste heat step recovery and intelligent sintering technology, which significantly improves the production efficiency and energy saving and emission reduction effect of sintering process. The process of blast furnace adopts 40%-60% high proportion pellet smelting, and the intelligent and visual operation technology is developed and applied to realize the long-term efficient and stable operation of blast furnace. In addition, through the deep integration of production scheduling system and ironmaking MES system, a whole-process integrated production control system has been built, creating conditions for efficient coordination and organization of iron pre-production. The integrated application of new processes and technologies in the iron front system of Tangsteel New District has significantly improved the level of production efficiency, energy efficiency utilization and environmental protection control, provided strong support for the sustainable development of Tangsteel, and set a typical example for the transformation and upgrading of steel enterprises and high-quality development.

Ammonia is a high hydrogen energy carrier and can be used as a highly efficient reducing gas in the iron and steel industry, with the advantages of easy liquefaction and convenient storage and transportation methods compared to hydrogen. Therefore, it proposes the idea of "ammonia instead of hydrogen" for direct reduction of iron ore powder, and combined with thermodynamic analysis, the direct reduction of magnetite by NH₃ at 900-1 100 ℃ is carried out. The variation of NH₃ reduction effect and the evolution of microscopic morphology and pore structure of reduction products under high temperature conditions were analyzed, and the reduction effect of H₂ was compared with that under the same conditions. The results show that compared with the thermodynamic results of H₂ reduction, Fe₃O₄ can be reduced to Fe under the same temperature with a lower relative partial pressure of NH₃, indicating that NH₃ has a certain reduction advantage. At 900 ℃, the TFe content of NH₃ reduction products was only about 4% different from that of H₂ reduction, indicating that NH₃ had the same reducing ability as H₂, while NH₃ reduction products were affected by the presence of iron nitride, resulting in a difference of about 14% in metallization rate compared with H₂ reduction. By increasing the temperature, the iron nitride in the product can be decomposed to achieve the effect of denitrogenation, and at 1 000-1 100 ℃, the reduction metallization rate of NH₃ is generally higher than that of H₂ reduction, and the metallization rate is as high as 98%. The microscopic morphology results showed that compared with H₂ reduction, the blind pores on the surface of NH₃ reduction products were more densely distributed, and the internal cross-linked pores were more complex, and the internal dendritic connection structure was conducive to the diffusion of NH₃ into the concentrate, and with the extension of time, the NH₃ reduction products showed a trend of transformation from honeycomb iron to dense iron. During the reduction process of NH₃, the proportion of microporous pore volume decreased by 11.79% and the fractal dimension decreased by 0.139, indicating that the surface of the sample changed from rough to smooth during the reduction process, and there was an obvious dynamic change in the pore structure. It proves that NH₃ reduction is more advantageous under high temperature conditions, and the complex pores and dynamically changing pore structure in the reduction process provide a diffusion channel for NH₃ to enter the concentrate. The results provide empirical data for the direct reduction ironmaking technology of "ammonia instead of hydrogen", and have certain significance for the sustainable development of low-carbon metallurgy.

The foundation of this study is established based on the actual raw material and fuel conditions of the blast furnace in a specific steel plant. It combines laboratory experiments, mathematical model construction, and production data analysis to experimentally investigate the effect of different types of scrap （including particle steel and slag steel particles） on the softening-melting-dripping property of the mixed burden. By establishing a material-heat balance mathematical model for the use of scrap in the blast furnace, the impact of scrap addition on fuel consumption was analyzed, and the results were compared and validated with actual production data from the steel plant's blast furnace. The softening-melting-dripping experimental results show that when 5.0%（mass fraction）scrap （including particle steel and slag steel particles） is added to the mixed burden, the composition and melting point of the dripping iron do not change significantly, but the FeO content in the dripping slag increases, the thickness of the cohesive zone narrows, and the permeability is improved dramatically. Among them, particle steel has a more pronounced effect on narrowing the cohesive zone and improving permeability than slag steel particles. When the scrap addition exceeds 5.0%（mass fraction）, the melting point of the dripping iron tends to increase, the thickness of the cohesive zone increases, and the permeability deteriorates. The material-heat balance model calculations show that when the scrap addition is 100 kg/t, the grade of the mixed burden increases by 2.68%, the ore ratio decreases by 171 kg/t, the coke ratio decreases by 38.5 kg/t, the slag ratio decreases by 32.94 kg/t, and the direct reduction rate slightly increases to 5.74%. Actual production data validation indicates that at this addition level, the daily output of the blast furnace increases by 600 t/d, the utilization coefficient improves by 6.6%, and the fuel ratio is reduced by 30 kg/t, while the gas utilization rate experiences a slight decline of approximately 1%. Furthermore, when the scrap addition is below 5.0%（mass fraction）, the permeability of the blast furnace can be significantly improved, but as the addition continues to increase, the permeability deteriorates. Comparative analysis shows good consistency between the laboratory experimental data, mathematical model calculation results, and production data. The research results provide a theoretical basis and technical support for the rational use of scrap in blast furnaces, offering important guidance for optimizing blast furnace smelting processes and reducing production costs.

Driven by the "the twin goals of carbon peak and carbon neutrality", the low-carbon innovation of blast furnace ironmaking process has become the core of iron and steel industry transformation. The path of oxygen/hydrogen enrichment in the blast furnace ironmaking process is focused on, and theoretically analyse the upper limit of oxygen enrichment and the working window of the three processes, namely, traditional blast furnace, injected circulating gas and injected coke oven gas, as well as the influence of oxygen enrichment rate, circulating gas and coke oven gas on the low-temperature zone thermal equilibrium, theoretical combustion temperature, coke ratio, and the amount of gas in the belly of the furnace by means of the thermal mass balance and the key constraint quantity model. The results show that, under the condition of 150 kg/t coal ratio, the oxygen enrichment rate threshold should be controlled within 7.9% to maintain the theoretical combustion temperature （t_f） at （2 150±50） ℃, and to ensure that the heat flux in the low-temperature zone is not less than 2.8 GJ/t, and with the use of the injected recycled gas process, the upper limit of the oxygen enrichment rate jumped to 52.23%, and the coke ratio was significantly reduced to 207.44 kg/t, which is 18.6% lower than the baseline condition. In the case of the injected coke oven gas process, the upper limit of the oxygen enrichment rate reaches 35.67% and the lowest value of the coke ratio is reduced to 183.91 kg/t, which is 11.3% lower than that of the recycled gas process, showing a better potential for carbon emission reduction under the low coal ratio condition. It was further found that, within the stable control range of tf, every 1% increase of oxygen enrichment rate, the amount of circulating gas at the air outlet can increase by 9.15 m³/t, and the maximum blowing amount reaches 458.26 m³/t, while the amount of coke oven gas blowing increases non-linearly with the oxygen enrichment rate, with the upper limit value of 293.27 m³/t. When the oxygen enrichment rate increases from 5% to 35%, the heat demand decreases by 22.4%, while the heat supply decreases by 31.7%, and this supply-demand imbalance puts forward new regulation requirements for the thermal management of the blast furnace. The triple synergistic effect of the coke oven gas injection process is revealed, the expansion of the process window through the increase of oxygen enrichment, the reduction of coke ratio is optimized by means of the H₂/CO （mass fraction ratio）, and the use of gas reforming to regulate the heat balance. It provides theoretical support for the construction of a new type of "oxygen-rich-hydrogen-rich-low coke ratio" blast furnace ironmaking system, which is of great value for promoting the low-carbon transformation of the iron and steel industry.

Data-driven approaches have achieved notable success in predicting the silicon content of hot metal in blast furnaces. Due to the complexity of the blast furnace, blast furnace data possesses strong coupling, large time delays, and multi-timescale characteristics, which make model training more challenging. It should be continuing researched and paid particular attention. A time windowing technique combined with principal component analysis （PCA） was employed to transform 22 minute-level features into heat-level parameters. A sliding window approach and the maximal information coefficient （C_imax） were then used to determine the optimal lag time for each parameter. To evaluate parameter importance under strong coupling conditions, a SHAP algorithm optimized with a random forest model was applied to 34 parameters, identifying 7 key influencing factors. Furthermore, SHAP was utilized to refine the self-attention mechanism of the transformer model, resulting in the development of the SHAP-Transformer model for predicting hot metal silicon content. The validity of the proposed model was checked using industrial data. The results showed that the SHAP-Transformer model, incorporating sliding window-based lag analysis and optimal coupled parameter selection, achieves the highest prediction accuracy. The hit rates of the prediction model within error ranges of -0.05-0.05 and -0.1-0.1 were 72.12% and 95.76%, respectively, surpassing the SHAP-Transformer model based on MIC-selected parameters by 26.67% and 21.21%, and outperforming the long short-term memory （LSTM） model with sliding window-based lag analysis and optimal coupled parameter selection by 17.57% and 9.7%. Moreover, the proposed SHAP-Transformer model exhibits high accuracy in forecasting silicon content trends, achieving 60.5% accuracy in trend classification and 87.3% accuracy in trend direction prediction. The model can provide a reliable basis for blast furnace operators to forecast temperature fluctuations and optimize furnace operation proactively.

The MR steel produced by basic oxygen furnace （BOF）→ ladle furnace （LF）→ continuous casting （CC） process in a domestic steel plant was taken as the research object. The whole process sampling was carried out at arriving at LF station, after slag melting, before calcium treatment, after calcium treatment, soft blowing for 3 min and 9 min, leaving LF station, and tundish. The total oxygen （T[O]） content, inclusion number density and other indicators were detected, while the evolution law of inclusions and its relationship with T[O] content were revealed. At the same time, the calcium treatment process was optimized by thermodynamic calculation. The results show that the content of T[O] continues to decrease from arriving at LF station to before calcium treatment, and the inclusions in molten steel are mainly deoxidized product Al₂O₃. After calcium treatment, the content of T[O] increases, while a part of Al₂O₃ in molten steel is modified into CaO-Al₂O₃. The content of T[O] continues to increase after the start of soft blowing, and the mass fraction of T[O] reaches 0.004 25% at soft blowing 9 min, indicating that the secondary oxidation of molten steel occurs during both calcium treatment and soft blowing. Al₂O₃-CaS and CaO-Al₂O₃-CaS inclusions begin to appear in the molten steel after soft blowing. From leaving LF station to tundish continuous casting stage, T[O] content shows a downward trend, and inclusions are mainly Al₂O₃ type （referring to Al₂O₃ and Al₂O₃-CaS） and CaO-Al₂O₃ type（referring to CaO-Al₂O₃ and CaO-Al₂O₃-CaS ）. The microscopic inclusions in the slab mainly include deoxidation product Al₂O₃, calcium treatment product calcium aluminate and their derivatives, while the macroscopic inclusions mainly origin from mold flux and Ti-containing particles. The evolution path of inclusions in the smelting process is Al₂O₃ → Al₂O₃+CaO-Al₂O₃ → Al₂O₃ type+CaO-Al₂O₃ type. The total number density, average size, total area, average area and other parameters of inclusions are consistent with the change trend of T[O] content in the process of arriving at LF station to leaving LF station, but they donot present so in the continuous casting process, which is related to the precipitation of CaS, TiN and other substances. For the calcium treatment process, thermodynamic calculations show that the liquid window of calcium treatment has no obvious correspondence with the（w（T[Ca]）/ w（T[Al]）） and w（T[Al]）/ w（T[O]） in molten steel, but is closely related to the w（T[Ca]） / w（T[O]） . When the w（T[Al]）） is 0.046%, the w（T[Ca]） / w（T[O]） range of the liquid window is 0.52-0.96. The inclusions formed under the current molten steel composition are located below the liquid window. The goal of theoretically modifying Al₂O₃ into liquid calcium aluminate can be achieved by increasing 0.000 94% w（T[Ca]） or controlling the w（T[O]） within 0.002 3%. In addition, CaS in the slab is mainly precipitated during the solidification of molten steel.

In order to improve the identification speed and accuracy of argon blowing grade detection in VD refining non-vacuum process, an improved YOLOv8n model for argon blowing grade detection in VD refining non-vacuum process was proposed. First, based on the YOLOv8n model, Coordinate Attention was introduced from the backbone network layer, and the characteristics of the attention mechanism was used to increase the recognition accuracy of the basic model. In order to increase the training speed and recognition speed of the origin model. Furthermore, the FasterBlock module inside the FasterNet was used to replace the Bottleneck processing module inside the feature extraction module C2f in the backbone network and neck network of the basic model, in order to increase the training speed and recognition speed of the original basic model. Secondly, since there was no relevant training weight for the recognition of argon blowing grade for ladle refining, the collected image set of ladle refining was marked and divided by annotation tools and video conversion tools. Then, according to the results of different training weights, the comparison test and ablation test of different improved models and basic models were carried out. The test results show that the improved YOLO8n model recognition accuracy （precision） is increased by 11 percent point, the average accuracy （P_mA） is increased by 6 percent point, and the image processing speed per second （F_PS） is increased by 6.4 f/s. This indicates that the improved model has significantly improved the recognition accuracy and speed. Finally, a large number of identification tests were carried out on the image set of the initial stage for VD refining-slag breaking stage-line feeding stage. The results show that the improved model can meet the real-time argon blowing grade detection in VD production process, and provide an effective intelligent detection model for standardized flow control under non-vacuum conditions in VD refining.

As the most common breakout accident, sticker breakout will not only damage the continuous casting equipment but also threaten the safety of operators. The conventional breakout prediction model mainly depends on the threshold value of process parameters for judgment and simple statistical analysis and does not make full use of the time series changes of data, which limits the accuracy of the model. To solve the above problems, a WOA（whale optimization algorithm）-LSTM（long short term memory） breakout prediction model based on deep learning is constructed by combining the whale optimization algorithm and the long short-term memory neural network. The temperature features, static geometric features, and dynamic features were extracted, and the Pearson correlation coefficient was used to screen out the characteristic parameters with high correlation with breakout accidents, including 11 features such as the average and maximum temperature change rate in the abnormal temperature rise and temperature drop regions. The whale optimization algorithm is used to optimize the hyperparameters of the long-term and short-term memory neural networks. The mean square error is used as the loss function of the model, and the optimal network hyperparameters are searched through cyclic iteration. In the process of model training, the sliding window technology is used to input training samples, so that the model can better learn and capture the time-series characteristics of process parameters in the continuous casting process. Finally, the actual production data of a steel plant was used for the test. Compared with BP, LSTM, and WOA-BP models, the WOA-LSTM prediction model performs well in multiple performance indicators, can more accurately capture the time-series change trend of characteristic data, and the model has fast convergence speed and high prediction accuracy. The prediction rate of the model is 98.4% and the prediction rate is 96.8%, which can meet the requirements of the actual production of the steel plant.

During the continuous casting process, the interaction between cerium-containing steel and mold flux significantly alters the metallurgical characteristics of the flux. This interaction may disrupt the smooth operation of continuous casting and affect the stability of slab quality. To gain deeper insights into the mechanism of Ce₂O₃ influence on SiO₂-based mold flux properties the rotating cylinder method was employed to measure viscosity variations in SiO₂-based mold fluxes with different Ce₂O₃ contents. Based on the measurements, the break temperature and viscous flow activation energy were analyzed. Additionally, Raman spectroscopy and X-ray diffraction were utilized to investigate structural characteristics of the slag. The results demonstrate that at temperatures above 1 673 K, the viscosity of the mold flux decreases with increase of Ce₂O₃ content, whereas below 1 573 K, viscosity increases with the increase of Ce₂O₃ concentrations. When Ce₂O₃ mass fraction rises from 0 to 15%, the viscosity at 1 573 K increases from 0.29 Pa·s to 0.35 Pa·s, the break temperature elevates from 1 454 K to 1 522 K, and the viscous flow activation energy increases from 130.3 kJ/mol to 157.9 kJ/mol. Raman spectroscopy reveals a shift in the main peak from 906 cm^-1 to 867 cm^-1 with increasing Ce₂O₃ content, indicating that Ce₂O₃ acting as a basic oxide can dissociate into Ce³⁺ and O^2-. The liberated O^2- disrupts the silicate tetrahedral structure in the melt, simplifying the mold flux network structure and thereby reducing viscous flow resistance, which macroscopically manifests as decreased high-temperature viscosity. However, when as the content of Ce₂O₃ in the mold flux increases, the low-melting-point phase （Ca₄Si₂O₂F） transforms into the high-melting-point phase [Ce_9.33（SiO₄）₆O₂], enhancing crystallization tendency and degrading lubrication performance, resulting in elevated low-temperature viscosity. Therefore, to maintain stable flux performance, Ce₂O₃ mass fraction in the mold flux should not exceed 15%. These findings provide crucial technical guidance for continuous casting production of cerium-containing steels.

Continuous manufacturing technology represents a pivotal direction in the development of hot-rolled strip production. By highly integrating traditionally separate processes （such as continuous casting, heating, and rolling）, it enables seamless production from molten steel to finished coils. This technology serves as a critical pathway for achieving low-carbon, high-efficiency, and high-quality development in the steel industry, embodying the essence of new quality productive forces. The process characteristics of MCCR （multi-mode continuous casting and rolling） production lines was systematically examined, key technological challenges in achieving fully continuous manufacturing were identified, and a comprehensive solution package grounded in industrial production practices was developed. Firstly, addressing the requirements of high-throughput continuous casting, an integrated technological system was developed incorporating ultra-high-speed casting, high-purity molten steel control, and flow-field stabilization, thus effectively breaking through the throughput-capacity and quality limitations inherent in conventional processes. Secondly, three core longevity technologies were simultaneously established for refractories, molds, and rolls, significantly reducing equipment wear while extending operational duration, thereby ensuring stable long-term operation of the continuous manufacturing line. In addition, the implementation of a tunnel-type soaking furnace-based multi-mode production system achieved flexible transitions between single-slab, semi-continuous and fully continuous modes, complemented by advanced online dynamic control systems for thickness and width regulation, effectively solving the problem of insufficient production flexibility of the continuous manufacturing line. Furthermore, a high-precision manufacturing system was created through the integration of profile control, stability enhancement, and advanced material processing, ultimately enabling the production of full-range high-strength steels with exceptional dimensional accuracy. It is concludes with an exploration of future development trajectories for continuous manufacturing technology, which involves deep integration of green/low-carbon solutions, high-quality production systems, and smart manufacturing enablers to drive the transformation and upgrading of the steel industry.

ESP rolling is a leading steel manufacturing process, which is regarded as an important technological breakthrough in the steel industry after oxygen converter steelmaking and continuous casting. Its headless rolling production line has significant advantages in improving production efficiency, reducing energy consumption and improving product quality. In order to further improve the efficiency of ESP headless rolling production line, it is very necessary to apply intelligent optimization scheduling and scheduling technology to ESP production. Aiming to address the batch scheduling and production scheduling problem of the ESP headless rolling production line in a steel plant, and considering factors such as production process regulations and order delivery times, a multi-objective optimization model for the ESP headless rolling production plan was established. The model uses thickness and hardness as the jump penalty function, with the optimization objectives being the minimization of both the jump penalty and the number of production units. Combined with the actual situation of the production line and the delivery date of the order, the order was divided into different rolling priorities, and then according to the production unit balance, the clustering algorithm was used to merge the same kind of steel in the lower priority, and finally the genetic-firefly hybrid optimization algorithm was developed and implemented. The hybrid algorithm cleverly combineed the powerful global search ability of genetic algorithm and the fine characteristics of firefly algorithm in local search, aiming to accurately and efficiently solve the production scheduling of the merged order, thereby realizing the multi-variety and multi-specification intelligent production scheduling of cross-process casting-rolling integration. Finally, the actual production data of the steel mill was used to verify the scheduling, and the optimization results of the traditional genetic algorithm, firefly algorithm and genetic-firefly hybrid algorithm were compared and analyzed, which confirmed the accuracy and effectiveness of the genetic-firefly hybrid algorithm optimization model proposed. The experimental results show that the hybrid algorithm can effectively reduce the jump penalty, reduce the production cost, and bring significant economic benefits to the steel mill under the premise of meeting the order and delivering on time.

Abnormal vibration in rolling mills not only degrades the surface quality of sheets and strips, accelerates roll wear, and reduces production efficiency, but also leads to issues such as steel stacking and strip breaking, thereby posing a significant threat to on-site production safety. However, in the currently widely used concentrated mass vibration model for rolling mill vibration research, the stiffness and damping of the rolling mill system cannot be precisely quantified and can only be approximated, which has large errors. This limitation makes it challenging for models based on this method to accurately describe the actual rolling process. To address the difficulty of suppressing vibrations during the rolling process, a non-model-based deep reinforcement learning control strategy was proposed targeting the strong nonlinearity of rolling mill vibrations and the challenges associated with establishing an accurate rolling mill system model. The deep deterministic policy gradient （DDPG） algorithm was adopted as the control strategy to design the controller, leveraging interactive learning based on the inherent characteristics of the algorithm. A dataset was obtained through signal processing, the network structure was saved after training, and an active controller was subsequently developed. By connecting the active controller in parallel with the original closed-loop control system within the automatic gauge control （AGC） system of the rolling mill, the control signal was derived from the interaction between the agent and the rolling mill state environment. The AGC system outputed the inertial force generated by external main power balance vibrations, and active control of vertical vibrations in the rolling mill was achieved through the negative work of the external main power and system damping. The feasibility of this method had been validated through simulation analysis. The results demonstrate that the root-mean-square （R_MS） reduction rate of the vibration signal after control reaches up to 80%, indicating that the active vibration control strategy based on DDPG exhibits superior vibration suppression capabilities. Additionally, this method features a simple parameter setting process and fast convergence speed, meeting control requirements and providing a novel approach for vibration control in complex industrial scenarios.

Galvannealed （GA） strip is widely used in automotive manufacturing due to its exceptional corrosion resistance and formability, with surface roughness being a critical quality indicator affecting coating adhesion and product aesthetics. To address the challenges of multi-process coupling effects and the insufficient accuracy of traditional single-point prediction in GA strip production, a galvannealing line was focused on. By analyzing the roughness evolution mechanisms across three stages, cold-rolled substrate, galvanizing, and temper rolling, and key influencing factors were identified, alloying temperature, alloying speed, Fe content in the coating layer （galvanizing stage）, and strip specifications/rolling parameters （temper rolling stage）. A machine learning prediction framework based on multi-objective independent modeling was developed, incorporating an optimized objective function for multi-target prediction accuracy, Bayesian hyperparameter optimization, and five-fold cross-validation. Using denoised industrial measurement data （6 551 records）, the performance of XGBoost （eXtreme gradient boosting）and CatBoost（categorical boosting） algorithms was compared, demonstrating that XGBoost outperformed the others, achieving an average coefficient of determination （R²） exceeding 0.94 at three measurement points on the upper and lower surfaces with root mean square error （R_MSE） below 0.046. Additionally, XGBoost exhibited 6-10 times faster training efficiency than the equally high-performance CatBoost, with average prediction accuracies over 96% for both surfaces. The proposed framework combined with XGBoost effectively enables multi-objective roughness prediction across the strip width direction, meeting industrial precision requirements and showing potential for extension to similar production lines and multi-target quality prediction scenarios.

With the rapid advancement of electronic information technology, the requirements for the comprehensive performance of ultra-thin strips have become increasingly stringent. In particular, controlling the surface topography of ultra-thin stainless steel strips has emerged as a critical technical challenge. To address this issue, single-pass rolling experiments were conducted on 0.08 mm thick 304 stainless steel ultra-thin strips using three types of rolls with distinct surface treatments （ground, shot-blasted, and polished） on a 12-high precision ultra-thin strip rolling mill. A tailored surface roughness transfer mechanism model （RTM） for ultra-thin strips was developed, with key parameters in the mechanistic model optimized through genetic algorithms （GA）. This investigation elucidated the influence laws of reduction ratio, tension, roll roughness, and initial strip roughness on surface roughness during rolling. Furthermore, a surface roughness prediction model constrained by the rolling transfer mechanism was proposed via the deep integration of the mechanistic model with machine learning techniques. Specifically, the deviation between the mechanistic model's predictions and actual values was extracted as input for machine learning, leveraging its nonlinear fitting capability to capture complex nonlinear features unexplained by traditional mechanistic models. Accurate surface roughness prediction was achieved through deviation correction. Experimental results demonstrate that the fused model effectively integrates the strengths of both approaches, taking the roughness average （R_a） index as an example, the optimal model achieves a prediction accuracy of 95.08% and a correlation coefficient of 0.934 7. Industrial field verification using production data confirms that prediction accuracy remains above 90% even under complex conditions. This model combines the efficient prediction performance of machine learning with the physical interpretability of rolling mechanisms, providing a novel direction for surface quality control of ultra-thin strips. It holds significant engineering value for further exploration of surface roughness formation mechanisms and process parameter optimization in ultra-thin strip production.

FeNi alloy has been widely used in transformers, generators, and magnetic shielding due to their excellent magnetic properties. Previous studies have shown that hot rolling and subsequent annealing can enhance the magnetic properties of FeNi alloy. However, the mechanism by which hot rolling influences the microstructure to improve magnetic properties remains unclear. Therefore, taking FeNi alloy as the object, the magnetic properties of FeNi alloy were optimized by adjusting the hot rolling and subsequent annealing process, and the magnetic properties of FeNi alloy comparable to literature reports were obtained under similar compositions.The microstructure evolution of FeNi alloy （such as grain size, high/low angle grain boundaries, dislocation density, etc.） and its impact on the alloy magnetic properties of FeNi alloy under different hot rolling temperatures （800, 1 000 ℃） and hot rolling deformation （30%, 50%） were studied, aiming to explore the mechanism of influence of hot rolling process on the magnetic properties of FeNi alloy. The results show that when the temperature of hot rolling is kept constant, the deformation of hot rolling increases from 30% to 50%, the permeability decreases, and the coercivity increases. This is because the increase of hot rolling deformation leads to grain refinement and dislocation density increasing, hinders the deflection of magnetic domains, and deteriorates the permeability and coercivity. Conversely, when the deformation is constant and the temperature increased from 800 ℃ to 1 000 ℃, the permeability of the FeNi alloy increase and coercivity decrease, as higher temperatures promote dislocation recovery and recrystallization, reducing overall dislocation density. It is also found that the storage of dislocations near low-angle grain boundaries during hot rolling is detrimental to magnetic performance, whereas higher hot rolling temperatures reduce low-angle grain boundaries, decreasing local dislocation accumulation and improving magnetic performance. Subsequently, after annealing at 1 100 ℃ for 4 h, the grain size of FeNi alloy increases, the dislocation density decreases significantly, and the magnetic properties are further improved. The result provides some guidance for optimizing the magnetic properties of FeNi alloy through hot rolling processes.

Austenitic stainless steel is one of the most widely used and produced stainless steels, playing a crucial role in various fields such as petrochemical, nuclear power, and food industries. However, the lower yield strength limits its broader applications. In recent years, the research and development of high-strength austenitic stainless steel has become one of the hot topics in the field of stainless steel. A new austenitic stainless steel with high strength and excellent ductility was prepared through processes such as vacuum smelting, cryogenic rolling, annealing, and aging heat treatment. The yield strength, tensile strength and elongation after fracture of the new austenitic steel were 676 MPa, 1 011 MPa and larger than 50%, respectively. Aging treatment plays an important role in improving strength, but the effect and mechanism of it on corrosion resistance are currently unclear. Therefore, the influence of aging treatment at 550 ℃ on the pitting corrosion performance of the new austenitic stainless steel by the methods of SEM/EBSD, XRD, XPS, TEM, and electrochemical test was investigated. The results show that the pitting potential （E_p） of the aged specimens is lower than that of the unageing one, moreover the E_p and polarization resistance （R_p） of the specimens gradually decreased with aging time, so the pitting resistance slowly declined. However, the minimum E_p value is also above 0.15 V, and the effect of aging treatment on the self-corrosion potential and self-corrosion current density is not significant. In addition, aging treatment reduces the relative content of Cr₂₃C₆ in the passivation film of specimens, while, there are negligible changes observed in the relative contents of NiO and MoO₃. This phenomenon can primarily be attributed to the precipitation and growth of secondary phases such as Cr₂₃C₆ and NbN during the aging process which induce localized heterogeneity in Cr and N element distribution within the specimens. This results in microelectrochemical heterogeneity that leads to the decrease of pitting potential alongside weakened re-passivation ability. Aging treatment does not affect the factors such as average grain size and matrix phases that influence the pitting corrosion of austenitic stainless steels. This provides a possible way for the development of advanced austenitic stainless steels with excellent comprehensive properties.

Hydrogen embrittlement is a critical factor in the service failure of high-strength steels exposed to hydrogen environments, with its mechanism being closely related to strain rate. The effects of hydrogen on the dynamic mechanical properties and microstructure of 18CrNiMo7-6 alloy steel were systematically investigated under high strain rates （400-2 700 s^-1） through split Hopkinson pressure bar （SHPB） experiments combined with electrochemical hydrogen charging technology, and the underlying mechanisms were revealed. Experimental results demonstrate that 18CrNiMo7-6 alloy steel exhibits significant strain rate strengthening effects at high strain rates. When the strain rate increases from 400 s^-1 to 2 700 s^-1, the yield strength increases by 107.72% and the plastic flow strain enhances by 387.88%.Microstructural analysis indicates that high strain rates promote the formation of low-angle grain boundaries, grain refinement, and significant increase in geometrically necessary dislocation （GND）density. The combined action of low-angle grain boundaries, refined grains, and GNDs creates a multi-scale obstacle network that substantially increases resistance to dislocation motion, resulting in the observed strain rate strengthening effect. The hydrogen effect on mechanical properties shows strain rate dependence. Hydrogen enhances dynamic yield strength but causes significant plastic loss at 400-1 100 s^-1, while reducing strength with negligible plastic loss at 1 900-2 700 s^-1. EBSD analysis reveals hydrogen's interaction with strain rate in deformation mechanisms. Hydrogen decreases low-angle grain boundary proportion, intensifies local strain concentration, and increases GND density. A proposed mechanism suggests that hydrogen aggravates local plastic deformation through dislocation pinning at high strain rates, but this pinning effect diminishes with increasing strain rate. This research provides new experimental evidence and theoretical interpretation for understanding hydrogen effects on 18CrNiMo7-6 steel under high strain rates, offering valuable insights for optimizing its applications in extreme service conditions.

The coupling of strain-induced precipitation hardening behavior and static recrystallization softening behavior between passes in the hot rolling production process determines the change of steel microstructure. For Nb-Ti composite microalloyed steel, the complexity of its precipitation behavior process and structure evolution leads to low accuracy of prediction results from traditional physical metallurgy models which are built based on hyptheses and experimental data. With the wide application of machine learning in the steel production process, it is introduced into the modeling process of physical metallurgy behavior during the heat deformation of Nb-Ti microalloyed steel. With the methodology being established based on the collection of text data, the components and process parameters that have significant influence on the strain-induced precipitation and recrystallization behaviors were screened as model input variables through correlation analysis, and three machine learning algorithms, including Random Forest （RF）, Extreme Gradient Boosting （XGBoost）, and Artificial Neural Networks （ANN）, were used to establish the time corresponding to a static recrystallization fraction of 0.5 （t_0.5） and material parameters （n） in the recrystallization model, as well as the starting time（t_ps） and finishing time （t_pf） of precipitation in the precipitation model, respectively. The root mean square error （R_MSE） of t_0.5,n,t_ps and t_pf calculated by the RF model were determined to be 2.25, 0.08, 49.50, and 1 252.8, respectively, which is the smallest compared with other machine learning algorithms. The two-pass compression experiment on 700XL steel finds that when the deformation temperature is 1 000 ℃, it is a typical recrystallization softening process. When the deformation temperature is 950 ℃ and 925 ℃, the strain-induced precipitation of microalloying elements and static recrystallization will appear at the same time, and the coupling effect of them leads to the softening rate curve appearing a "plateau". The experimental data validation shows that the established machine learning model is better than the traditional physical metallurgy model in the calculation accuracy of the starting and finishing time of precipitation, while the calculated softening rate curve presents the interaction law of recrystallization and precipitation very well.

The two-roll extrusion coating of silicon steel is a process of liquid-solid coupling between the coating liquid, the coating roll and the strip steel. The deformation of coating roll with micro-grooving on the roll surface is highly nonlinear. In order to improve the coating uniformity of wide strip steel, the distribution of roll surface contact stress and groove area after roll system was loaded. Under the condition of no fluid, the mechanical analysis of roll system was carried out. In view of the composite structure of rubber and steel cylinder, the calculation method of the total pressure between the rolls of strip steel and rubber, rubber and rubber contact was proposed for the first time. Based on the constitutive model of superplastic rubber, a finite element model suitable for multi-groove analysis was established by adopting the whole-local strategy and combining the actual parameters. A visual testing platform for the deformation of rollers under load was designed and built for the first time. The accuracy of the finite element method was verified by applying different loads and using image processing algorithm to extract the indentation width. The effects of load, rubber hardness, band width and thickness on roll surface contact stress and groove area distribution were revealed. The results show that when the load is 20 kN, the contact stress of the contact zone increases first and then decreases from the middle of the groove to the end. The bottom and side edges of grooves are irregularly raised after loading. The maximum contact stress is 2.457 times of the middle contact stress. The reduction rate of groove area at different positions of roll surface is 29.609%-31.176%. When the load is 10-30 kN, the stress at the end of coater in the contact zone between the strip and the coater is increased by 4.079%- 9.112%, and the grooving area is reduced by 1.144%-3.421%. When the bandwidth is 1 000-1 400 mm, the end stress increases by 4.988%-9.208% compared with the middle stress, and the area decreases by 1.552%-3.016%. The greater the rubber hardness, the more uneven the stress distribution, and the area distribution tends to be uniform. The influence of strip thickness on stress and groove area distribution is negligible. The research results provide an important test for the subsequent fluid-structure coupling analysis.

Aiming at the problems of poor accuracy and low detection speed in casting surface defect detection, a casting surface defect detection algorithm named SPI based on the improved YOLOv8n is proposed, and the algorithm is trained via distillation using Channel-Wise Knowledge Distillation （CWD）.First, the S-Fasternet network is designed and replaces the backbone network of YOLOv8n. This network reduces redundant computations and memory access while capturing rich contextual information in both horizontal and vertical dimensions to preserve subtle casting defect features. Subsequently the P-C2f module is proposed to enhance the network's feature representation while reducing information loss. An Inner-CIoU bounding box regression loss function is introduced to improve localization performance. Furthermore, based on the SPI object detection network, CWD technology is utilized to further improve the model's detection accuracy.Finally, combining the CWD-SPI detection algorithm with the MySQL database, a casting surface defect detection system is developed. This system enables real-time execution of multiple defect detection tasks while storing and analyzing results, facilitating real-time statistics of defect information. It demonstrates stronger flexibility and adaptability in addressing various detection requirements. Experimental results show that the CWD-SPI algorithm achieves 85.9% mean average precision on casting defect datasets, representing a 5.9% point improvement over the original YOLOv8n, with a 5% point increase in precision. The enhanced algorithm not only improves defect detection performance but also satisfies real-time requirements for industrial applications.