High performance die steel and method of manufacture

Abstract

The present application relates to the technical field of high-performance die steel material, and particularly relates to a high-performance die steel and a manufacturing method, Sn, Ba and Zn are added in the die steel, through the synergistic effect of Sn, Ba and Zn, the effect of inhibiting carbide coarsening and reducing residual stress are balanced, meanwhile, through nitrogen and niobium composite micro-alloying, the grain is further refined, the thermal stability and strength are improved, a higher-performance die steel is formed, Nb(C, N) is precipitated in the rolling process, the grain boundary is strongly pinned, the austenite grain growth is inhibited, the carbonitride such as stable NbC, NbN and VN is difficult to coarsen and dissolve at high temperature, can long-term pin the grain boundary and dislocation, by adjusting the N content, the strength and wear resistance of the material can be 'customized' in a certain range, a higher-performance die steel is formed.

Description

Technical Field

[0001] This invention relates to the field of high-performance mold steel materials, and particularly to a high-performance mold steel and its manufacturing method. Background Technology

[0002] Alloy pressure casting technology can precisely mold high-strength parts with complex structures, an advantage that makes it play a vital role in key fields such as automobile manufacturing and aerospace. As the core tool in aluminum alloy die casting production, die casting mold steel has undergone long-term technological iterations and has become an indispensable key material in industrial production.

[0003] In aluminum alloy casting, die-casting mold steel must continuously withstand the dual effects of thermomechanical stress and frictional loads. This requires the mold steel to possess excellent thermal fatigue properties, stable thermal stability, and reliable resistance to aluminum thermal melting. Simultaneously, to meet processing and forming requirements, it must also have excellent machinability. Furthermore, residual stress in the mold may cause stress concentration leading to cracking failure during service, and may also induce pitting corrosion, thus affecting the mold's surface quality and service life. Therefore, scientifically controlling the residual stress in die-casting mold steel is a crucial step in ensuring the reliability of the mold during service.

[0004] Traditional H13 steel is a widely used hot-work die steel, hailed as the "king of hot-work die steels." However, its core shortcomings in aluminum alloy casting are that its thermal stability, resistance to aluminum hot melting, and thermal fatigue performance are insufficient to fully meet stringent requirements, limiting its service performance. Specifically, its thermal fatigue performance is inadequate; the repeated heating-cooling cycles of aluminum alloy casting generate thermal stress within the H13 steel, easily leading to the initiation and gradual propagation of hot cracks. Its resistance to aluminum hot melting is limited; at high temperatures, molten aluminum reacts chemically with the H13 steel, causing surface erosion and aluminum adhesion, affecting casting accuracy and die life. Its poor thermal stability means that under prolonged high-temperature conditions, the steel's hardness and strength gradually decrease, increasing the risk of die deformation. Finally, residual stress control is difficult; post-processing residual stress easily superimposes with operating stress during service, accelerating die cracking or pitting failure.

[0005] Therefore, there is an urgent need for a high-performance mold steel that can further improve the toughness, thermal stability and high-temperature performance of materials. Summary of the Invention

[0006] To address the above problems, this invention provides a high-performance mold steel and its manufacturing method. By adding Sn, Ba, and Zn to the mold steel, the synergistic effect of Sn, Ba, and Zn balances the effect of inhibiting carbide coarsening and reducing residual stress. At the same time, by combining nitrogen and niobium composite microalloying, the grains are further refined, and the thermal stability and strength are improved, resulting in a higher-performance mold steel.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A high-performance mold steel, comprising the following components by weight percentage: C: 0.30%-0.38%, Si: 1.00%-1.20%, Mn: 1.10%-1.30%, Cr: 3.0%-3.4%, Mo: 3.3%-3.5%, V: 1.00%-1.30%, Co: 0.80%-1.00%, N: 0.02%-0.05%, Nb: 0.03%-0.06%, Sn: 0.10%-0.15%, Ba: 0.05%-0.10%, Zn: 0.10%-0.20%, P: ≤0.012%, S: ≤0.008%, balance Fe and unavoidable impurities; In mold steel, the ratio of the sum of Ba and Zn content to Sn content is controlled at 1.8-2.2; the ratio of Mo content to Co content is controlled at 3.5-4.2.

[0008] As an improvement, the manufacturing method of the above-mentioned high-performance mold steel includes the following steps: Primary refining, refining, vacuum treatment, continuous casting, slab heating, rolling, cooling, spheroidizing annealing, quenching, cyclic cryogenic treatment and tempering; The refining process includes LF furnace refining and VD furnace refining. Fe-Nb is added in the early stage of LF furnace refining. In VD furnace refining, manganese nitride or chromium-nitrogen alloy is added under vacuum to adjust the nitrogen content. In the later stage of VD furnace refining, cored wires of Sn, Ba and Zn are added using wire feeding technology. When adding cored wires of Sn, Ba and Zn using wire feeding technology, FeBa wire, Zn wire and Sn wire are fed in sequence.

[0009] As an improvement, in the VD furnace refining step, the amount of nitrogen added is controlled by a nitrogen yield prediction model. The method for establishing the nitrogen yield prediction model includes the following steps: The data acquisition step involves real-time acquisition of multi-source process data during the VD refining process. The multi-source process data includes steel composition data, temperature data, vacuum degree data, stirring parameter data, and nitrogen alloying operation data. The feature engineering step involves constructing a feature set for predicting nitrogen yield based on the multi-source process data. The feature set includes basic features, derived features, and interactive features. The model training steps involve training an ensemble machine learning model using historical data, with the feature set as input and nitrogen yield as output. The real-time prediction step involves inputting the real-time acquired process parameters into a trained integrated machine learning model during the VD refining process to predict the nitrogen yield of the current furnace. The control output step calculates and outputs the recommended amount of nitrogen alloy addition based on the predicted nitrogen yield.

[0010] As an improvement, in the data acquisition step, the steel composition data includes at least the real-time contents of C, Si, Mn, Cr, Mo, V, Al, O, and S; the temperature data includes the LF outlet temperature, VD start temperature, and nitrogen alloying temperature; the vacuum degree data includes the vacuum degree change curve and holding time; the stirring parameter data includes the bottom-blown argon flow rate and stirring mode; and the nitrogen alloying operation data includes the timing, method, and batch of nitrogen alloy addition.

[0011] As an improvement, the derived feature includes a nitrogen solubility parameter, which is calculated as follows: f(T,P,composition)=K×exp(ΔH / RT)×√P_N2, where T is the temperature of the molten steel, P is the vacuum degree, R is the gas constant, and K and ΔH are constants related to the composition of the molten steel. The deoxidation degree index is calculated based on the [O] and [Al] content in the molten steel. Historical average alloy yield, calculated based on the historical average yield of the same steel grade; Process stability index is calculated based on the parameter fluctuations of preceding processes.

[0012] As an improvement, the interaction feature includes a temperature-vacuum interaction term: T×log(P), where T is the temperature of the molten steel and P is the vacuum degree; Alloy type - characteristics of the combination of addition timing; Stirring intensity - vacuum degree synergy coefficient.

[0013] As an improvement, the ensemble machine learning model in the model training step adopts a two-layer structure; The first layer includes multiple heterogeneous base prediction models, which are selected from at least two of XGBoost, gradient boosting tree and random forest; The second layer is a meta-learner, whose input is the output of each basic prediction model in the first layer. The meta-learner uses a multilayer perceptron neural network.

[0014] The architecture of the nitrogen yield prediction system includes: Data acquisition module: Responsible for collecting process data from multiple data sources in real time: Process control system interface: Connects to the basic automation system to obtain real-time data. Laboratory management system interface: Obtaining chemical composition analysis results Data caching layer: Temporarily stores and preprocesses raw data Quality inspection module: Verifies the integrity and accuracy of data. Feature engineering module: Transform the raw data into features usable by the model: Feature computation engine: Performs computation of derived features and interaction features. Feature repository: stores feature definitions and computation logic Feature monitoring: Tracking changes in the quality and distribution of feature data. Feature version management: Managing feature sets for different versions Model storage module: Responsible for storing and managing machine learning models: Model repository: Stores trained model files and configuration information. Version control system: manages model versions and history Model serialization: Converting model objects into a storable format. Fast loading mechanism: Supports fast loading of models in production environments Prediction calculation module: Perform real-time nitrogen yield prediction: Model inference engine: Loads the model and performs prediction calculations. Batch processing support: Supports single-line and batch prediction modes. Resource management: Allocate computing resources rationally to ensure response speed. Anomaly Handling: Properly handle various abnormal situations during the forecasting process. Result output module: Transform prediction results into operational guidelines: Result formatting: Organize output information according to a standard format. Interface adaptation: Adapting to the display requirements of different terminal devices. Alarm Trigger: Trigger the corresponding level of alarm based on the risk level. Historical records: Save complete prediction records and actual results.

[0015] Also includes: Model update module: To achieve automated management and optimization of the model: Performance monitoring: Continuously track the model's predictive performance in the production environment. Attenuation Detection: Automatically detect model performance degradation and concept drift Update scheduling: Intelligently schedules model update times to avoid impacting production. A / B testing: Supports parallel running and performance comparison of new and old models. Human-computer interaction interface: Provide a user-friendly interactive experience for operators: Real-time dashboard: Displays key process parameters and forecast results Historical trend chart: Showing the historical trend of nitrogen yield. Operation log: Records all manual interventions and feedback information. Mobile support: Access system functions via mobile devices.

[0016] As an improvement, the model training step adopts the time series cross-validation method, which divides the training set and validation set according to the time sequence of the furnace batches to ensure the time sequence of model validation.

[0017] As an improvement, it also includes an online learning step for the model, which collects nitrogen yield data from actual production and its corresponding process parameters; when the amount of collected data reaches a predetermined threshold, it uses new data to incrementally learn or retrain the integrated machine learning model; and it updates the model parameters to adapt to changes in process conditions.

[0018] As an improvement, in the control output step, the recommended amount of nitrogen alloy added is calculated using the following formula: Recommended addition amount = (target nitrogen content - current nitrogen content) / (predicted yield / 100) Simultaneously, the control output step also outputs a confidence assessment of the prediction result, which is based on the similarity calculation between the input features and the training data distribution.

[0019] The beneficial effects of this invention are as follows: (1) This invention adds Sn, Ba and Zn to the mold steel. Through the synergistic effect of Sn, Ba and Zn, the coarsening inhibition effect of carbides and the reduction of residual stress are balanced. At the same time, the grains are further refined and the thermal stability and strength are improved by nitrogen and niobium composite microalloying, forming a higher performance mold steel. Nb(C,N) precipitates during the rolling process, strongly pinning the grain boundaries and inhibiting the growth of austenite grains. Stable carbonitrides such as NbC, NbN and VN are extremely difficult to coarsen and dissolve at high temperatures. They can pin the grain boundaries and dislocations for a long time. By adjusting the N content, the strength and wear resistance of the material can be "customized" within a certain range. (2) In the process of preparing mold steel, the present invention uses wire feeding technology to inject Sn, Ba, Zn and other elements into the depth of the molten steel in the form of cored wire in the later stage of VD treatment, so as to avoid the oxidation of these elements and also avoid the possibility of a large amount of element burn-off caused by the high oxygen level of the molten steel. At the same time, the use of wire feeding technology can ensure that the temperature of the molten steel in the later stage of VD treatment can still dissolve Sn, Ba and Zn. (3) This invention introduces a nitrogen yield prediction model in the process of manufacturing mold steel, and uses the nitrogen yield prediction model to accurately control the nitrogen absorption rate in the manufacturing process, thereby reducing the mean absolute error (MAE) of nitrogen yield prediction from ±8% based on human experience to within ±2%.

[0020] In summary, the present invention has the advantages of producing turbine rotor blades with high purity, low oxidation, ideal internal metal microstructure, and uniform structure, and is particularly suitable for the field of high-temperature alloy processing of turbine rotor blades. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the manufacturing method of the present invention; Figure 2 Microstructure morphology of the mold steel obtained in Example 1; Figure 3 The microstructure morphology of the mold steel obtained in Example 2 is shown in the image. Figure 4 Microstructure morphology of the mold steel obtained in Example 3; Figure 5 Microstructure morphology of the mold steel obtained in Example 4; Figure 6 Microstructure morphology of the mold steel obtained in Example 5; Figure 7 This is a schematic diagram of the nitrogen yield prediction model establishment method of the present invention; Figure 8 This is a schematic diagram illustrating the principle of the XGBoost model of this invention; Figure 9 This is a schematic diagram illustrating the principle of the gradient boosting tree model of the present invention; Figure 10 This is a schematic diagram illustrating the principle of the random forest model of this invention; Figure 11 This is a schematic diagram illustrating the principle of the meta-learner of this invention; Figure 12 This is a schematic diagram of the architecture of the nitrogen yield prediction system of the present invention; Figure 13 This is a comparison curve of nitrogen yield stability in this invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1: A high-performance mold steel, comprising the following components by weight percentage: C: 0.30%-0.38%, Si: 1.00%-1.20%, Mn: 1.10%-1.30%, Cr: 3.0%-3.4%, Mo: 3.3%-3.5%, V: 1.00%-1.30%, Co: 0.80%-1.00%, N: 0.02%-0.05%, Nb: 0.03%-0.06%, Sn: 0.10%-0.15%, Ba: 0.05%-0.10%, Zn: 0.10%-0.20%, P: ≤0.012%, S: ≤0.008%, balance Fe and unavoidable impurities; In mold steel, the ratio of the sum of Ba and Zn content to Sn content is controlled at 1.8-2.2; the ratio of Mo content to Co content is controlled at 3.5-4.2.

[0024] It should be noted that Sn, by segregating at grain boundaries, acts like a "roadblock" to hinder the diffusion of carbide growth elements, but its individual effect may lead to grain boundary brittleness or stress problems. The synergistic addition of Ba and Zn seems to play a "lubricating" and "buffering" role, effectively offsetting the potential negative effects of Sn (such as increased residual stress) while suppressing carbide coarsening, thus achieving a unity of the two contradictory goals of "suppressing coarsening" and "reducing stress".

[0025] Meanwhile, a key role of Co is to reduce the solubility of Mo in the steel matrix, thereby "forcing" more Mo into carbide precipitation. This ratio range ensures that there is enough Mo to form reinforcing carbides, while just the right amount of Co to maximize the process, avoiding cost waste or performance degradation caused by excessive Co.

[0026] Furthermore, the addition of Nb results in the formation of highly stable NbC (niobium carbide) and NbN (niobium nitride) compounds in steel. These compounds do not completely dissolve in austenite during austenitizing heating. The undissolved Nb(C,N) particles act like "nails" anchoring themselves at the austenite grain boundaries, strongly hindering grain boundary migration. Even at higher austenitizing temperatures, austenite grains are difficult to grow, resulting in fine, pristine austenite grains. According to the Hall-Page equation, finer grains simultaneously increase both the strength and toughness of the material. This is the primary reason for the high toughness of Nb. Moreover, during the cooling or tempering process of heat treatment, Nb precipitates from the matrix as extremely fine, dispersed nanoscale NbC. These nanoscale precipitates effectively impede dislocation movement, significantly improving the strength and hardness of the material. Due to their small size and dispersed distribution, their impact on toughness is far less than that of coarse carbides. Furthermore, during thermomechanical rolling, Nb inhibits the recrystallization of austenite through the solid solution dragging effect in austenite and the strain-induced precipitation of Nb(C,N), which elongates and refines the deformed austenite grains, resulting in a finer ferrite or bainite structure after phase transformation.

[0027] This invention also incorporates trace amounts of nitrogen (N), a strong austenite-forming and stabilizing element. N can dissolve in the steel matrix, causing lattice distortion, providing solid solution strengthening, and improving the hardenability of the steel. More importantly, it alters the solubility and distribution of other alloying elements in austenite and ferrite, indirectly affecting precipitation behavior. N has a strong affinity for elements such as v, nb, and chromium, preferentially forming vanadium nitride (VN), niobium nitride (NbN), and composite (Nb,V)(C,N) carbonitrides. VN is one of the hardest known dispersed precipitates; its precipitation during tempering brings a strong secondary hardening effect, significantly improving the material's hot strength, red hardness, and wear resistance. The addition of N causes the originally formed volatile organic compounds (VC) to be replaced by VN, or to form composite carbonitrides. These carbonitrides have higher stability (less soluble and less prone to coarsening) than simple carbides.

[0028] Furthermore, N and Nb do not act independently but rather mutually reinforcingly. The presence of N promotes the formation of NbN and (Nb,V)(C,N), while the presence of Nb also fixes N, preventing it from escaping in the form of bubbles or forming harmful BN (boron nitride). This mutual promotion significantly increases the number of fine, stable, and dispersed precipitates in the steel. Moreover, because nitrides / carbonitrides such as Nb(C,N) and VN are extremely stable at high temperatures, their coarsening rate is very slow. In molds operating under high-temperature environments for extended periods (such as die casting), these fine precipitates can effectively pin grain boundaries and dislocations, preventing dislocation climb and recrystallization. After long-term service at high temperatures, the mold steel experiences a smaller decrease in hardness and a more stable microstructure. This means that the mold can maintain high precision and strength for a longer lifespan and has stronger resistance to thermal fatigue. Additionally, Nb, through Nb(C,N) pinning grain boundaries, and N, through promoting the formation of more VN and other precipitates, also indirectly contribute to grain refinement. The two work together to achieve an even more refined grain size.

[0029] In summary, the combined addition of N and Nb, through the synergistic effect of multiple mechanisms such as grain refinement strengthening, precipitation strengthening, and solid solution strengthening, especially the formation of a series of highly stable nanoscale carbonitride precipitates, fundamentally optimizes the microstructure of mold steel, thereby achieving a breakthrough improvement in its comprehensive properties such as strength, toughness, thermal stability, and wear resistance.

[0030] Example 2: like Figure 1 As shown, the manufacturing method of the high-performance mold steel of Embodiment 2 of the present invention, as described with reference to Embodiment 1, includes the following steps: Primary refining, refining, vacuum treatment, continuous casting, slab heating, rolling, cooling, spheroidizing annealing, quenching, cyclic cryogenic treatment and tempering; The refining process includes LF furnace refining and VD furnace refining. Fe-Nb is added in the early stage of LF furnace refining. In VD furnace refining, manganese nitride or chromium-nitrogen alloy is added under vacuum to adjust the nitrogen content. In the later stage of VD furnace refining, cored wires of Sn, Ba and Zn are added using wire feeding technology. When adding cored wires of Sn, Ba and Zn using wire feeding technology, FeBa wire, Zn wire and Sn wire are fed in sequence.

[0031] Specifically, step one: raw materials and smelting Electric Arc Furnace (EAF) Primary Refining: Raw materials: High-quality low-phosphorus and low-sulfur scrap steel and pig iron.

[0032] Process: Oxygen blowing dephosphorization, with a final carbon control of ≥0.10%.

[0033] Tack temperature: ≥1630℃.

[0034] LF furnace refining: Slag formation: to produce high-alkalinity white slag.

[0035] Alloying: The composition is coarsely adjusted by adding bulk alloys such as Fe-Cr, Fe-Mo, Ni, and Fe-V.

[0036] Desulfurization: [S] is reduced to ≤0.008% by strong stirring.

[0037] Temperature control: Heat the molten steel to 1580-1600℃.

[0038] Microalloying: Before tapping, Fe-Nb alloy is added and stirred thoroughly to dissolve it evenly.

[0039] VD Vacuum Refining: Vacuuming: Start the vacuum pump and raise the vacuum level to ≤0.5mbar (50Pa) within 15-20 minutes, and maintain it for 15-20 minutes.

[0040] Degassing: Deep removal of [H] to ≤1.5ppm and [O] to ≤15ppm.

[0041] Nitrogen alloying: Under vacuum conditions, the nitrogen content is precisely adjusted by adding manganese nitride (MnN) or chromium nitride (CrN) alloy.

[0042] Component fine-tuning: Based on the final rapid analysis results, final fine-tuning is performed on easily oxidized elements such as Sn, Ba, and Zn.

[0043] Secondary vacuuming: Start the vacuum pump and raise the vacuum level to ≤0.5mbar (50Pa) within 5-8 minutes, and maintain it for 5-8 minutes.

[0044] Continuous casting: Protective casting: Long nozzle + argon sealing is used to prevent secondary oxidation.

[0045] Superheat: Strictly control the superheat of molten steel between 25-35℃.

[0046] Cooling: Electromagnetic stirring combined with a weak cooling mode is used to obtain a continuous casting billet with uniform composition and small central porosity.

[0047] Cross-section: Cast into a rectangular blank of the required dimensions.

[0048] Step 2: Thermomechanical rolling.

[0049] Slab heating: Temperature: 1180℃-1220℃.

[0050] Holding time: ≥2 hours / ton, to ensure uniform heating of the billet and full dissolution of alloying elements.

[0051] Controlled rolling: Rolling temperature: ≤1150℃.

[0052] Recrystallization zone rolling: The first few passes are carried out at temperatures above 1050℃ to fully break down the as-cast structure.

[0053] Non-recrystallization zone rolling: The final 3-5 passes are rolled in the range of 950℃-850℃. At this temperature, Nb(C,N) precipitates as fine particles, strongly pinning grain boundaries, inhibiting recrystallization, and accumulating deformation energy, thereby forming extremely fine ferrite / bainite grains in subsequent processes.

[0054] Final rolling temperature: controlled at 850℃-880℃.

[0055] Post-rolling cooling: After rolling, an ACC (Accelerated Cooling) system is used to rapidly cool the steel plate to 550℃-600℃ at a cooling rate of 15-25℃ / s.

[0056] It is then slowly cooled to room temperature in a slow-cooling pit. This process transforms austenite into a fine bainite structure.

[0057] Step 3: Preliminary heat treatment - spheroidizing annealing Objective: To reduce hardness (HB≤240) to facilitate machining and prepare the microstructure for final heat treatment.

[0058] Process: Isothermal spheroidizing annealing is adopted.

[0059] Heat to 780℃-800℃ at a rate of ≤100℃ / h and hold for 4-6 hours.

[0060] Then the furnace is cooled to 690℃-710℃ and held at that temperature for 4-6 hours.

[0061] Finally, the furnace is cooled to below 500°C before being removed and air-cooled.

[0062] Structure: A uniform granular pearlite structure was obtained.

[0063] Step 4: Final heat treatment Quenching: Austenitizing temperature: 1050℃-1080℃. Higher temperatures facilitate the complete solid solution of Mo and V carbides and Nb carbonitrides, providing a reserve of solutes for secondary hardening.

[0064] Holding time: Calculated based on the effective thickness of the workpiece, typically 1.0-1.5 min / mm (in an air furnace).

[0065] Cooling: Cooling is carried out in a vacuum high-pressure gas quenching furnace. High-purity nitrogen gas at a quenching pressure of 8-12 bar is used. Gas quenching significantly reduces workpiece deformation and yields a uniform martensitic structure.

[0066] Cryogenic treatment: Procedure: Immediately after quenching, transfer the workpiece to the cryogenic chamber.

[0067] Parameters: Cool to -150℃ to -196℃ at a rate of ≤5℃ / min, and hold for 1-3 hours. Then remove and allow to return to room temperature in air. Repeat this process twice.

[0068] Function: Deeply eliminates residual austenite, promotes the precipitation of fine carbides, and improves dimensional stability and wear resistance.

[0069] Tempering: Process: Three tempering cycles are employed.

[0070] Temperature: Select according to the required hardness, preferably 560℃-600℃.

[0071] Time: Each heat treatment lasts 2.5-3.5 hours, followed by air cooling after tempering.

[0072] Function: To achieve sufficient secondary hardening, relieve stress, and stabilize the microstructure. After the third tempering, the hardness can reach 48-52 HRC, and it also has extremely high toughness reserves.

[0073] Preparation Examples 1-5: Table 1. Chemical composition (wt%) of each preparation example of the present invention. Table 2 compares the test results of the mold steels prepared in each example of the present invention with those of conventional H13 steel. like Figures 2-6 As shown in the comparison above, Preparation Example 3 performed best, with a hardness loss of only 0.8 HRC after holding at 580℃ for 100 hours, which is far better than the 3.5 HRC loss of H13 steel.

[0074] All prepared samples exhibited a high-temperature hardness exceeding 42 HRC at 600°C, which is more than 20% higher than that of H13 steel (35 HRC).

[0075] High levels of N and Nb form stable Nb(C,N) and VN precipitates with V, effectively pinning grain boundaries and inhibiting microstructure coarsening at high temperatures.

[0076] Preparation Example 2 achieved an impact toughness of 35 J / cm², which is 40% higher than that of H13 steel, while maintaining a high hardness of 49 HRC.

[0077] Preparation Example 5 maintains excellent toughness of 32 J / cm² at a hardness of 50 HRC, achieving the best combination of strength and toughness.

[0078] A balanced design of fine grain strengthening (Nb effect) and moderate alloying avoids excessive coarse carbides from compromising toughness.

[0079] Preparation Example 4 achieved the highest hardness of 52.5 HRC and the best wear resistance (volume loss of only 35%). The combined effect of cryogenic treatment and high Mo content promoted the precipitation of fine secondary hardening phases.

[0080] All preparations showed a 45-65% improvement in wear resistance compared to H13 steel.

[0081] All prepared samples showed more negative residual stress values ​​(deeper compressive stress), which is beneficial for suppressing fatigue crack initiation.

[0082] Preparation Example 4 achieved the deepest residual compressive stress (-365 MPa) due to cryogenic treatment, which is 30.4% better than that of H13 steel.

[0083] Example 3: like Figures 7-13 As shown, this invention adds nitrogen (N) to the mold steel. The absorption rate of N is affected by various factors such as the temperature of the molten steel, the content of [O] and [S], and the stirring intensity, making it difficult to predict and control precisely. Furthermore, the nitrogen content must be controlled within a very narrow range of 0.02%-0.05%, requiring extremely high precision. Too low a content results in insufficient microalloying, while too high a content may lead to porosity or macroscopic segregation in the cast billet. Although oxygen and nitrogen probes can be used for online monitoring, their response still has a certain lag, making it difficult to achieve true "real-time closed-loop control."

[0084] Therefore, it is necessary to establish an accurate nitrogen yield prediction model to calculate the amount of nitrogen to be added in real time based on parameters such as steel composition, temperature, and vacuum degree.

[0085] Specifically, we first define the input feature vector X for the nitrogen yield prediction model: X=[x1,x2,x3,...,x n ] T Each feature component includes: Basic process characteristics x1: Temperature of molten steel T (°C) x2: Vacuum degree P (Pa) x3: Bottom-blown argon flow rate Q_Ar (NL / min) x4: Aluminum content [Al] (wt%) in molten steel x5: Oxygen content [O] (wt%) in molten steel x6: Carbon content [C] (wt%) in molten steel x7: Timing of alloy addition t_add(min).

[0086] Derivative metallurgical characteristics x8: Deoxygenation Index (DEI) DEI=K×log 10 ([%Al] / ([%O]+ε)), where [%Al]: the mass percentage of dissolved aluminum in the molten steel, [%O]: the mass percentage of dissolved oxygen in the molten steel, K: the proportionality coefficient, with a value ranging from 0.8 to 1.2, usually taken as 1.0, and ε: a minimum constant, usually taken as 0.0001 to avoid the denominator being zero.

[0087] x9: Nitrogen solubility parameter NSP NSP = K_N × exp(ΔH / (R × T)) × √P_N2, where T is the temperature of the molten steel, P is the vacuum degree, R is the gas constant, and K and ΔH are constants related to the composition of the molten steel.

[0088] x 10 Process stability index (PSI) PSI = 1 / (1+CV), where CV is the coefficient of variation of the key parameters of the preceding process.

[0089] x 11 Historical average alloy yield, calculated based on the historical average yield of the same steel grade.

[0090] Interactive features X 12 Temperature-vacuum interaction term TVI = T × log(P), where T is the temperature of the molten steel and P is the vacuum degree. x 13 Stirring-vacuum synergy coefficient SVC=(Q_Ar / 100)×(1-P / 101325) x 14 Alloy-Timing Combination Characteristics Coding combination of alloy type and timing of addition Tag data Y: Actual nitrogen yield, calculated retrospectively using the following formula: Actual nitrogen recovery rate (%) = (nitrogen content of molten steel after alloying - nitrogen content of molten steel before alloying) / total weight of added nitrogen element × weight of molten steel × 100%.

[0091] Establish nitrogen yield prediction functions for each basic model. XGBoost model 1. For nitrogen yield prediction, the functional expression of the XGBoost model is: _xgb(X)=Σ[k=1toK]f_k(X); In this context, the predicted value of each tree f_k(X) is the weight w_{q_k(X)} of the leaf node where sample X falls, and q_k(X) represents the index of the leaf node where sample X falls in the k-th tree.

[0092] 2. Objective function of XGBoost in nitrogen yield prediction: L_xgb=Σ[i=1toN]l( _xgb(X_i),y_i)+Σ[k=1toK]Ω(f_k); Wherein, the regularization term Ω(f_k) = γT + ½λ‖w‖², T is the number of leaf nodes, and ‖w‖² is the sum of squared weights of the leaves.

[0093] 3. Calculation of optimal weights for leaf nodes (MSE loss for nitrogen yield): w_j =-(Σ[i∈I_j]g_i) / (Σ[i∈I_j]h_i+λ); Where g_i= l / h_i= ²l / ².

[0094] Unlike traditional GBDT which only uses the first-order gradient, XGBoost uses the second-order Taylor expansion to more accurately approximate the loss function, thereby achieving faster convergence and higher prediction accuracy.

[0095] The mathematical principle is as follows: 1. Model prediction function XGBoost's nitrogen yield prediction function is an additive model with K trees: _xgb(X)=Σ[k=1 to K] f_k(X); The predicted value of f_k(X) for each tree is: f_k(X) = w_{q_k(X)}, q_k(X), The index of the leaf node where sample X falls in the k-th tree. w_{q_k(X)}: The weight value of this leaf node; 2. Objective function optimization XGBoost's objective function includes a loss function and a regularization term: L_xgb =Σ[i=1 to N]l( _xgb(X_i), y_i)+Σ[k=1 to K]Ω(f_k); Regularization term: Ω(f_k) = γT + ½λ‖w‖²; γ: Complexity control parameter, penalizing the creation of new leaf nodes. T: The number of leaf nodes in the tree. λ: L2 regularization coefficient, controlling the weight of the leaf elements. ||w||²: The sum of squares of the weights of all leaf nodes; 3. Calculation of optimal weights for leaf nodes For leaf node j, its optimal weight is obtained by minimizing the objective function: w_j =-(Σ[i∈I_j] g_i) / (Σ[i∈I_j]h_i+λ); Gradient calculation (MSE loss for nitrogen yield): g_i = h_i = = 1; Therefore, for the MSE loss function, the formula for calculating the leaf node weights simplifies to: w_j =-(Σ[i∈I_j]( _i-y_i)) / (|I_j| + λ); 4. Tree splitting criteria XGBoost uses a greedy algorithm to select the optimal splitting features and splitting points. The splitting gain is calculated using the following formula: Gain =½[(Σ_{left}g_i)² / (|I_left|+λ)+(Σ_{right}g_i)² / (|I_right| +λ)-(Σ_{parent}g_i)² / (|I_parent|+λ)]-γ.

[0096] Specific example: The entire process of nitrogen yield prediction Training data preparation Suppose we have 4 batches of historical data as a training set: Model training process First iteration: Construct the first tree (k=1) Step 1.1: Initial Prediction （0） = (80.5 + 75.2 + 78.8 + 77.5) / 4 = 78.0% For all samples: _i （0） = 78.0%; Step 1.2: Calculate the gradient (assuming λ = 1.0) g1= 78.0 - 80.5 = -2.5, h1= 1, g2= 78.0 - 75.2 = 2.8, h2= 1, g3 = 78.0 - 78.8 = -0.8, h3= 1, g4= 78.0 - 77.5 = 0.5, h4= 1; Step 1.3: Finding the optimal split Attempting to split the characteristic "temperature" at 1595℃: Left node: Temperature < 1595, containing sample 2 → Σg_left = 2.8, |I_left| = 1. Right node: Temperature ≥ 1595, containing samples 1, 3, 4 → Σg_right = -2.8, |I_right| = 3. Calculate the split gain: Gain = ½ [ (2.8)² / (1+1) + (-2.8)² / (3+1) - (0)² / (4+1) ] - γ = ½ [ 7.84 / 2 + 7.84 / 4 - 0 ] - 0 = ½ [ 3.92 + 1.96 ] = ½ × 5.88 = 2.94, When the gain is positive, it accepts splitting; Step 1.4: Calculate the weights of the leaf nodes The weight of the left leaf, w_left, is -(2.8) / (1+1) = -1.4. The weight of the right leaf node w_right = -(-2.8) / (3+1) = 0.7; Step 1.5: Update the predicted values ​​(assuming a learning rate η = 0.1) Samples 1, 3, and 4: _i （1） =78.0 + 0.1 × 0.7 = 78.07% Sample 2: _i （1） =78.0 + 0.1 × (-1.4) = 77.86%, Second iteration: Construct the second tree (k=2); Step 2.1: Recalculate the gradient g1= 78.07 - 80.5 = -2.43, h1= 1, g2= 77.86 - 75.2 = 2.66, h2= 1, g3= 78.07 - 78.8 = -0.73, h3= 1, g4= 78.07 - 77.5 = 0.57, h4= 1; Step 2.2: Finding the optimal split Try splitting the feature “[Al]%” at 0.034: Left node: [Al]% < 0.034, containing samples 2 and 4 → Σg_left = 2.66 + 0.57 = 3.23, |I_left| = 2 Right node: [Al]%≥0.034, containing samples 1,3 → Σg_right = -2.43-0.73=-3.16, |I_right|=2 Calculate the split gain: Gain = ½ [ (3.23)² / (2+1) + (-3.16)² / (2+1) - (0)² / (4+1) ] - 0 = ½ [ 10.43 / 3 + 9.99 / 3 ] = ½ [ 3.48 + 3.33 ] = 3.41 Step 2.3: Calculate the weights of the leaf nodes The weight of the left leaf node, w_left, is -(3.23) / (2 + 1) = -1.08 The weight of the right leaf node, w_right, is -(-3.16) / (2 + 1) = 1.05 Step 2.4: Update the predicted values Samples 1, 3: _i （2） = 78.07 + 0.1 × 1.05 = 78.18% Samples 2,4: _i （2） = 77.86 + 0.1 × (-1.08) = 77.75% Third iteration: Construct the third tree (k=3) Step 3.1: Recalculate the gradient g1= 78.18 - 80.5 = -2.32, h1= 1, g2= 77.75 - 75.2 = 2.55, h2= 1, g3= 78.18 - 78.8 = -0.62, h3= 1, g4= 77.75 - 77.5 = 0.25, h4= 1; Step 3.2: Finding the optimal split Attempting to split the characteristic "vacuum degree" at 75 Pa: Left node: Vacuum degree ≤ 75, containing samples 1, 3 → Σg_left = -2.32 - 0.62 = -2.94, |I_left| = 2 Right node: Vacuum degree > 75, containing samples 2, 4 → Σg_right = 2.55 + 0.25 = 2.80, |I_right| = 2 Calculate the split gain: Gain = ½ [ (-2.94)² / (2+1) + (2.80)² / (2+1) - (0)² / (4+1) ] - 0 = ½ [ 8.64 / 3 + 7.84 / 3 ] = ½ [ 2.88 + 2.61 ] = 2.75 Step 3.3: Calculate the weights of the leaf nodes The weight of the left leaf node, w_left, is -(-2.94) / (2 + 1) = 0.98. The weight of the right leaf node, w_right, is -(2.80) / (2 + 1) = -0.93 Step 3.4: Update the predicted values Samples 1, 3: _i （3） = 78.18 + 0.1 × 0.98 = 78.28% Samples 2,4: _i （3） = 77.75 + 0.1 × (-0.93) = 77.66% Final prediction model After three rounds of iteration, the prediction function of the XGBoost model is: _xgb(X)=78.0+0.1×[ (X)+ (X)+ (X)] in: f1(X): Tree based on temperature splitting f2(X): Tree splitting based on aluminum content f3(X): Tree based on vacuum degree splitting Example of new furnace prediction Assuming the process parameters for the new furnace: X_new = [1605, 70, 0.036] # Temperature, Vacuum level, Aluminum content Prediction process: Through the first tree f1: Temperature 1605 ≥ 1595 → Falls onto the right leaf → (X_new) = 0.7 Through the second tree f2: [Al]%=0.036 ≥ 0.034 → Falls into the right leaf → (X_new) = 1.05 Through the third tree f3: Vacuum degree 70 ≤ 75 → Falls into the left leaf → (X_new) = 0.98 Final prediction: _pred = 78.0 + 0.1 × (0.7 + 1.05 + 0.98) = 78.0 + 0.1 × 2.73 = 78.0 + 0.273 = 78.27% Gradient Boosting Tree (GBDT) model 1. Nitrogen yield prediction function of GBDT model: _gbdt(X)= +η×Σ[m=1toM]T_m(X); in, Let η be the initial predicted value (the average nitrogen yield of the training set), η be the learning rate, and T_m(X) be the predicted value of the m-th tree.

[0097] 2. The pseudo-residual of the m-th iteration (for nitrogen yield): r_im=- 3. For the squared loss function : .

[0098] GBDT does not build a complex model all at once. Instead, it uses multiple simple trees, each of which learns the residuals from the previous prediction, to gradually approximate the true nitrogen yield.

[0099] The mathematical principle is as follows: 1. Initialize predicted values: ; Where N: number of training furnaces, y_i: actual nitrogen yield of the i-th furnace.

[0100] In the absence of other information, the most reasonable initial forecast for any new furnace run is the historical average yield.

[0101] 2. Iteratively construct the decision tree (m = 1, 2, ..., M). For the m-th iteration: Calculate pseudo residuals ; Where y_i: the actual nitrogen yield of the i-th furnace. r_im: The predicted value of the i-th furnace in the previous model round; r_im: The difference between the current predicted value and the actual value.

[0102] False residuals indicate which samples the current model has a large prediction bias on and that require focused correction.

[0103] 3. Train the new tree to fit the pseudo residuals. Train a new decision tree T_m using the feature vectors X_i and pseudo residuals r_im of all training samples.

[0104] The goal of tree splitting is to find the optimal features and split points that minimize the variance of pseudo-residuals within the split subsets.

[0105] 4. Update the prediction model ; Where η is the learning rate (usually 0.05-0.2), which controls the contribution of each tree.

[0106] 5. Final prediction function After M iterations, the final predicted nitrogen yield is: ; Specific example: The entire process of nitrogen yield prediction Suppose we have 3 batches of historical data as a training set: Feature vector: = [1620, 67, 0.035] = [1580, 85, 0.028] = [1600, 72, 0.041]; Model training process First iteration (m=1) Step 1.1: Initialization = (80.5 + 75.2 + 78.8) / 3 = 78.17% Initial prediction: for all furnace batches =78.17%; Step 1.2: Calculate the pseudo residuals =80.5-78.17=2.33 =75.2-78.17=-2.97 =78.8-78.17=0.63; Step 1.3: Training the first tree T1 Use features [ and pseudo residuals Training decision trees.

[0107] Assume the splitting rule for T1: If the temperature is ≥1590: Predicted value = (2.33 + 0.63) / 2 = 1.48 # Average residual of the high-temperature group; otherwise: Predicted value = -2.97# Low temperature group average residual Step 1.4: Update the predicted value (assuming η=0.1) = 78.17 + 0.1 × 1.48 = 78.32% # Furnace 1 = 78.17 + 0.1 × (-2.97) = 77.87% # Furnace 2 = 78.17 + 0.1 × 1.48 = 78.32% # Furnace 3; Second iteration (m=2) Step 2.1: Calculate the new pseudo residual = 80.5 - 78.32 = 2.18 = 75.2 - 77.87 = -2.67 = 78.8 - 78.32 = 0.48; Step 2.2: Training the second tree T2 T2 was trained using the new pseudo residual.

[0108] Assume that the splitting rule of T2 is based on the aluminum content: If [Al] ≥ 0.03: Predicted value = (2.18 + 0.48) / 2 = 1.33 # High Alumina Group otherwise: Predicted value = -2.67# Low Aluminum Group; Step 2.3: Update the predicted values = 78.32 + 0.1 × 1.33 = 78.45% = 77.87 + 0.1 × (-2.67) = 77.60% = 78.32 + 0.1 × 1.33 = 78.45%; Third iteration (m=3) Step 3.1: Calculate the new pseudo residuals = 80.5 - 78.45 = 2.05 = 75.2 - 77.60 = -2.40 = 78.8 - 78.45 = 0.35; Step 3.2: Training the third tree T3 Based on vacuum degree splitting: If the vacuum degree is ≤ 75: Predicted value = (2.05 + 0.35) / 2 = 1.20 # High vacuum group otherwise: Predicted value = -2.40 # Low vacuum group; Step 3.3: Update the predicted values = 78.45 + 0.1 × 1.20 = 78.57% = 77.60 + 0.1 × (-2.40) = 77.36% = 78.45 + 0.1 × 1.20 = 78.57%; Final prediction model After three rounds of iteration, the prediction function of the GBDT model is: = 78.17 + 0.1 × [ ]; Example of new furnace prediction Assuming there is a new furnace batch, the process parameters are as follows: X_new = [1610, 70, 0.038] # Temperature, Vacuum level, Aluminum content Prediction process: pass Temperature 1610 ≥ 1590 → Predicted value = 1.48 pass [Al] = 0.038 ≥ 0.03 → Predicted value = 1.33 pass Vacuum degree 70 ≤ 75 → Predicted value = 1.20 Final prediction: = 78.17 + 0.1 × (1.48 + 1.33 + 1.20) = 78.17 + 0.1 × 4.01 = 78.17 + 0.401 = 78.57% Random Forest Model 1. Nitrogen yield prediction function for random forests: ; Where T_b(X) is the predicted value of the b-th decision tree, which is constructed through Bootstrap sampling and random feature selection; 2. The prediction process for a single tree T_b(X): T_b(X)=Σ[j=1toJ]c_jb×I(X∈R_jb); Where R_jb is the j-th region (leaf node) of the b-th tree, c_jb=(1 / |R_jb|)×Σ[i∈R_jb]y_i is the average nitrogen yield of the training samples in this region, and I(·) is the indicator function.

[0109] Random forests construct diverse decision tree populations by introducing dual randomness (data sampling randomness + feature selection randomness), thereby reducing prediction variance through collective decision-making.

[0110] The mathematical principle is as follows: 1. Bootstrap sampling to build training set For a training set containing N batches, B distinct training subsets are constructed through sampling with replacement: For each tree T_b (b = 1, 2, ..., B): Randomly select N samples (with replacement) from the original N samples. Each Bootstrap sample set contains approximately 63.2% of the original samples. The remaining approximately 36.8% of the samples constitute "out-of-bag data" (OOB); Physical significance: Each tree is trained on a different subset of data, increasing the diversity of the model.

[0111] 2. Random selection of features When splitting at each node of each tree, m features (usually m ≈ √M) are randomly selected from all M features as candidate splitting features.

[0112] 3. Prediction process of a single decision tree The prediction value of the b-th decision tree for the feature vector X is: T_b(X)=Σ[j=1 to J]c_jb ×I(X ∈ R_jb); Detailed breakdown: R_jb: The feature space region corresponding to the j-th leaf node of the b-th tree. I(X ∈ R_jb): Indicator function, with a value of 1 when X falls into the region R_jb, and 0 otherwise; c_jb: The predicted value of leaf node j, calculated using the following formula: c_jb = (1 / |R_jb|) × Σ[i∈R_jb] y_i; That is, the average nitrogen yield of the training samples that fall into this leaf node.

[0113] 4. Random Forest Ensemble Prediction The final nitrogen yield prediction is a simple average of the predictions for all trees: .

[0114] Specific example: The entire process of nitrogen yield prediction Training data preparation Assume the training set contains historical data from 6 furnaces: Random forest construction process (assuming B = 3 trees) Construction of the first tree T1 Bootstrap sampling: 6 samples are randomly selected (with replacement). Selected samples: 1, 2, 2, 3, 5, 6 (Sample 2 was selected repeatedly, while sample 4 was not selected) train (Assuming the features are randomly selected as "temperature" and "[Al]%"): Assumption Split structure: Root node: Temperature ≥ 1595°C ├── Yes → Node A: [Al]% ≥ 0.037? │ ├── is → Leaf 1: Contains samples 1, 5 → c 11 =(80.5 + 79.8) / 2=80.15% │ └──No→ Leaf 2: Contains sample 3→c 21 = 78.8% └── No → Leaf 3: Contains samples 2, 2, 6 → c 31 = (75.2 + 75.2 + 74.5) / 3 = 75.0%; Construction of the second tree T2 Bootstrap sampling: Selected samples: 1, 3, 4, 4, 5, 6 (Sample 4 was selected repeatedly, while sample 2 was not selected) Training T2 (assuming features are randomly selected as "vacuum degree" and "temperature"): Assume a split structure for T2: Root node: Vacuum degree ≤ 75? ├── Yes → Node A: Temperature ≥ 1605? │ ├── is → Leaf 1: Contains sample 1 → c 12 =80.5% │ └── No → Leaf 2: Contains samples 3 and 5 → c 22 = (78.8 + 79.8) / 2 = 79.3% └── No → Leaf 3: Contains samples 4, 4, 6 → c 32 = (77.5 + 77.5 + 74.5) / 3 = 76.5%; Construction of the third tree T3 Bootstrap sampling: Selected samples: 2, 3, 4, 5, 6, 6 (Sample 6 was selected repeatedly, while sample 1 was not selected) Training T3 (assuming features are randomly selected as "[Al]%" and "vacuum degree"): Assume a split structure for T3: Root node: [Al]% ≥ 0.030? ├── is → Node A: Vacuum degree ≤ 80? │ ├── is → Leaf 1: Contains samples 3, 5 → c 13 = (78.8 + 79.8) / 2 = 79.3% │ └──No→Leaf 2: Contains sample 4 →c 23 = 77.5% └──No→Leaf 3: Contains samples 2, 6, 6 →c 33 = (75.2 + 74.5 + 74.5) / 3 = 74.73%; Example of new furnace prediction Assuming the process parameters for the new furnace: X_new=[1605,70,0.036] # Temperature, vacuum level, aluminum content; Step 1: Predict each tree independently T1 Prediction: Temperature 1605 ≥1595? → Yes [Al]% 0.036 ≥0.037? → No →Falling into Leaf 2→T1(X_new)=c 21 =78.8%; T2 Prediction: Vacuum degree 70≤75? → Yes Temperature ≥ 1605? → Yes →Falling into leaf 1→T2(X_new)=c12=80.5%; T3 Prediction: [Al]%0.036 ≥0.030?→Yes Vacuum degree 70≤80? → Yes →Falling into Leaf 1→T3(X_new)=c 13 =79.3%; Step 2: Random Forest Ensemble Prediction _rf(X_new)=(1 / 3)×[T1(X_new)+T2(X_new)+T3(X_new)] =(1 / 3)×[78.8+80.5+79.3] =(1 / 3)×238.6 =79.53% Establish meta-learner ensemble The three basic prediction results are used as input to the meta-learner (multilayer perceptron), which has a 50×25×1 structure. The input consists of the prediction results of the three basic models (XGBoost, GBDT, and random forest), and the output is the final nitrogen yield prediction.

[0115] Z = [ ] T Stacked feature vectors First hidden layer: h1 = ReLU(W1Z + b1) Second hidden layer: h2 = ReLU(W2h1 + b2) Output layer: _final=W3h2+b3 in: ∈ ^(50×3), ∈ ^50 ∈ ^(25×50), ∈ ^25 ∈ ^(1×25), ∈ Where W1 is a 50×3 matrix and b1 is a 50-dimensional vector; W2 is a 25×50 matrix and b2 is a 25-dimensional vector; W3 is a 1×25 matrix and b3 is a scalar.

[0116] The mathematical principle is as follows: 1. Stacked Feature Construction The prediction results of the three base models are combined into a stacked feature vector: Z = [ _xgb(X), _gbdt(X), _rf(X)]ᵀ ∈ 2. Forward Propagation Process of Meta-Learner First hidden layer computation (50 neurons): = ReLU( ), ∈ First layer weight matrix (50 rows × 3 columns) ∈ First layer bias vector ReLU(x) = max(0, x): Activation function Second hidden layer computation (25 neurons): = ReLU( ), Second layer weight matrix Second layer bias vector Output layer computation: _final = W3h2 + b3; 3. Complete mathematical expression Complete ensemble function for nitrogen yield prediction: _final(X)=W3×ReLU(W2×ReLU(W1×[ _xgb(X), _gbdt(X), _rf(X)] T +b1)+b2)+b3; Intelligent Combination Mechanism of Meta-Learners The three basic models have different predictive properties: Adaptive weighting of meta-learners Meta-learners automatically learn context-aware weight assignments through neural networks: Effective weights = f(Base model prediction, prediction consistency, historical performance pattern) Specific example: The entire process of a meta-learner predicting nitrogen yield. Basic model prediction results Assuming that for a certain heat process parameter X, the prediction results of the three basic models are as follows: Stacked feature vectors Z = [78.5, 77.8, 78.2] T Meta-learner prediction process Step 1: Calculation of the first hidden layer Assume the trained weights and biases are as follows: = [[0.12, -0.08, 0.15], # Weights of neuron 1 [0.05, 0.20, -0.10], # Weights of neuron 2 [-0.15, 0.12, 0.08], # Weights of neuron 3 ... # A total of 50 neurons [0.10, -0.05, 0.18]] # Weights of neuron 50 = [0.02, -0.01, 0.05, ..., 0.03]ᵀ # 50-dimensional bias Calculate the activation of the first hidden layer: a1 = W1Z + b1 = [[0.12, -0.08, 0.15] · [78.5, 77.8, 78.2] T + 0.02, [0.05, 0.20, -0.10] · [78.5, 77.8, 78.2] T - 0.01, ...] Calculate the first neuron: = 0.12×78.5 + (-0.08)×77.8 + 0.15×78.2 + 0.02 = 9.42 - 6.224 + 11.73 + 0.02 = 14.946 Calculate the second neuron: = 0.05×78.5 + 0.20×77.8 + (-0.10)×78.2 - 0.01 = 3.925 + 15.56 - 7.82 - 0.01 = 11.655 ...Continue calculating all 50 neurons Applying the ReLU activation function: h1=ReLU(a1)=[max(0,14.946),max(0,11.655), ...] T =[14.946,11.655, ...] T Step 2: Calculation of the second hidden layer Assume the second layer parameters: W2∈ 25X50 , b2∈ 25 Calculate the second hidden layer: a2 = W2h1 + b2 h2 = ReLU(a2) Step 3: Output layer calculation Assume the output layer parameters are: W3 = [0.08, -0.12, 0.15, ..., 0.06] # 1×25 weight matrix b3 = 0.35 Final prediction calculation: _final = W3h2+ b3 = 0.08×h 21 + (-0.12)×h 22 + 0.15×h 23 + ... + 0.06×h 225 + 0.35 Assuming the calculation yields: _final=78.15% Furthermore, the meta-learner can also assess prediction confidence based on the degree of dispersion of the base model's predictions: Overall confidence assessment (based on base model predictions, historical accuracy, and feature anomalies): "Multi-dimensional confidence assessment" # 1. Consistency confidence level (40% weight) Predicted mean = np.mean(baseline model prediction) Prediction standard deviation = np.std(baseline model prediction) Coefficient of variation = Predicted standard deviation / max(Predicted mean, 1e-6) # Avoid division by zero Consistency confidence score = max(0, 1 - 3) Coefficient of variation # 2. Historical accuracy confidence level (30% weight) # Weighted average of each model's performance on the validation set Model weights = np.array([0.35, 0.33, 0.32]) # Historical accuracy of XGBoost, GBDT, and RF Weighted standard deviation = np.sqrt(np.average((baseline model predicted value - predicted mean)) 2, weights = model weights) Historical accuracy confidence = max(0, 1 - weighted standard deviation / 5) #Assume 5% is the standard threshold # 3. Feature Anomaly Confidence (30% weight) # Based on the similarity between input features and training data distribution Confidence of feature anomaly = 1 - min(1.0, feature anomaly) # Overall confidence level Overall confidence level = (0.4) Consistency confidence level +0.3 Historical accuracy confidence level +0.3 (Feature anomaly confidence level) Return the overall confidence level; Practical application examples Confidence calculation under different conditions Scenario 1: Highly Consistent Predictions Predicted values ​​= [78.5, 78.3, 78.4] # Highly consistent Mean = 78.4, Standard Deviation = 0.08 Coefficient of variation = 0.08 / 78.4 ≈ 0.00102 # Comprehensive assessment (assuming normal characteristics) Consistency confidence level = 1 - 3 × 0.00102 = 0.9969 Historical accuracy confidence level = 0.98 Confidence level of anomaly characteristic = 0.95 Overall confidence level = 0.4 × 0.9969 + 0.3 × 0.98 + 0.3 × 0.95 = 0.977 (97.7%) Scenario 2: Significantly divergent predictions Predicted values ​​= [75.0, 80.0, 78.0] # Disagreement exists Mean = 77.67, Standard Deviation = 2.08 Coefficient of variation = 2.08 / 77.67 ≈ 0.0268 Consistency confidence level = 1 - 3 × 0.0268 = 0.9196 Historical accuracy confidence level = 0.85 # Accuracy expected to decrease due to discrepancies. Feature anomaly confidence level = 0.90 Overall confidence level = 0.4 × 0.9196 + 0.3 × 0.85 + 0.3 × 0.90 = 0.892 (89.2%) Predicted values ​​= [75.0, 80.0, 78.0] # Disagreement exists Mean = 77.67, Standard Deviation = 2.08 Coefficient of variation = 2.08 / 77.67 ≈ 0.0268 Consistency confidence level = 1 - 3 × 0.0268 = 0.9196 Historical accuracy confidence level = 0.85 # Accuracy expected to decrease due to discrepancies. Feature anomaly confidence level = 0.90 Overall confidence level = 0.4 × 0.9196 + 0.3 × 0.85 + 0.3 × 0.90 = 0.892 (89.2%) Scenario 3: Extreme Disagreement Prediction Predicted values ​​= [70.0, 82.0, 76.0] # Significant divergence Mean = 76.0, Standard Deviation = 4.90 Coefficient of variation = 4.90 / 76.0 ≈ 0.0645 Consistency confidence level = 1 - 3 × 0.0645 = 0.8065 Historical accuracy confidence level = 0.70 Feature anomaly confidence level = 0.80 Overall confidence level = 0.4 × 0.8065 + 0.3 × 0.70 + 0.3 × 0.80 = 0.773 (77.3%) Confidence Level Classification and Coping Strategy Table The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for manufacturing high-performance mold steel, comprising the following components: C: 0.30%-0.38%, Si: 1.00%-1.20%, Mn: 1.10%-1.30%, Cr: 3.0%-3.4%, Mo: 3.3%-3.5%, V: 1.00%-1.30%, Co: 0.80%-1.00%, N: 0.02%-0.05%, Nb: 0.03%-0.06%, Sn: 0.10%-0.15%, Ba: 0.05%-0.10%, Zn: 0.10%-0.20%, P: ≤0.012%, S: ≤0.008%, balance Fe and unavoidable impurities; In mold steel, the ratio of the sum of Ba and Zn content to Sn content is controlled at 1.8-2.2; the ratio of Mo content to Co content is controlled at 3.5-4.

2. Its features are, The manufacturing method includes the following steps: The primary refining process uses high-quality low-phosphorus and low-sulfur scrap steel and pig iron, which are smelted in an electric arc furnace, dephosphorized by oxygen blowing, with the final carbon content controlled at ≥0.10% and the tapping temperature at ≥1630℃. Refining is carried out in an LF furnace to produce high-basicity white slag. Fe-Cr, Fe-Mo, Ni, and Fe-V are added for rough composition adjustment. The S content is reduced to ≤0.008% by strong stirring. The temperature of the molten steel is heated to 1580-1600℃. Before tapping, Fe-Nb alloy is added and stirred thoroughly to make it uniformly dissolved. In the VD furnace refining process, the vacuum pump is started, and the vacuum level is increased to ≤0.5 mbar within 15-20 minutes and maintained for 15-20 minutes to deeply remove H content to ≤1.5 ppm and O content to ≤15 ppm. Under vacuum, the nitrogen content is adjusted by adding manganese nitride or chromium-nitrogen alloy. Sn, Ba, and Zn cored wires are added using wire feeding technology, and the easily oxidized elements Sn, Ba, and Zn are finely adjusted. Vacuum treatment: Start the vacuum pump and raise the vacuum level to ≤0.5mbar within 5-8 minutes, and maintain it for 5-8 minutes; Continuous casting employs a long nozzle and argon seal protection for pouring to prevent secondary oxidation. The superheat of the molten steel is strictly controlled at 25-35℃. Electromagnetic stirring combined with a weak cooling mode is used to obtain a continuous casting billet with uniform composition and small central porosity, which is then cast into a rectangular billet of the required size. Slab heating, 1180℃-1220℃, holding time ≥2 hours / ton; Rolling: the initial rolling temperature is ≤1150℃, the first few passes are carried out at 1050℃ or above to fully break up the as-cast structure, and the final 3-5 passes are rolled in the range of 950℃-850℃, with the final rolling temperature controlled at 850℃-880℃. After cooling, the steel plate is rapidly cooled to 550℃-600℃ using the ACC system at a cooling rate of 15-25℃ / s, and then slowly cooled to room temperature in a slow cooling pit. Spheroidizing annealing involves heating to 780℃-800℃ at a rate of ≤100℃ / h, holding at that temperature for 4-6 hours, then furnace cooling to 690℃-710℃, isothermal holding for 4-6 hours, and finally furnace cooling to below 500℃ before air cooling. Quenching is performed in an air furnace at 1050℃-1080℃, with the holding time calculated based on the effective thickness of the workpiece, at 1.0-1.5 min / mm. After quenching, the workpiece is cooled in a vacuum high-pressure gas quenching furnace with a quenching pressure of 8-12 bar of high-purity nitrogen. Cyclic cryogenic treatment: After quenching, the workpiece is immediately transferred to a cryogenic chamber and cooled to -150℃ to -196℃ at a rate of ≤5℃ / min. It is then held at this temperature for 1-3 hours and then removed to be allowed to return to room temperature in the air. This process is repeated twice. Tempering is performed in three stages, with tempering temperatures ranging from 560℃ to 600℃, each stage lasting 2.5 to 3.5 hours, followed by air cooling.

2. The manufacturing method according to claim 1, characterized in that: In the VD furnace refining step, the amount of nitrogen added is controlled by a nitrogen yield prediction model. The method for establishing the nitrogen yield prediction model includes the following steps: The data acquisition step involves real-time acquisition of multi-source process data during the VD refining process. The multi-source process data includes steel composition data, temperature data, vacuum degree data, stirring parameter data, and nitrogen alloying operation data. The feature engineering step involves constructing a feature set for predicting nitrogen yield based on the multi-source process data. The feature set includes basic features, derived features, and interactive features. The model training steps involve training an ensemble machine learning model using historical data, with the feature set as input and nitrogen yield as output. The real-time prediction step involves inputting the real-time acquired process parameters into a trained integrated machine learning model during the VD refining process to predict the nitrogen yield of the current furnace. The control output step calculates and outputs the recommended amount of nitrogen alloy addition based on the predicted nitrogen yield.

3. The manufacturing method according to claim 2, characterized in that: In the data acquisition step, the molten steel composition data includes at least the real-time contents of C, Si, Mn, Cr, Mo, V, Al, O, and S; the temperature data includes the LF outlet temperature, VD start temperature, and nitrogen alloying temperature; the vacuum degree data includes the vacuum degree change curve and holding time; the stirring parameter data includes the bottom-blown argon flow rate and stirring mode; and the nitrogen alloying operation data includes the timing, method, and batch of nitrogen alloy addition.

4. The manufacturing method according to claim 2, characterized in that: The derived characteristics include a nitrogen solubility parameter, which is calculated as: f(T,P,composition)=K×exp(ΔH / RT)× Where T is the temperature of the molten steel, P is the vacuum degree, R is the gas constant, and K and ΔH are constants related to the composition of the molten steel; The deoxidation degree index is calculated based on the [O] and [Al] content in the molten steel. Historical average alloy yield, calculated based on the historical average yield of the same steel grade; Process stability index is calculated based on the parameter fluctuations of preceding processes.

5. The manufacturing method according to claim 3, characterized in that: The interaction feature includes a temperature-vacuum interaction term: T×log(P), where T is the temperature of the molten steel and P is the vacuum degree; Alloy type - characteristics of the combination of addition timing; Stirring intensity - vacuum degree synergy coefficient.

6. The manufacturing method according to claim 3, characterized in that: The ensemble machine learning model in the model training step adopts a two-layer structure; The first layer includes multiple heterogeneous base prediction models, which are selected from at least two of XGBoost, gradient boosting tree and random forest; The second layer is a meta-learner, whose input is the output of each basic prediction model in the first layer. The meta-learner uses a multilayer perceptron neural network.

7. The manufacturing method according to claim 3, characterized in that: The model training steps employ a time-series cross-validation method, dividing the training and validation sets according to the furnace time sequence to ensure the time sequence of model validation.

8. The manufacturing method according to claim 3, characterized in that: It also includes an online learning step for the model, collecting nitrogen yield data from actual production and its corresponding process parameters; when the amount of collected data reaches a predetermined threshold, the integrated machine learning model is incrementally learned or retrained using new data; Update the model parameters to adapt to changes in process conditions.

9. The manufacturing method according to claim 3, characterized in that: In the control output step, the recommended amount of nitrogen alloy added is calculated using the following formula: Recommended addition amount = (target nitrogen content - current nitrogen content) / (predicted yield / 100) Simultaneously, the control output step also outputs a confidence assessment of the prediction result, which is based on the similarity calculation between the input features and the training data distribution.

Images