Thick steel plate heat treatment performance prediction method and system based on gradient sensing mechanism and physical constraint

Abstract

The invention relates to the technical field of material science and artificial intelligence, and discloses a thick steel plate heat treatment performance prediction method and system based on a gradient sensing mechanism and physical constraints, and the method comprises the steps: obtaining chemical component parameters and heat treatment process parameters of a steel plate; and predicting the heat treatment performance of the steel plate by using the constructed gradient sensing neural network model. Accurate prediction of the heat treatment performance of the thick steel plate is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of material science and artificial intelligence, and particularly relates to a thick steel plate heat treatment performance prediction method based on a gradient perception mechanism and physical constraints. BACKGROUND

[0002] As a key structural material for marine engineering equipment, the accurate prediction of the mechanical properties of thick steel plates is of great significance to ensure engineering safety. Due to the large thickness specification (usually 50-100 mm), low-alloy high-strength steel plates for marine engineering have a significant thickness-direction temperature gradient during quenching-tempering heat treatment, resulting in a complex spatial distribution of properties. This property gradient not only affects the evaluation of structural load-carrying capacity, but also directly relates to the distribution of welding residual stress and fatigue life prediction.

[0003] The traditional thick steel plate performance prediction methods have limitations in dealing with gradient distribution problems:

[0004] (1) Empirical formula method is difficult to capture the continuous change in the thickness direction;

[0005] (2) Finite element numerical simulation can obtain detailed performance distribution, but the calculation cost is high and depends on the accurate determination of material parameters;

[0006] (3) Continuous cooling phase transformation kinetics model has limited accuracy in dealing with complex composition systems;

[0007] (4) Ultrasonic testing can detect full-thickness defects, but cannot directly measure mechanical property parameters;

[0008] (5) Hardness testing is limited to surface measurement, and the standard error of the hardness-strength empirical formula is as high as ±100-112 MPa;

[0009] (6) The detection depth of magnetic powder and eddy current testing is only 2-18 mm, and it is powerless for internal performance evaluation of 50-100 mm thick plates;

[0010] (7) Destructive testing usually takes discrete samples at 1 / 4t and 1 / 2t positions, and the sampling volume is only 0.001% of the total volume of the plate, which is difficult to represent the continuous performance gradient.

[0011] Studies have shown that the surface cooling rate of thick plates during water quenching can reach 89°C / s, while the center is only 8°C / s, resulting in a change of 13% in the tensile strength of a 130 mm thick Q690 steel plate from the surface to the center. Even the Z-direction tensile test specified in the international standard ASTM A770 is only applicable to the thickness range of 20-80 mm, and only provides a single reduction of area index. There is a clear gap between these detection methods and the high-resolution, continuous performance distribution characterization capability required by engineering, which has become a technical bottleneck for thick plate quality control.

[0012] Machine learning techniques bring new opportunities for material performance prediction. Deep neural networks have made significant progress in predicting the mechanical properties of steel materials. However, existing research has the following shortcomings:

[0013] (1) Traditional machine learning models (such as random forests, XGBoost, etc.) cannot effectively capture the performance gradient characteristics in the thickness direction, and can only predict single-point performance or average performance;

[0014] (2) Simple neural networks lack location awareness and are difficult to model spatially continuous changes in the thickness direction;

[0015] (3) Pure data-driven "black box" models often ignore the basic physical laws of materials science, and may produce prediction results that violate metallurgical common sense, such as yield strength greater than tensile strength, and unreasonable combinations of high strength and high plasticity;

[0016] (4) Existing physical information neural network (PINN) research mostly targets a single physical process, and how to build a comprehensive physical constraint system suitable for thick plate gradient performance prediction is still an open problem;

[0017] (5) Lack of quantitative evaluation of prediction uncertainty in engineering applications, making it difficult to provide reliability reference for decision-making. SUMMARY

[0018] In view of the above shortcomings in the prior art, the present application provides a thick steel plate heat treatment performance prediction method and system based on gradient perception mechanism and physical constraints to solve the technical problem that traditional methods are difficult to accurately predict the performance gradient distribution in the thickness direction of thick steel plates.

[0019] In order to achieve the above invention purpose, the technical scheme adopted by the present application is:

[0020] A thick steel plate heat treatment performance prediction method based on gradient perception mechanism and physical constraints, comprising the following steps:

[0021] Obtaining the chemical composition parameters and heat treatment process parameters of the steel plate;

[0022] Constructing a gradient perception neural network model; the gradient perception neural network model includes a shared feature extraction network, a multi-head gradient attention network, a location-specific encoding network, a gradient consistency network, and an attribute prediction network;

[0023] Using the shared feature extraction network to extract high-level shared features from the input features composed of chemical composition parameters and heat treatment process parameters;

[0024] The multi-head gradient attention network and the position-specific encoding network use a multi-head gradient attention mechanism and position encoding to explicitly model the thickness direction performance gradient features, and a gradient bias matrix is used to strengthen the exchange of information between adjacent positions.

[0025] The gradient consistency network is used to learn the relationship between positions, and the thickness direction performance gradient features are output after gradient correction.

[0026] The attribute prediction network is used to output the heat treatment performance prediction results based on the thickness direction performance gradient features after gradient correction.

[0027] In an implementation, the chemical composition parameters of the steel plate include alloy element content parameters such as C, Mn, Si, Cr, Ni, and Mo.

[0028] The heat treatment process parameters include plate thickness, austenitizing temperature, holding time, surface cooling rate, cooling rate distribution along the thickness direction of the steel plate, tempering temperature, and tempering time. The cooling rate distribution along the thickness direction of the steel plate is calculated based on the surface cooling rate, and the calculation formula is as follows:

[0029]

[0030] where v(z) is the cooling rate at the thickness direction coordinate z, v s is the surface cooling rate of the steel plate, erfc is the complementary error function, z is the depth coordinate from the surface of the steel plate to the core direction, a is the thermal diffusivity of the steel, t c is the effective cooling time, Bi is the Biot number, and is the correction coefficient.

[0031] In an implementation, in addition to the chemical composition parameters and the heat treatment process parameters of the steel plate, derived features calculated based on the chemical composition parameters and the heat treatment process parameters of the steel plate are also included, including carbon equivalent, hardenability index, cooling intensity, tempering parameter, and carbon-manganese interaction term.

[0032] In an implementation, the multi-head gradient attention network uses 8 parallel attention heads to learn different gradient patterns, including surface hardening layer features, transition zone features, and core softening zone features.

[0033] In an implementation, the multi-head gradient attention network adds a gradient bias matrix to the calculation of the scaled dot-product attention in each attention head, so that the attention weight gives priority to the information of adjacent positions; the multi-head gradient attention network outputs the final attention features through concatenation and linear transformation.

[0034] ​In one implementation, the location-specific encoding network discretizes the thickness direction into multiple standard locations for encoding; and employs a corresponding number of location-specific encoders to fuse the features output by the multi-head gradient attention network with the corresponding location codes, processing the features of the corresponding thickness locations separately. Further, the location-specific encoding network discretizes the thickness direction into five standard locations for location encoding, including the surface, 1 / 8 depth, 1 / 4 depth, 3 / 8 depth, and core; and employs five independent location-specific encoders to process the features of the five thickness locations separately based on the attention features output by the multi-head gradient attention network and the location codes.

[0035] In one feasible implementation, the property prediction network includes a microstructure type classification head, a yield strength regression head, a tensile strength regression head, an elongation regression head, and an impact toughness regression head; the property prediction network outputs heat treatment performance prediction results based on the gradient-corrected thickness-direction performance gradient features, including:

[0036] The tissue type is obtained based on the microstructure prediction results;

[0037] The predicted mechanical properties are inversely normalized to obtain the true values; the predicted mechanical properties include predicted values ​​for yield strength, tensile strength, elongation, and impact toughness.

[0038] Output the predicted values ​​of microstructure type, yield strength, tensile strength, elongation and impact toughness at the corresponding thickness location.

[0039] In one possible implementation, the gradient-aware neural network model is trained by constructing a total loss function using a weighted sum of data fitting loss, gradient consistency loss, and metallurgical constraint loss; wherein the data fitting loss includes microstructure mean square error loss and mechanical property SmoothL1 loss; the gradient consistency loss includes gradient direction consistency loss and gradient smoothness loss; and the metallurgical constraint loss includes yield / tensile strength ratio constraint loss, strength-ductility balance constraint loss, monotonicity constraint loss, property range constraint loss, and gradient magnitude constraint loss.

[0040] In one possible implementation, the gradient direction consistency loss uses cosine similarity measurement and includes gradient consistency loss for strength-type attributes and gradient consistency loss for plasticity-type attributes; the gradient smoothness loss is calculated by penalizing the second-order gradient; the strength-ductility balance constraint is calculated based on the Considère criterion; the monotonicity constraint loss includes strength monotonicity constraint and plasticity monotonicity constraint, calculated based on the linear correction function ReLU; the attribute range constraint loss includes yield strength range constraint loss, tensile strength range constraint loss, and impact toughness range constraint loss, all of which are calculated based on the linear correction function ReLU; and the gradient magnitude constraint loss is calculated based on the linear correction function ReLU.

[0041] In one feasible approach, the method for predicting the heat treatment performance of thick steel plates based on gradient sensing mechanisms and physical constraints further includes uncertainty quantification, specifically including:

[0042] Multiple independently initialized gradient-aware neural network models are trained using a deep ensemble method, and the uncertainty is estimated by the variance of the prediction results.

[0043] Alternatively, a Bayesian neural network method can be used to quantify uncertainty by modeling the probability distribution of the weights;

[0044] Alternatively, the MCDropout method can be used to maintain Dropout activation during inference and estimate uncertainty through multiple forward propagations of the prediction variance.

[0045] Furthermore, the use of deep integration methods for uncertainty quantification specifically includes:

[0046] Multiple independently initialized gradient-aware neural network models are trained using a deep ensemble method;

[0047] Each gradient-aware neural network model is used to make predictions, resulting in multiple sets of prediction results.

[0048] The mean of multiple prediction results is calculated as the final prediction value;

[0049] The standard deviation of multiple sets of prediction results is calculated as the prediction uncertainty.

[0050] The present invention also provides a system for predicting the heat treatment performance of thick steel plates based on gradient sensing mechanism and physical constraints for implementing the above method, comprising:

[0051] The data generation module is used to generate the chemical composition parameters and heat treatment process parameters of the steel plate;

[0052] The gradient-aware neural network model comprises a shared feature extraction network, a multi-head gradient attention network, a location-specific encoding network, a gradient consistency network, and an attribute prediction network. The shared feature extraction network extracts high-level shared features from input features composed of chemical composition parameters and heat treatment process parameters. The multi-head gradient attention network and location-specific encoding network explicitly model the thickness-direction performance gradient features using a multi-head gradient attention mechanism and location encoding, and enhance information exchange between adjacent locations through a gradient bias matrix. The gradient consistency network learns the interrelationships between locations and outputs gradient-corrected thickness-direction performance gradient features. Finally, the attribute prediction network outputs heat treatment performance prediction results based on the gradient-corrected thickness-direction performance gradient features.

[0053] In one possible implementation, the thick steel plate heat treatment performance prediction system based on gradient sensing mechanism and physical constraints further includes:

[0054] Bayesian models are used for uncertainty quantification.

[0055] The deep ensemble model is obtained by integrating multiple gradient-aware neural network models configured in parallel.

[0056] In one possible implementation, the thick steel plate heat treatment performance prediction system based on gradient sensing mechanism and physical constraints further includes a data verification unit and a physical constraint verification unit:

[0057] The data verification unit is used to verify whether the input chemical composition parameters and heat treatment process parameters are within the preset valid range before prediction.

[0058] Physical constraint verification is used to verify whether the prediction results meet the physical constraints.

[0059] The present invention also provides a web prediction service interface for loading the thick steel plate heat treatment performance prediction system based on gradient perception mechanism and physical constraints.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] (1) Accurate prediction of the performance gradient in the thickness direction of thick steel plates was achieved, with an R² value of 0.9439 and a gradient direction prediction accuracy of 100%;

[0062] (2) The physical constraint satisfaction rate exceeds 94%, ensuring that the prediction results conform to the basic laws of materials science;

[0063] (3) Compared with baseline methods such as LightGBM and XGBoost, it significantly outperforms in overall prediction accuracy and successfully captures the performance gradient features in the thickness direction.

[0064] (4) Millisecond-level inference time (approximately 6.1ms per sample) and lightweight model size (approximately 1.2MB) meet the needs of industrial real-time monitoring;

[0065] (5) Provides a quantitative assessment of prediction uncertainty, with a correlation of 0.262 between prediction uncertainty and absolute error, enhancing the reliability of engineering applications;

[0066] (6) The model version management and checkpoint saving mechanism ensure the traceability of the training process and the automatic selection of the best model. Attached Figure Description

[0067] Figure 1 This is a schematic diagram of the process for predicting the heat treatment performance of thick steel plates based on gradient sensing mechanism and physical constraints according to the present invention.

[0068] Figure 2 This is a schematic diagram of the gradient-aware neural network model structure;

[0069] Figure 3 This is a diagram illustrating the working principle of the multi-head gradient attention mechanism.

[0070] Figure 4 This is a schematic diagram illustrating the orthogonality of positional coding.

[0071] Figure 5 This is a structural diagram of the physical constraint loss function;

[0072] Figure 6 A diagram illustrating the monitoring and checkpoint saving process during training;

[0073] Figure 7 The following is a comparison chart of gradient prediction performance; where (a) is the gradient distribution prediction result of yield strength, (b) is a schematic diagram of probability density gradient distribution, (c) gradient direction accuracy, (d) physical constraint satisfaction rate, and (e) position error.

[0074] Figure 8 The diagram shows the results of the physical constraint verification; where (a) is YS / TS, (b) is the strength-ductility balance, (c) is the cooling rate, (d) is the gradient monotonicity, (e) is the Hall-Petch relationship, and (f) is the physical score.

[0075] Figure 9 A flowchart for model version management;

[0076] Figure 10 This is a flowchart for data validation. Detailed Implementation

[0077] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0078] Example 1

[0079] like Figure 1 As shown, an embodiment of the present invention provides a method for predicting the heat treatment performance of thick steel plates based on gradient sensing mechanism and physical constraints, comprising the following steps S1 to S6:

[0080] S1. Obtain the chemical composition parameters and heat treatment process parameters of the steel plate;

[0081] S2. Construct a gradient-aware neural network model; the gradient-aware neural network model includes a shared feature extraction network, a multi-head gradient attention network, a location-specific encoding network, a gradient consistency network, and an attribute prediction network;

[0082] S3. Utilize a shared feature extraction network to extract advanced shared features from the input features composed of chemical composition parameters and heat treatment process parameters;

[0083] S4. Employ multi-head gradient attention network and location-specific coding network to explicitly model thickness direction performance gradient features using multi-head gradient attention mechanism and location coding, and enhance information exchange between adjacent locations through gradient bias matrix;

[0084] S5. Utilize a gradient consistency network to learn the interrelationships between locations and output the thickness direction performance gradient features after gradient correction.

[0085] S6. The attribute prediction network is used to output the heat treatment performance prediction results based on the gradient-corrected thickness direction performance gradient features.

[0086] In step S1, the chemical composition parameters of the steel plate are obtained, including the content parameters of alloying elements such as C, Mn, Si, Cr, Ni, and Mo; the heat treatment process parameters of the steel plate are obtained, including plate thickness, austenitizing temperature, holding time, surface cooling rate, cooling rate distribution along the thickness direction of the steel plate, tempering temperature, and tempering time. The cooling rate distribution along the thickness direction of the steel plate is calculated based on the surface cooling rate using a physical model.

[0087] The cooling rate distribution employs an enhanced model based on an error function:

[0088] ;

[0089] Where v(z) is the cooling rate at coordinate z in the thickness direction, v s Let be the surface cooling rate of the steel plate, erfc be the residual error function, z be the depth coordinate measured from the surface of the steel plate towards the core, α be the thermal diffusivity of the steel, and t be the thermal diffusivity of the steel. c Bi represents the effective cooling time, and Bi is the Biot number. This is a correction factor.

[0090] The advantage of this model is that it simultaneously considers the error function characteristics of transient heat transfer and the exponential decay characteristics of steady-state heat transfer.

[0091] Step S1 also includes derived characteristics calculated based on the chemical composition parameters and heat treatment process parameters of the steel plate, including carbon equivalent, hardenability index, cooling intensity, tempering parameters, and carbon-manganese interaction term. The specific calculation method is as follows:

[0092] Carbon equivalent: Wherein, C, Mn, Cr, Mo, and Ni represent the content of the corresponding elements;

[0093] Hardenability Index: Wherein, C, Mn, Cr, and Mo represent the content of the corresponding elements;

[0094] Cooling intensity: , where v s t represents the surface cooling rate, and t represents the plate thickness.

[0095] Tempering parameters: , among which, T t This refers to the tempering temperature;

[0096] Carbon-manganese interaction term: Where C and Mn represent the content of the corresponding elements, respectively.

[0097] In step S2, the constructed gradient-aware neural network model (GANN) is configured as follows: Figure 2 As shown in Table 1.

[0098] Table 1 Gradient-aware neural network model configuration

[0099]

[0100] The shared feature extractor uses a fully connected network with layer-wise dimensionality reduction. The input dimension is 17, including 12 original features and 5 derived features. It extracts high-level feature representations through three-layer transformation. The network structure is 512→256→128. LayerNorm layer normalization and Dropout regularization are applied after each layer. The Dropout rate is 0.15, and the GELU function is used as the activation function.

[0101] Multi-head gradient attention networks, such as Figure 3 As shown, eight parallel attention heads are used, each with a key dimension d. k =64, each head learns different relationships between positions through an independent parameter matrix. Guided by the gradient bias mechanism, each head may develop specialized attention capabilities to gradient patterns in different regions such as the surface hardening layer, transition region, and core softening region. The final attention output is obtained through splicing and linear transformation:

[0102] ;

[0103] Where MultiHead is the final attention feature, X is the input feature, Concat is the concatenation function, and head is the head function. l W represents the independent output feature of the l-th attention head. o This is the linear transformation weight matrix.

[0104] Multi-head gradient attention networks introduce gradient bias terms on top of standard scaled dot product attention. The design of the gradient bias matrix is ​​based on the continuity principle in materials science.

[0105] .

[0106] in, Let B be the learnable matrix; i,j=0,1,2,3,4, and B be the gradient bias matrix. grad The row and column indices correspond to the specified thickness positions on the steel plate, respectively.

[0107] This design ensures that attention weights prioritize information from neighboring locations during computation, thereby better capturing gradient features.

[0108] Therefore, the formula for calculating each attention head is:

[0109] ;

[0110] Where Q, K, and V are the query matrix, key matrix, and value matrix, respectively, and d k Let be the dimension of the key, and λ = 0.1 be the gradient bias coefficient. Each head has an independent projection matrix. Learn different linear transformations.

[0111] The location-specific encoding network, based on the output of the multi-head gradient attention network, uses 5 independent encoders to process features at 5 thickness locations. Each encoder consists of two fully connected layers with a structure of 128→64→32 and a dropout rate of 0.1.

[0112] The position encoder adopts the position encoding concept of the Transformer architecture, and is designed with a position encoding mechanism suitable for discrete sampling in the thickness direction. Position encoding uses a combination of sine and cosine functions to encode five positions in the thickness direction (surface, 1 / 8 depth, 1 / 4 depth, 3 / 8 depth, and core), ensuring that the encoded vectors of different positions are unique and distinguishable, and that the encoding of adjacent positions is continuous, conforming to the physical law of gradual performance changes. The position encoding formula is:

[0113] ;

[0114] ;

[0115] in Indicates the normalized thickness locations, corresponding to the surface, 1 / 8 depth, 1 / 4 depth, 3 / 8 depth, and core, respectively, d model Let q represent the total dimension of the encoded vector, and q represent the dimension index of the encoded vector.

[0116] Each position-specific encoder fuses the positional encoding with the features output by the multi-head gradient attention network, and obtains the corresponding thickness positional features through two fully connected layers. Thus, the initial thickness-direction performance gradient features are obtained through the position-specific encoding network.

[0117] like Figure 4 As shown, the positional encoding orthogonality analysis indicates that all diagonal elements of the matrix are 64.00, and the inner product value between adjacent positions gradually decreases from 62.09 to 48.59 at distant positions, showing a clear decreasing pattern of spatial correlation.

[0118] The gradient consistency network receives 160 dimensions of features from all locations. It learns the relationships between these locations through a three-layer fully connected network (160→256→128→25), outputting a 25-dimensional gradient correction vector. The final predicted thickness-direction performance gradient feature output is y. final for:

[0119]

[0120] Among them, y initial The initial thickness-direction performance gradient features represent the preliminary predictions made by the location-specific coding network based on the input features (chemical composition, process parameters, and location coding); y gradientcorrection This is the gradient correction vector, with a coefficient of 0.05 determined through optimization on the validation set. It is used to ensure the effectiveness of gradient correction while avoiding instability caused by over-correction.

[0121] The attribute prediction heads are designed separately for different attributes, including a microstructure type classification head, a yield strength regression head, a tensile strength regression head, an elongation regression head, and an impact toughness regression head. Microstructure prediction uses a classification head (32→16→4) and Softmax output, corresponding to four microstructure types: 1-ferrite + pearlite, 2-bainite + ferrite, 3-martensite, and 4-martensite + bainite. Mechanical property prediction uses a regression head (32→16→1) to predict yield strength, tensile strength, elongation, and impact toughness respectively.

[0122] In an optional embodiment of the present invention, when training the gradient-aware neural network model, 20,000 physically consistent virtual samples are generated based on physical metallurgical principles, covering a wide range of composition and process parameters. The chemical composition range is based on typical marine engineering steel specifications, referencing DNV-GL specifications and China Classification Society standards: C (0.05-0.40wt%), Mn (0.30-1.40wt%), Si (0.10-0.50wt%), Cr (0.20-0.40wt%), Ni (0.05-0.30wt%), Mo (0.05-0.20wt%).

[0123] The selection range of process parameters is based on industrial production practice: plate thickness of 30-100mm covers the main specifications of steel for marine engineering; austenitizing temperature of 750-980°C includes sub-temperature quenching and conventional quenching processes; holding time of 0.5-4.0h considers the heat homogenization requirements of plates of different thicknesses; surface cooling rate of 0.3-80°C / s covers various cooling methods from air cooling to water quenching; tempering temperature of 0-650°C and tempering time of 0-3.0h reflect the requirements for different strength and toughness matching.

[0124] This embodiment then performs data preprocessing, calculating 5 derived features, and performing Z-score standardization on the input features consisting of 12 original features and 5 derived features. The calculation formula is as follows: , where μ is the mean of the training set and σ is the standard deviation of the training set.

[0125] like Figure 5 As shown, this embodiment encodes materials science knowledge into the training of the neural network, including: data fitting loss (mean squared error loss for microstructure and SmoothL1 loss for mechanical properties); gradient consistency loss (including gradient direction consistency loss and gradient smoothness loss); metallurgical constraint loss (including yield / tensile strength ratio constraint loss, strength-ductility balance constraint loss, monotonicity constraint loss, property range constraint loss and gradient magnitude constraint loss).

[0126] (1) Data fitting loss An attribute-adaptive loss function strategy is adopted. The mean squared error loss is used for microstructure, and the SmoothL1 loss (also known as Huber loss) is used for mechanical properties. It combines the advantages of MSE and MAE, and performs as MSE when the prediction error is less than 1, and degenerates into MAE when the error is greater than 1.

[0127] Data fitting loss Represented as:

[0128] ;

[0129] in, Indicates loss at the microscopic organizational level; This represents the loss of the m-th type of mechanical property.

[0130] Microorganism loss Represented as:

[0131] ;

[0132] in, Batch size; Number of thickness locations; For the first The sample at the th Predicted microstructure values ​​for each location; This corresponds to the actual value.

[0133] mechanical property loss Represented as:

[0134] ;

[0135] The SmoothL1 function is defined as a piecewise function:

[0136] ;

[0137] This loss function combines the advantages of MSE and MAE: when the prediction error is less than 1, it exhibits the smoothing properties of MSE, and when it is greater than 1, it degenerates into the robustness of MAE, effectively suppressing the influence of outliers.

[0138] (2) Gradient consistency loss includes gradient direction consistency loss and gradient smoothness loss. Gradient direction consistency loss ensures that the predicted performance gradient direction is consistent with the target gradient direction, that is, strength-related properties (yield strength, tensile strength) decrease monotonically from the surface to the core, and plastic properties (elongation, impact toughness) increase monotonically. Gradient smoothness loss ensures a smooth transition of the performance curve by penalizing the second-order gradient.

[0139] Gradient consistency loss Represented as:

[0140] ;

[0141] in, This represents the gradient direction consistency loss; This represents the gradient smoothness loss;

[0142] Gradient direction consistency loss Cosine similarity is used as the metric. For strength-related attributes (yield strength, tensile strength, ...), ... ):

[0143] ;

[0144] The gradient vector is defined as:

[0145] ;

[0146] Negative gradient reward term The intensity of the encouragement decreases from the surface to the core.

[0147] For plastic properties (elongation, impact toughness, ... ):

[0148] ;

[0149] Positive gradient reward items Encourage plasticity to increase from the surface to the core.

[0150] Gradient smoothing loss ensures a smooth transition in the performance curve by penalizing the second-order gradient.

[0151] ;

[0152] The first-order gradient is:

[0153] ;

[0154] The second gradient is:

[0155] .

[0156] (3) Metallurgical constraint loss encodes materials science knowledge into neural network training, including yield / tensile strength ratio constraint (reasonable range 0.6-0.95), strength-ductility balance constraint (based on the Considère criterion), property range constraint loss, and monotonicity constraint, expressed as:

[0157] ;

[0158] in, Indicates the yield / tensile strength ratio constraint loss; This represents the strength-ductility balance constraint loss; This represents the loss due to attribute range constraints. This indicates a gradient magnitude constraint.

[0159] The formula for calculating the yield / tensile strength ratio constraint loss (i.e., the theoretical tensile strength constraint loss) is as follows:

[0160] ;

[0161] in, For yield strength, This refers to tensile strength. A reasonable strength ratio range is 1.2-1.4 (corresponding to a yield strength of 0.71-0.83).

[0162] The strength-ductility balance constraint is based on the Considère criterion; high-strength steel typically corresponds to a lower elongation. The formula for calculating the strength-ductility balance constraint loss is:

[0163] ;

[0164] The expected elongation rate is based on an empirical formula:

[0165] ;

[0166] The 5 percentage point tolerance takes into account the influence of factors such as microstructure type and grain size.

[0167] Monotonicity constraints, including strength monotonicity constraints and plasticity monotonicity constraints, are expressed as:

[0168] ;

[0169] Intensity monotonicity constraint (decreasing from surface to core):

[0170] ;

[0171] A small increase in tolerance of 5 MPa is allowed to accommodate fluctuations in actual measurements.

[0172] Plastic monotonicity constraint (increasing from surface to core):

[0173] ;

[0174] A slight reduction in tolerance of 2% is allowed.

[0175] Property range constraints, based on the typical performance range of low-alloy high-strength steel for marine engineering:

[0176] ;

[0177] Yield strength range constraint (450-600MPa):

[0178] ;

[0179] Tensile strength range constraint (600-900MPa):

[0180] ;

[0181] Impact toughness range constraints (290-500J):

[0182] ;

[0183] Gradient magnitude constraint limits the magnitude of performance variation between adjacent locations, ensuring that the prediction results conform to physical continuity:

[0184] ;

[0185] Where M represents the number of attributes, and in this embodiment, M=5 (microstructure, yield strength, tensile strength, elongation, impact toughness). This represents the maximum allowed gradient, which is set according to the model type.

[0186] The total loss function is obtained by weighted summation of all losses:

[0187] ;

[0188] Where λ1=0.001, λ2=0.0001.

[0189] Training was performed using the Adam optimizer with an initial learning rate of 0.001, employing a cosine annealing learning rate scheduling strategy. Validation set performance was evaluated after each training epoch, and checkpoints were saved when the validation loss decreased, including model weights, optimizer state, and learning rate scheduler state. The training batch size was set to 32; the number of training epochs was 100-200; an early stopping strategy was used, stopping training when the validation set loss did not decrease for 20 consecutive epochs; the dataset was partitioned as follows: training set 70%, validation set 20%, and test set 10%.

[0190] In step S3, a shared feature extraction network is used to extract the input features consisting of the chemical composition parameters and heat treatment process parameters of the steel plate obtained in step S1, so as to obtain advanced shared features.

[0191] In step S4, firstly, a multi-head gradient attention network with an introduced gradient bias matrix is ​​used to extract gradient features from the high-level shared features. Then, a position-specific encoding network is used to encode the position features of different thicknesses of the steel plate to obtain the initial thickness direction performance gradient features.

[0192] In step S5, the gradient consistency network is used to learn the features of all positions of the steel plate to obtain the gradient correction vectors of different thickness positions. Combined with the initial thickness direction performance gradient features, the gradient-corrected thickness direction performance gradient features are obtained.

[0193] S6. The attribute prediction network is used to output the heat treatment performance prediction results based on the gradient-corrected thickness direction performance gradient features.

[0194] like Figure 6 As shown, the training process in this embodiment uses the following configuration:

[0195] (1) Batch size: 32;

[0196] (2) Training cycles (epochs): 100-200;

[0197] (3) Optimizer: Adam, initial learning rate 0.001;

[0198] (4) Learning rate scheduling: Cosine Annealing (LR);

[0199] (5) Data set partitioning: 70% training set, 20% validation set, and 10% test set.

[0200] Checkpoint saving strategy: At the end of each training cycle, the system automatically evaluates the validation set performance. When the validation loss decreases, the complete training state is saved, including:

[0201] (1) Current epoch number;

[0202] (2) Model state dictionary (model_state_dict);

[0203] (3) Optimizer state dictionary (optimizer_state_dict);

[0204] (4) Learning rate scheduler state dictionary (scheduler_state_dict);

[0205] (5) Training loss and validation loss;

[0206] (6) Validation set performance metrics (R², MAE, MSE, etc.).

[0207] After training, the checkpoint with the lowest validation loss is automatically loaded as the final model and persistently saved.

[0208] like Figure 7 and Figure 8As shown, the present invention was verified on 20,000 physically constrained enhanced synthetic samples.

[0209] (1)Overall performance metrics

[0210] The GANN model achieved an R² value of 0.9439, with MSE of 39.545, RMSE of 6.289, and MAE of 3.500. Compared with the baseline methods: the R² of LightGBM was 0.9371, XGBoost was 0.9294, GradientBoosting was 0.9077, RandomForest was 0.9017, and SimpleNN was 0.6543. GANN significantly led in overall prediction accuracy.

[0211] (2)Prediction accuracy for each attribute

[0212] Microstructure: R² = 0.9627; Yield strength: R² = 0.9842, MAPE = 1.156%; Tensile strength: R² = 0.9704, MAPE = 0.955%; Elongation: R² = 0.9806, MAPE = 3.054%; Impact toughness: R² = 0.9704, MAPE = 1.228%.

[0213] (3)Gradient prediction performance

[0214] Gradient direction accuracy: Yield strength 100%, Tensile strength 100%, Elongation 99.69%, Impact toughness 97.44%. Monotonicity retention rate: Yield strength 100%, Tensile strength 100%, Elongation 98.77%, Impact toughness 91.50%.

[0215] (4)Physical constraint satisfaction rate

[0216] YS < TS constraint: 100%; Strength - ductility balance: 99.5%; Gradient monotonicity: 99.0%; Hall - Petch relationship: 100%.

[0217] (5)Ablation experiment

[0218] The ablation experiment showed that after removing the physical constraints, the MSE increased by 316.3%, the R² decreased by 11.39%, and the physical violation rate increased from 0% to 16%, demonstrating the crucial role of physical constraints in model performance.

[0219] (6)Computational efficiency

[0220] The model has a total of 307,985 parameters, 0.000320 GFLOPs, a memory footprint of approximately 1.2 MB, a CPU inference time of approximately 6.1 ms for a single sample, and a CPU inference time of approximately 7.05 ms for a batch of 32 samples, meeting the requirements of real-time industrial applications.

[0221] As can be seen, this embodiment achieves simultaneous performance prediction of thick steel plates at five locations from the surface to the core, with an R² of 0.9439, a gradient direction accuracy of approximately 100%, a physical constraint satisfaction rate of over 94%, and an inference time of approximately 6ms. It can be applied to the intelligent production and quality control of thick steel plates for marine engineering.

[0222] This embodiment also proposes an uncertainty quantification method for the constructed gradient-aware neural network model, which can be a deep ensemble method, a Bayesian neural network method, or an MCDropout method.

[0223] (1) Multiple independently initialized gradient-aware neural network models are trained using a deep ensemble method, and the uncertainty is estimated by the standard deviation of multiple prediction results.

[0224] The use of deep integration methods for uncertainty quantification specifically includes:

[0225] Multiple independently initialized gradient-aware neural network models are trained using a deep ensemble method;

[0226] Each gradient-aware neural network model is used to make predictions, resulting in multiple sets of prediction results.

[0227] The mean of multiple prediction results is calculated as the final prediction value;

[0228] The standard deviation of multiple sets of prediction results is calculated as the prediction uncertainty.

[0229] (2) The uncertainty is quantified by modeling the probability distribution of the weights using the Bayesian neural network method, that is, the prediction variance obtained by sampling the posterior distribution of the weights.

[0230] (3) The MCDropout method is used to maintain Dropout activation during inference and to estimate uncertainty through multiple forward propagation of the prediction variance.

[0231] The correlation coefficient between prediction uncertainty and absolute error (i.e., the Pearson correlation coefficient between prediction uncertainty and absolute error) obtained based on the Bayesian neural network method is ρ=0.262, indicating that the model can identify cases with high prediction difficulty. The expected calibration error (ECE) obtained based on the Bayesian neural network method is 10.7%, which is within an acceptable range.

[0232] Example 2

[0233] This embodiment also provides a heat treatment performance prediction system for thick steel plates based on gradient sensing mechanism and physical constraints, used to implement the method in Embodiment 1. The system includes:

[0234] The data generation module is used to generate the chemical composition parameters and heat treatment process parameters of the steel plate;

[0235] The gradient-aware neural network model comprises a shared feature extraction network, a multi-head gradient attention network, a location-specific encoding network, a gradient consistency network, and an attribute prediction network. The shared feature extraction network extracts high-level shared features from input features composed of chemical composition parameters and heat treatment process parameters. The multi-head gradient attention network and location-specific encoding network explicitly model the thickness-direction performance gradient features using a multi-head gradient attention mechanism and location encoding, and enhance information exchange between adjacent locations through a gradient bias matrix. The gradient consistency network learns the interrelationships between locations and outputs gradient-corrected thickness-direction performance gradient features. Finally, the attribute prediction network outputs heat treatment performance prediction results based on the gradient-corrected thickness-direction performance gradient features.

[0236] The data generation module includes a chemical composition parameter generation unit, a heat treatment process parameter generation unit, and a derived feature generation unit. The chemical composition parameter generation unit generates the chemical composition parameters of the steel plate to be predicted. The heat treatment process parameter generation unit generates the heat treatment process parameters of the steel plate to be predicted. The derived feature generation unit calculates derived features based on the chemical composition parameters and heat treatment process parameters of the steel plate.

[0237] In an optional embodiment, the above-mentioned thick steel plate heat treatment performance prediction system based on gradient sensing mechanism and physical constraints further includes:

[0238] Bayesian models are used for uncertainty quantification.

[0239] In an optional embodiment, the above-mentioned thick steel plate heat treatment performance prediction system based on gradient sensing mechanism and physical constraints further includes:

[0240] The deep ensemble model is obtained by integrating multiple gradient-aware neural network models configured in parallel.

[0241] In an optional embodiment, the above-mentioned thick steel plate heat treatment performance prediction system based on gradient sensing mechanism and physical constraints further includes:

[0242] The prediction report generation module is used to generate a heat treatment performance prediction result report based on the gradient-aware neural network model or the deep ensemble model, and can also generate the uncertainty of the prediction result through the Bayesian model.

[0243] Example 3

[0244] This embodiment is a further improvement based on Embodiment 2, such as... Figure 9 As shown, the thick steel plate heat treatment performance prediction system provided in this embodiment also includes a model management module for saving, loading, version control, and selecting the best model. The specific functions of this module include:

[0245] (a) Model weight storage: The model state dictionary is stored in PyTorch's .pth format, which includes model parameters, model type, timestamps and performance metrics;

[0246] (b) Preprocessor persistence: Use the joblib library to save the data preprocessor (containing normalized parameters) as a .pkl format file;

[0247] (c) Metadata management: Store model type, timestamp, model path, preprocessor path, performance metrics, configuration parameters, and feature names in JSON format;

[0248] (d) Automatic selection of the best model: The link file of the best model is automatically updated by comparing the R² scores on the validation set;

[0249] (e) Version backtracking: All available models are obtained by scanning the metadata catalog, and historical version queries are supported by sorting by timestamp.

[0250] This embodiment uses a global model registry to uniformly manage the three model types:

[0251] (1) gradient_aware: The gradient-aware neural network model (referred to as the gradient-aware model) is the core prediction model;

[0252] (2) Bayesian: Bayesian neural network, used for uncertainty quantification;

[0253] (3) ensemble: Deep ensemble model, which improves prediction stability by integrating multiple gradient-aware neural network models.

[0254] Example 4

[0255] This embodiment is a further improvement based on Embodiment 2 or Embodiment 3. The thick steel plate heat treatment performance prediction system in this embodiment also includes a data verification unit and a physical constraint verification unit.

[0256] The data verification unit is used to verify whether the input chemical composition parameters and heat treatment process parameters are within the preset valid range (as shown in Table 2) before prediction. Only input parameters within the preset valid range can be used for subsequent operations.

[0257] Physical constraint verification is used to verify whether the prediction results meet the physical constraints. Only prediction results that meet the physical constraints can be output. If the physical constraints are not met, the prediction results will still be output and a warning message will be marked.

[0258] Physical constraints include:

[0259] (a) Yield strength is less than tensile strength;

[0260] (b) The intensity gradient decreases from the surface to the core; surface > 1 / 8 > 1 / 4 > 3 / 8 > core;

[0261] The plastic gradient increases from the surface to the core: surface < 1 / 8 < 1 / 4 < 3 / 8 < core;

[0262] (c) Consistency between microstructure and mechanical properties: Verify that martensitic microstructure (type 3) corresponds to high strength (yield strength > 550 MPa) and low plasticity (elongation < 18%), bainitic + ferrite microstructure (type 2) corresponds to medium strength and plasticity, and ferrite + pearlite microstructure (type 1) corresponds to lower strength (yield strength < 520 MPa) and higher plasticity (elongation > 20%), etc., which are metallurgical laws. The consistency judgment method is as follows: based on the predicted microstructure type, check whether the predicted mechanical property value at the corresponding thickness position falls within the typical performance range of the microstructure type. If the predicted mechanical property value exceeds the typical range of the corresponding microstructure type, it is determined that the microstructure and mechanical properties are inconsistent.

[0263] Table 2 Valid range of input parameters

[0264]

[0265] Example 5

[0266] This embodiment is a further improvement on embodiment 4. This embodiment provides a Web prediction service interface that can load the above-mentioned thick steel plate heat treatment performance prediction system, that is, a Flask-based Web interface that supports single-sample prediction and batch prediction.

[0267] Users can input chemical composition and heat treatment process parameters, and the web-based prediction service interface will output the microstructure type and mechanical property prediction values ​​of steel plates at different thickness locations.

[0268] The operation process of the above Web prediction service interface is as follows:

[0269] (1) Verify whether the input chemical composition parameters are within the preset valid range; if yes, proceed to the next step; otherwise, return the error message "Composition out of range";

[0270] (2) Verify whether the input heat treatment process parameters are within the preset valid range; if yes, proceed to the next step; otherwise, return the error message "process parameters out of range";

[0271] (3) Execute model (gradient-aware neural network model or deep ensemble model) prediction;

[0272] (4) Verify whether the prediction results meet the physical constraints; if the physical constraints are met, output the prediction results (microstructure type and mechanical property value at 5 thickness locations); if the physical constraints are not met, mark the warning message, still output the prediction results but with a prompt.

[0273] The output prediction results can also be formatted according to design requirements, and the final output prediction results can be formatted.

[0274] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0275] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0276] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0277] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0278] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for predicting the heat treatment performance of thick steel plates based on a gradient-aware mechanism and physical constraints, characterized in that, The method comprises the following steps: obtaining chemical composition parameters and heat treatment process parameters of the steel plate; constructing a gradient-aware neural network model; the gradient-aware neural network model comprises a shared feature extraction network, a multi-head gradient attention network, a position-specific encoding network, a gradient consistency network and an attribute prediction network; extracting high-level shared features from input features composed of the chemical composition parameters and the heat treatment process parameters by using the shared feature extraction network; using the multi-head gradient attention network and the position-specific encoding network to explicitly model the thickness-direction performance gradient features by using a multi-head gradient attention mechanism and position encoding, and to strengthen the exchange of adjacent position information through a gradient bias matrix; learning the mutual relationship between positions by using the gradient consistency network, and outputting the thickness-direction performance gradient features after gradient correction; outputting the heat treatment performance prediction results according to the thickness-direction performance gradient features after gradient correction by using the attribute prediction network.

2. The method for predicting the heat treatment performance of thick steel plates based on a gradient-aware mechanism and physical constraints according to claim 1, characterized in that, The chemical composition parameters of the steel plate include alloy element content parameters such as C, Mn, Si, Cr, Ni and Mo; The heat treatment process parameters include plate thickness, austenitizing temperature, holding time, surface cooling rate, cooling rate distribution along the thickness direction of the steel plate, tempering temperature and tempering time.

3. The method for predicting the heat treatment performance of thick steel plates based on a gradient-aware mechanism and physical constraints according to claim 1 or 2, characterized in that, In addition to obtaining the chemical composition parameters and the heat treatment process parameters of the steel plate, the method further comprises calculating derived features according to the chemical composition parameters and the heat treatment process parameters of the steel plate; the derived features include carbon equivalent, hardenability index, cooling intensity, tempering parameter and carbon-manganese interaction term. 4.The method of predicting heat treatment performance of thick steel plate based on gradient-aware mechanism and physical constraints according to claim 1, wherein, The multi-head gradient attention network adds a gradient bias matrix to the calculation of scaled dot-product attention in each attention head, so that the attention weight gives priority to the information of adjacent positions; the multi-head gradient attention network outputs the final attention features through splicing and linear transformation. 5.The method for predicting the heat treatment performance of thick steel plates based on a gradient-aware mechanism and physical constraints according to claim 1, characterized in that, The position-specific encoding network discretizes the thickness direction into multiple standard positions for position encoding; and a corresponding number of position-specific encoders are used to process the features of the corresponding thickness positions by taking the output features of the multi-head gradient attention network and the position encoding of the corresponding positions as inputs. 6.The method for predicting the heat treatment performance of thick steel plates based on a gradient-aware mechanism and physical constraints according to claim 1, characterized in that, The attribute prediction network comprises a microstructure type classification head, a yield strength regression head, a tensile strength regression head, an elongation regression head and an impact toughness regression head; Outputting the heat treatment performance prediction results according to the thickness-direction performance gradient features after gradient correction by using the attribute prediction network comprises: obtaining the microstructure type according to the microstructure prediction results; performing inverse standardization on the mechanical property prediction results to obtain the true values; the mechanical property prediction results include yield strength, tensile strength, elongation and impact toughness prediction values; outputting the microstructure type, yield strength, tensile strength, elongation and impact toughness prediction values of the corresponding thickness positions.

7. The method for predicting the heat treatment performance of thick steel plates based on a gradient-aware mechanism and physical constraints according to claim 1, characterized in that, The gradient-aware neural network model adopts a weighted sum of a data fitting loss, a gradient consistency loss and a metallurgical constraint loss to construct a total loss function during training; the data fitting loss includes a microstructure mean square error loss and a mechanical property SmoothL1 loss; the gradient consistency loss includes a gradient direction consistency loss and a gradient smoothness loss; and the metallurgical constraint loss includes a yield / tensile strength ratio constraint loss, a strength-ductility balance constraint loss, a monotonicity constraint loss, an attribute range constraint loss and a gradient amplitude constraint loss.

8. The method for predicting the heat treatment performance of thick steel plates based on a gradient-aware mechanism and physical constraints according to claim 7, characterized in that, The gradient direction consistency loss adopts cosine similarity measurement and includes a strength attribute gradient consistency loss and a plastic attribute gradient consistency loss; the gradient smoothness loss is calculated by penalizing a second-order gradient; the strength-ductility balance constraint is calculated based on the Considère criterion; the monotonicity constraint loss includes a strength monotonicity constraint and a plastic monotonicity constraint and is calculated based on a linear modification function ReLu; the attribute range constraint loss includes a yield strength range constraint loss, a tensile strength range constraint loss and an impact toughness range constraint loss, and each of the yield strength range constraint loss, the tensile strength range constraint loss and the impact toughness range constraint loss is calculated based on the linear modification function ReLu; and the gradient amplitude constraint loss is calculated based on the linear modification function ReLu. 9.The method for predicting the heat treatment performance of thick steel plates based on a gradient-aware mechanism and physical constraints according to claim 1, characterized in that, Further comprising uncertainty quantification, specifically including: training multiple independently initialized gradient-aware neural network models using a deep ensemble method, and estimating uncertainty through variance of prediction results; or using a Bayesian neural network method to model probability distribution of weights to realize uncertainty quantification; or using an MCDropout method to maintain Dropout activation during inference, and estimating uncertainty through prediction variance of multiple forward propagations.

10. The method for predicting the heat treatment performance of thick steel plates based on a gradient-aware mechanism and physical constraints according to claim 9, characterized in that, Uncertainty quantification using a deep ensemble method specifically includes: training multiple independently initialized gradient-aware neural network models using a deep ensemble method; performing prediction on each of the gradient-aware neural network models to obtain multiple sets of prediction results; calculating a mean of the multiple sets of prediction results as a final prediction value; calculating a standard deviation of the multiple sets of prediction results as prediction uncertainty.

11. A system for predicting the heat treatment performance of thick steel plates based on a gradient-aware mechanism and physical constraints, characterized in that, Further comprising: a data generation module configured to generate chemical composition parameters and heat treatment process parameters of a steel plate; a gradient-aware neural network model including a shared feature extraction network, a multi-head gradient attention network, a position-specific encoding network, a gradient consistency network and an attribute prediction network; the shared feature extraction network is used to extract high-level shared features from input features composed of the chemical composition parameters and the heat treatment process parameters; the multi-head gradient attention network and the position-specific encoding network are used to explicitly model thickness-direction performance gradient features using a multi-head gradient attention mechanism and position encoding, and to strengthen exchange of information between adjacent positions through a gradient bias matrix; the gradient consistency network is used to learn mutual relationships between positions and output gradient-corrected thickness-direction performance gradient features; the attribute prediction network is used to output heat treatment performance prediction results according to the gradient-corrected thickness-direction performance gradient features.

12. The thick steel plate heat treatment performance prediction system based on a gradient-aware mechanism and physical constraints according to claim 11, characterized in that, Further comprising: a Bayesian model configured to quantify uncertainty. The deep integration model is obtained by integrating a plurality of gradient perception neural network models arranged in parallel.

13. The thick steel plate heat treatment performance prediction system based on a gradient-aware mechanism and physical constraints according to claim 11 or 12, characterized in that, It also includes a data verification unit and a physical constraint verification unit: The data verification unit is used to verify whether the input chemical composition parameters and heat treatment process parameters are within the preset effective range before prediction. Physical constraint verification is used to verify whether the prediction result meets the physical constraint.

14. A Web prediction service interface for loading the thick steel plate heat treatment performance prediction system based on the gradient perception mechanism and physical constraint of claim 13.

Images