An amorphous alloy material performance prediction method and system based on multi-task learning

Abstract

This invention discloses a method and system for predicting the performance of amorphous alloy materials based on multi-task learning, relating to the field of materials informatics. The method includes: constructing a task set according to the prediction target; repeatedly executing the multi-task learning steps according to the task set until the multi-task learning model converges; the multi-task learning steps include: using a feature extractor to extract features from the sample data to obtain specific feature information for each task; using an expert hybrid network to extract features from the sample data to obtain M shared feature information; fusing the specific feature information and the shared feature information; and performing multi-task prediction based on the fused feature information to obtain the prediction result for each task; and updating the parameters of the multi-task learning model based on the prediction results. This invention fuses task-specific feature information and shared feature information, enhancing the model's generalization ability and interpretability, and by employing a multi-task learning approach, improving the prediction accuracy of performance parameters that are difficult to predict.

Description

Technical Field

[0001] This invention relates to the field of materials informatics, and in particular to a method and system for predicting the properties of amorphous alloy materials based on multi-task learning. Background Technology

[0002] With the extensive research on bulk amorphous alloys, the outstanding properties of this novel high-performance metallic material, such as high strength, high hardness, corrosion resistance, low melting point, and excellent elastic modulus, have attracted widespread attention in the scientific community. However, the design and optimization of bulk amorphous alloys remain challenging due to their high-dimensionality and complex compositional space. Using a trial-and-error approach involves cumbersome experimental procedures, long development cycles, and wasted resources. Data-driven machine learning algorithms offer a novel approach to the design of bulk amorphous alloys. However, the different properties of bulk amorphous alloys vary in prediction difficulty due to their different formation mechanisms. For example, with the same amount of data, predicting the glass transition temperature of bulk amorphous alloys is much more difficult than predicting their glass-forming ability. For performance parameters that are more difficult to predict, accurate prediction cannot be achieved with a limited number of samples. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for predicting the properties of amorphous alloy materials based on multi-task learning, which can improve the prediction accuracy of performance parameters that are difficult to predict.

[0004] To achieve the above objectives, the present invention provides the following solution:

[0005] A method for predicting the properties of amorphous alloy materials based on multi-task learning, the method comprising:

[0006] A task set is constructed based on the prediction target; the task set includes a main task and at least one auxiliary task, the main task being the task of predicting the prediction target, and the auxiliary task being the task of predicting performance parameters associated with the prediction target;

[0007] The multi-task learning steps are repeated according to the task set until the multi-task learning model converges; wherein, the multi-task learning model includes multiple feature extractors, expert hybrid networks, attention mechanisms and regression models; the multiple feature extractors correspond one-to-one with multiple tasks in the task set;

[0008] The steps of multi-task learning include:

[0009] Each feature extractor corresponding to each task in the task set is used to extract features from the sample data to obtain the unique feature information of each task in the task set; the sample data is the composition information of amorphous alloy materials, and the labels of the sample data are the performance parameters of amorphous alloy materials that need to be predicted in each task in the task set;

[0010] Each expert network in the expert hybrid network is used to extract features from the sample data to obtain M shared feature information; M is the total number of expert networks included in the expert hybrid network;

[0011] By using an attention mechanism, the unique feature information of each task in the task set is fused with all the shared feature information to obtain the fused feature information of each task in the task set.

[0012] The fused feature information of each task is input into the regression model to obtain the prediction results for each task;

[0013] The parameters of the multi-task learning model are updated based on the prediction results for each task.

[0014] Optionally, the step of using an attention mechanism to fuse the unique feature information of each task in the task set with all shared feature information to obtain fused feature information for each task in the task set specifically includes:

[0015] Based on the unique feature information of each task in the task set, a query vector is generated for each task in the task set.

[0016] Based on M shared feature information, generate M key vectors and M value vectors; where K m =V m m = 1, 2, ..., M, K m and V m These are the m-th key vector and the m-th value vector, respectively.

[0017] Let n = 1, 2, ..., N, where N is the number of tasks in the task set, and perform the following steps:

[0018] The query vector of the nth task is multiplied by the M key vectors to obtain the attention weights of the nth task in each shared feature information, and the attention vector of the nth task is formed.

[0019] The attention value of the nth task is obtained by multiplying the M value vectors by the attention vector of the nth task.

[0020] The attention value of the nth task is concatenated with the query vector of the nth task to obtain the fused feature information of the nth task.

[0021] Optionally, the attention value of the nth task is concatenated with the query vector of the nth task to obtain the fused feature information of the nth task. This process also includes:

[0022] The query vector of the nth task is constrained using the following formula;

[0023] Q' n =λ n Q n ;

[0024] Among them, Q' n Let Q be the Query vector after constraint processing for the nth task. n Let λ be the query vector for the nth task. n Let λ be the hyperparameter of the nth task, 0 ≤ λ n ≤1, and when the nth task is the main task, λ n =1.

[0025] Optionally, a dynamically weighted loss function can be used in the multi-task learning step.

[0026] Optionally, the dynamically weighted loss function is:

[0027]

[0028] Where L(t) is the dynamic weighted loss function during the t-th iteration of training, w n (t) represents the weight of the nth task during the t-th iteration of training, L n (t) represents the loss of the nth task during the t-th iteration of training, where N is the number of tasks in the task set.

[0029] Optionally, the weights are updated using the following steps:

[0030] A parameter matrix is ​​constructed based on the parameters of the last layer of the feature extractor and the expert hybrid network corresponding to each task.

[0031] Based on the year-on-year change rate of the parameters of the last layer of the feature extractor and the expert hybrid network corresponding to each task, a year-on-year change rate matrix is ​​constructed.

[0032] Based on the parameter matrix and the year-on-year change rate matrix, calculate the parameter vector consisting of the gradient change rate of the parameters of the feature extractor corresponding to each task;

[0033] Based on the parameter vector, the gradient L2 norm of the loss of each task in the t-th iteration training process is calculated using the following formula;

[0034]

[0035] in, Let L2 norm be the gradient of the loss of the nth task during the t-th iteration of training. This represents the weighted loss of the nth task during the t-th iteration of training, expressed in parameter W. n Partial derivative on, Wn The gradient rate of change of the parameters of the feature extractor corresponding to the nth task;

[0036] Calculate the average L2 norm of the loss of all tasks in the t-th iteration of training based on the gradient L2 norm of the loss of each task in the t-th iteration of training.

[0037] Based on the gradient L2 norm of the loss of each task in the t-th iteration training process and the average of the L2 norms of the loss of all tasks, the regularized gradients of all tasks in the t-th iteration training process are calculated using the following formula.

[0038]

[0039] in, This represents the gradient after task regularization during the t-th iteration of training. p represents the average L2 norm of the gradients of the losses of all tasks during the t-th iteration of training. n (t) represents the intermediate parameter. α is a hyperparameter used to control the magnitude of the change;

[0040] Based on the gradients of all tasks after regularization during the t-th iteration of training, the weights are updated using the following formula;

[0041]

[0042] Where lr is the learning rate during iterative training. This represents the gradient of all tasks after regularization during the t-th iteration of training, in w. n The partial derivative on (t), w n (t+1) represents the weight of the nth task during the (t+1)th iteration of training.

[0043] Optionally, feature extraction is performed on the sample data using the feature extractor corresponding to each task in the task set to obtain the specific feature information of each task in the task set. This also includes:

[0044] Based on the composition expression of amorphous alloy materials, the composition information of amorphous alloy materials is obtained;

[0045] The component composition information is reduced in dimensionality to obtain the sample data.

[0046] A computer system includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0047] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the above-described method.

[0048] A computer program product includes a computer program that, when executed by a processor, implements the steps of the above method.

[0049] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0050] A method and system for predicting the performance of amorphous alloy materials based on multi-task learning is disclosed. The method includes: constructing a task set according to the prediction target; repeatedly executing the multi-task learning steps according to the task set until the multi-task learning model converges; the multi-task learning steps include: extracting features from the sample data using the feature extractor corresponding to each task in the task set to obtain the specific feature information of each task in the task set; extracting features from the sample data using each expert network in the expert hybrid network to obtain M shared feature information; M being the total number of expert networks included in the expert hybrid network; fusing the specific feature information of each task in the task set with all the shared feature information using an attention mechanism to obtain the fused feature information of each task in the task set; inputting the fused feature information of each task into a regression model to obtain the prediction result of each task; and updating the parameters of the multi-task learning model according to the prediction result of each task. This invention integrates task-specific feature information and shared feature information, enhances the generalization ability and interpretability of the model, and improves the prediction accuracy of performance parameters that are difficult to predict by adopting a multi-task learning approach. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 A flowchart illustrating a method for predicting the properties of amorphous alloy materials based on multi-task learning, provided in an embodiment of the present invention;

[0053] Figure 2 A flowchart for constructing a task set provided in an embodiment of the present invention;

[0054] Figure 3 A flowchart of a multi-task learning algorithm provided in an embodiment of the present invention;

[0055] Figure 4 An internal structure diagram of a computer system provided for an embodiment of the present invention. Detailed Implementation

[0056] 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.

[0057] The purpose of this invention is to provide a method and system for predicting the properties of amorphous alloy materials based on multi-task learning, which can improve the prediction accuracy of performance parameters that are difficult to predict.

[0058] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0059] Research has shown that different properties of bulk amorphous alloys are related based on domain knowledge, necessitating the integration of domain knowledge into the feature information required for model building. Based on this, this invention proposes a novel multi-task learning algorithm to address this problem. This algorithm cleverly integrates task-specific and shared feature information, enhancing the model's generalization ability and interpretability, and improving the prediction accuracy of the target task.

[0060] The purpose of this invention is to address the challenge of single-objective prediction in the field of bulk amorphous alloys, where small-sample learning convergence issues make accurate prediction of glass-forming ability based on alloy composition difficult. This invention proposes a deep multi-task learning model, DPAMTL, which integrates primary and auxiliary tasks. It mainly comprises three parts: an expert hybrid network, an attention mechanism, and a dynamically weighted loss. The expert hybrid network integrates information from multiple tasks, compensating for the inability to extract task-specific information in hard parameter sharing. The attention mechanism assigns appropriate expert network weights to each task, thereby better extracting shared task information. The dynamically weighted loss balances the magnitude differences between losses for different tasks, preventing high-magnitude losses from dominating the training of the entire network. This provides a novel and effective method for predicting glass-forming ability.

[0061] In one embodiment, a method for predicting the properties of amorphous alloy materials based on multi-task learning is provided, such as... Figure 1 As shown, the method includes:

[0062] A task set is constructed based on the prediction target; the task set includes a main task and at least one auxiliary task, the main task being the task of predicting the prediction target, and the auxiliary task being the task of predicting performance parameters associated with the prediction target.

[0063] The process of constructing the task set is as follows: Figure 2 As shown, it specifically includes:

[0064] Step 1.1: Based on the target requirements, the glass forming ability (GFA) prediction based on the alloy composition of bulk amorphous alloys is set as the main task.

[0065] Step 1.2: Based on domain knowledge, select different combinations of glass transition temperature (Tg), glass initiation crystallization temperature (Tx), and glass liquefaction temperature (Tl) as auxiliary tasks.

[0066] Step 1.3: Based on Steps 1.1 and 1.2, the following four task sets were designed. Dmax is the main task, and the rest are auxiliary tasks.

[0067] Task Set 1: {[Dmax],(Tg,Tx,Tl)};

[0068] Task Set2: {[Dmax],(Tx-Tg,Tx / Tl)};

[0069] Task Set 3: {[Dmax],(Tg,Tx-Tg,Tx / Tl)};

[0070] Task Set4: {[Dmax],(Tg,Tx-Tg)};

[0071] Step 1.4: Select a task set in sequence for multi-task learning, and finally select the model with the best performance.

[0072] The multi-task learning steps are repeated based on the task set until the regression model converges; wherein the regression model is used to make predictions for each task in the task set.

[0073] The steps of multi-task learning include: extracting features from the sample data using the feature extractor corresponding to each task in the task set to obtain the specific feature information of each task in the task set; extracting features from the sample data using each expert network in the expert hybrid network to obtain M shared feature information; M is the total number of expert networks included in the expert hybrid network; fusing the specific feature information of each task in the task set with all the shared feature information using an attention mechanism to obtain the fused feature information of each task in the task set; and training the regression model using the fused feature information of each task in the task set.

[0074] For example, such as Figure 3 As shown, the specific steps of multi-task learning include:

[0075] Step 2.1: Preprocess the bulk amorphous alloy material data and then input it into the network.

[0076] Step 2.1 in this embodiment of the invention specifically includes: generating 94-dimensional features based on the alloy's composition expression and domain knowledge. To reduce training difficulty and improve model performance, methods such as R2 filtering and distance coefficients are used to remove features with low information content, which are then used as sample data.

[0077] Step 2.2: Use N feature extractors to extract task-specific feature information. Here, a fully connected neural network is used as the feature extractor.

[0078] Step 2.2 in this embodiment of the invention specifically includes: extracting task-specific feature information using multiple feature extractors, and constraining the feature information of different tasks using parameter λ. The feature information output in this step can be used as a query vector for subsequent use.

[0079] During iterative training, each feature extractor updates its parameters based on a different loss, as shown in Equation (1).

[0080]

[0081] Where t represents the number of iterations in the training process, The mean square error is used for calculation, that is... Where m is the number of samples; y j The actual value; These are predicted values.

[0082] Step 2.3: Extract shared feature information from multiple tasks using a hybrid expert network. There are M expert networks, where M>=N, meaning the number of expert networks is no less than the number of tasks. The shared feature information output in this step can be used as Key and Value vectors for subsequent use. The hybrid expert network extracts the shared feature information and maps it to a space with the same dimension as the task-specific feature information. Multiple expert networks rely on the same loss during updates, as shown in Equation (2).

[0083]

[0084] Where t represents the number of training iterations. Where L all Same as L(t), that is

[0085] In step 2.3, assuming the number of tasks is N, M (≥N) expert networks are constructed to extract shared feature information from multiple tasks. The feature information output in this step can be used as a key vector for subsequent applications.

[0086] Step 2.4: Step 2.2 yielded N Query vectors, and Step 2.3 yielded M Key vectors and M Value vectors, where Key = Value.

[0087] Step 2.5: In order to obtain the fused feature vector for each task, the dot product of a Query vector and each Key vector is performed sequentially to obtain the attention weight of each task in the shared feature information.

[0088] Step 2.6: Multiply the Value vector by the attention weights to obtain the attention value.

[0089] Step 2.7: Then, concatenate the attention value with the Query vector to obtain the fused feature information for a specific task. It's important to note that the Query vector needs to be constrained before concatenation by multiplying it by the hyperparameter λ (0 ≤ λ ≤ 1). For the Query vector of the main task Dmax, λ = 1.

[0090] In steps 2.5-2.7, based on the expert hybrid network and the query vector and key vector obtained in steps 2.3 and 2.4, the weights of different expert networks are calculated using the attention mechanism, and the weighted shared information is concatenated with the proprietary information.

[0091] The feature information fusion in this embodiment of the invention aims to integrate proprietary feature information and shared feature information, ensuring that the information of the main task dominates the fusion, specifically including:

[0092] Step A1: Construct the Query vector. Select the output of the feature extractor for task τ as the Query vector.

[0093] Step A2: Construct the Key vector. The output of the expert network is used as the Key vector. The attention weight vector α is obtained by multiplying the Query vector by each Key vector.

[0094] Step A3: Multiply the output of the expert network by the attention weights to obtain the weighted shared feature information.

[0095] Step A4: Concatenate the proprietary and shared feature information of task τ.

[0096] Step 2.8: Repeat steps 2.5 to 2.7 until the fusion feature information of each task is obtained.

[0097] Step 2.9: Input the fused feature information of each task into the prediction layer (i.e., the regression model) to jointly predict multiple targets. Continue until the model converges.

[0098] In multi-task learning, there are often differences between tasks, such as varying data volumes and difficulty levels. If left unaddressed, these differences may lead to some tasks being overlooked or treated unfairly. This invention proposes a dynamic weighting method for multi-task loss, ensuring that the norm of the gradient for each task remains the same. This guarantees that each task has an equal impact on model updates, thereby improving the model's prediction accuracy on the primary task.

[0099] In this embodiment of the invention, dynamic weighting of multi-task losses is used to address issues such as inconsistent training speeds for different tasks, and the possibility that tasks with larger losses may dominate the training of the entire multi-task network during each parameter update. Specific steps include:

[0100] Step B1: Formula for dynamic loss function: The value of n ranges from [1, N], where N is the number of tasks.

[0101] Step B2: Select the parameters of the last layer of the feature extractor and the expert hybrid network [W] n,1 W n,1 ,…,W n,k ]. Calculation parameter W n,i The year-on-year change rate r n,i And for [r] n,1 ,r n,2 ,…,r n,k Normalization is performed. Then, the parameters used to calculate the gradient rate of change are determined. Here, the weight vector and parameter vector are multiplied and added together to obtain the parameters W, where k represents the number of feature extractors and expert hybrid networks. [W] n,1 W n,1 ,…,W n,k The vector represents the combination of the feature extractor for the nth task and the parameters of the last layer of the expert hybrid network. In this embodiment, they are arranged side by side into a matrix. The softmax function returns a 1×k row vector (where each element is a value for each r). i The probabilities after softmax transformation are then multiplied by a k×N matrix (based on the column vector [W]). n,1 W n,1 ,…,W n,k ] T Composition), the result is a 1×n row vector W = [W1W2,…,W N ].

[0102] Step B3: Calculate the loss L for each single task. n (t) is the L2 norm of the gradient at the t-th training round. It is calculated as follows: Where w n (t) is a numerical value representing the loss weights for task n in the current t-th training round. The partial derivative of the weighted loss with respect to the parameters W is then used to obtain the L2 paradigm of the gradient. This can typically be accomplished using automatic differentiation techniques (such as in frameworks like TensorFlow or PyTorch).

[0103] Step B4: Calculate the average gradient L2 norm for all tasks during the t-th training round.

[0104] Step B5: Calculate the gradients after task regularization for all tasks. The rules here are the calculation methods shown in the formula. Among them... α is used as a hyperparameter to control the magnitude of the change.

[0105] Step B6: Calculate the gradient Then update the loss weight of task n. Where lr is the learning rate during training. This indicates that the regularized gradient is applied to w during the t-th training round. n Find the partial derivative. This can usually be done using automatic differentiation techniques (such as in frameworks like TensorFlow or PyTorch).

[0106] In one embodiment, a computer system is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the steps of the methods described in the foregoing embodiments. The computer system may be a database, and its internal structure diagram may be as follows: Figure 4As shown, the system includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores pending transactions. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements the method described in the preceding embodiments.

[0107] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0108] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0109] It should be noted that the object information (including but not limited to object device information, object personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the object or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0110] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0111] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0112] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting the properties of amorphous alloy materials based on multi-task learning, characterized in that, The method includes: A task set is constructed based on the prediction target; the task set includes a main task and at least one auxiliary task, the main task being the task of predicting the prediction target, and the auxiliary task being the task of predicting performance parameters associated with the prediction target; The multi-task learning steps are repeated according to the task set until the multi-task learning model converges; wherein, the multi-task learning model includes multiple feature extractors, expert hybrid networks, attention mechanisms and regression models; the multiple feature extractors correspond one-to-one with multiple tasks in the task set; The steps of multi-task learning include: Each feature extractor corresponding to each task in the task set is used to extract features from the sample data to obtain the unique feature information of each task in the task set; the sample data is the composition information of amorphous alloy materials, and the labels of the sample data are the performance parameters of amorphous alloy materials that need to be predicted in each task in the task set; Each expert network in the expert hybrid network is used to extract features from the sample data to obtain M shared feature information; M is the total number of expert networks included in the expert hybrid network; By using an attention mechanism, the unique feature information of each task in the task set is fused with all the shared feature information to obtain the fused feature information of each task in the task set. The fused feature information of each task is input into the regression model to obtain the prediction results for each task; The parameters of the multi-task learning model are updated based on the prediction results for each task.

2. The method for predicting the properties of amorphous alloy materials based on multi-task learning according to claim 1, characterized in that, The method of using an attention mechanism to fuse the unique feature information of each task in the task set with all shared feature information to obtain the fused feature information of each task in the task set specifically includes: Based on the unique feature information of each task in the task set, a query vector is generated for each task in the task set. Based on the M shared feature information, M Key vectors and M Value vectors are generated; wherein, K m = V m , m = 1, 2, …, M, K m and V m are the mth Key vector and the mth Value vector respectively; Let n = 1, 2, ..., N, where N is the number of tasks in the task set, and perform the following steps: The query vector of the nth task is multiplied by the M key vectors to obtain the attention weights of the nth task in each shared feature information, and the attention vector of the nth task is formed. The attention value of the nth task is obtained by multiplying the M value vectors by the attention vector of the nth task. The attention value of the nth task is concatenated with the query vector of the nth task to obtain the fused feature information of the nth task.

3. The method for predicting the properties of amorphous alloy materials based on multi-task learning according to claim 2, characterized in that, The attention value of the nth task is concatenated with the query vector of the nth task to obtain the fused feature information of the nth task. This process also includes: The query vector of the nth task is constrained using the following formula; Q' n = λ n Q n ; Q n is the constraint processed Query vector of the nth task, Q n is the Query vector of the nth task, λ n is the hyper-parameter of the nth task, 0 ≤ λ n ≤ 1, and λ n = 1 when the nth task is the main task.

4. The method for predicting the properties of amorphous alloy materials based on multi-task learning according to claim 1, characterized in that, A dynamic weighted loss function is used in the multi-task learning process.

5. The method for predicting the properties of amorphous alloy materials based on multi-task learning according to claim 4, characterized in that, The dynamically weighted loss function is: Where L(t) is the dynamic weighted loss function during the t-th iteration of training, w n (t) represents the weight of the nth task during the t-th iteration of training, L n (t) represents the loss of the nth task during the t-th iteration of training, where N is the number of tasks in the task set.

6. The method for predicting the properties of amorphous alloy materials based on multi-task learning according to claim 5, characterized in that, The weights are updated using the following steps: A parameter matrix is ​​constructed based on the parameters of the last layer of the feature extractor and the expert hybrid network corresponding to each task. Based on the year-on-year change rate of the parameters of the last layer of the feature extractor and the expert hybrid network corresponding to each task, a year-on-year change rate matrix is ​​constructed. Based on the parameter matrix and the year-on-year change rate matrix, calculate the parameter vector consisting of the gradient change rate of the parameters of the feature extractor corresponding to each task; Based on the parameter vector, the gradient L2 norm of the loss of each task in the t-th iteration training process is calculated using the following formula; in, Let L2 norm be the gradient of the loss of the nth task during the t-th iteration of training. This represents the weighted loss of the nth task during the t-th iteration of training, expressed in parameter W. n Partial derivative on, W n The gradient rate of change of the parameters of the feature extractor corresponding to the nth task; Calculate the average L2 norm of the loss of all tasks in the t-th iteration of training based on the gradient L2 norm of the loss of each task in the t-th iteration of training. Based on the gradient L2 norm of the loss of each task in the t-th iteration training process and the average of the L2 norms of the loss of all tasks, the regularized gradients of all tasks in the t-th iteration training process are calculated using the following formula. in, This represents the gradient after task regularization during the t-th iteration of training. p represents the average L2 norm of the gradients of the losses of all tasks during the t-th iteration of training. n (t) represents the intermediate parameter. α is a hyperparameter used to control the magnitude of the change; Based on the gradients of all tasks after regularization during the t-th iteration of training, the weights are updated using the following formula; Where lr is the learning rate during iterative training. This represents the gradient of all tasks after regularization during the t-th iteration of training, in w. n The partial derivative on (t), w n (t+1) represents the weight of the nth task during the (t+1)th iteration of training.

7. The method for predicting the properties of amorphous alloy materials based on multi-task learning according to claim 5, characterized in that, Each task in the task set uses its corresponding feature extractor to extract features from the sample data, obtaining the unique feature information for each task in the task set. This process also includes: Based on the composition expression of amorphous alloy materials, the composition information of amorphous alloy materials is obtained; The component composition information is reduced in dimensionality to obtain the sample data.

8. A computer system, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the method according to any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method described in any one of claims 1-7.

Images