A quantum-complex hybrid neural network-based alloy phase structure prediction system
By constructing a multi-path dataset and combining it with a quantum-complex value hybrid neural network, the problems of low prediction accuracy and insufficient generalization ability of alloy phase structure are solved, realizing efficient and accurate prediction of alloy phase structure, which is suitable for the rapid design and optimization of various alloy materials.
Patent Information
- Application Number
- CN202610736290.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-26
- Publication Date
- 2026-11-17
AI Technical Summary
Existing methods for predicting alloy phase structures suffer from low accuracy, low efficiency, and insufficient generalization ability. Traditional real-valued neural networks do not fully extract the features of alloy elements, resulting in severe information loss and insufficient model generalization ability.
A multi-path dataset is constructed, combining quantum neural networks and complex-valued neural networks. Amplitude encoding is used to convert classical data to quantum space. By leveraging the information retention capabilities of quantum parallel computing and complex-valued neural networks, composite activation functions are employed to enhance feature extraction, thus constructing a quantum-complex-valued hybrid neural network model.
It significantly improves the accuracy and generalization performance of alloy phase structure prediction, with a prediction accuracy of over 94%, shortens the prediction cycle, reduces alloy R&D costs, and is suitable for rapid and accurate prediction of various alloy materials.
Smart Images

Figure CN123120124A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of alloy performance prediction and artificial intelligence. Specifically, it relates to an alloy phase structure prediction system based on quantum-complex value hybrid neural network, which is applicable to the rapid and accurate prediction of phase structure of various alloys, providing technical support for alloy material design and performance optimization. Background Technology
[0002] The phase structure of an alloy is a core factor determining its mechanical, physical, and chemical properties. Accurate prediction of the alloy phase structure is of great significance for shortening the alloy research and development cycle, reducing research and development costs, and optimizing alloy composition design. Traditional alloy phase structure prediction methods mainly rely on empirical formulas, phase diagram analysis, and extensive experimental verification. These methods suffer from limited prediction accuracy, low efficiency, and narrow applicability, making it difficult to meet the needs of efficient research and development of modern alloy materials.
[0003] With the development of artificial intelligence technology, machine learning and deep learning methods have been gradually applied to the field of alloy phase structure prediction, enabling rapid prediction of phase structure by constructing datasets and training models. However, existing deep learning-based prediction methods mostly employ traditional real-valued neural networks, which suffer from shortcomings such as insufficient extraction of alloy element features, severe information loss, and insufficient model generalization ability.
[0004] Quantum neural networks, with their powerful parallel computing capabilities and ability to represent complex data, offer a new approach to solving the aforementioned problems. Complex-valued neural networks, on the other hand, can effectively preserve the phase information of data, reduce information loss, and perfectly adapt to the characteristics of quantum computing, which is entirely conducted in the complex domain. Combining the two can fully leverage their respective advantages. Currently, there is no technical solution that combines quantum neural networks and complex-valued neural networks, along with multi-path dataset construction, to achieve high-precision and high-generalization prediction of alloy phase structures. Therefore, there is an urgent need for a method for predicting alloy phase structures that balances accuracy and generalization ability. Summary of the Invention
[0005] To address the problems of low accuracy, low efficiency, and insufficient generalization ability in the prediction of alloy phase structures in existing technologies, this invention provides an alloy phase structure prediction system based on a quantum-complex valued hybrid neural network. By constructing high-quality datasets through multiple approaches and combining the advantages of quantum neural networks and complex valued neural networks, it achieves accurate prediction of alloy phase structures while improving the generalization performance of the model.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] Step 1: Construct a dataset related to alloy phase structure prediction
[0008] Through experiments, simulations, literature reviews, and database searches, comprehensive information on the composition of alloys and their corresponding phase structure characteristics was collected. The alloy composition information included the types and percentages of each alloying element, while the phase structure characteristics included the types of phases. Based on the collected information, an alloy phase structure prediction dataset was established. This dataset was randomly divided into a training set and a validation set, with the division ratio adjustable according to actual needs, preferably 70%-80% for the training set and 20%-30% for the validation set. Simultaneously, a phase structure test set, entirely different from the predicted dataset, was constructed to ensure its independence and was used to verify the generalization performance of the subsequent model.
[0009] The experimental approach includes determining the composition and phase structure characteristics of alloy samples using metallographic analysis, X-ray diffraction, scanning electron microscopy, and other experimental methods. The simulation approach includes simulating the phase structure of alloys with different compositions using first-principles calculations, molecular dynamics simulations, and other methods. The literature review and database retrieval approach includes extracting publicly available information on the correspondence between alloy composition and phase structure from academic literature, patent literature, and professional alloy databases related to alloy materials, ensuring the comprehensiveness and diversity of the dataset and reducing data bias.
[0010] Step 2: Construct and train a quantum-complex hybrid neural network model.
[0011] Using the training set constructed in step 1 as input, where the input features of the training set are the percentage of each alloying element, the training process includes the following sub-steps:
[0012] 2.1 Inputting into a neural network to construct trainable element weights: The percentage of alloy elements in the training set is input into a traditional neural network. The network is trained to adaptively adjust and generate trainable element weights. These element weights are used to characterize the degree of influence of different alloy elements on the phase structure, thereby achieving preliminary enhancement and screening of alloy element characteristics.
[0013] 2.2 Amplitude Encoding and Quantum Neural Network Input: The trainable element weights obtained in step 2.1 are fused with the percentages of the corresponding alloy elements to obtain a fused feature vector; the fused feature vector is encoded into a quantum state using an amplitude encoding scheme and input into the quantum neural network to complete the conversion from classical data to quantum space, and the parallel computing capability of the quantum neural network is used to mine the complex correlation between features.
[0014] 2.3 Density Matrix Extraction and Complex-Valued Neural Network Input: The quantum neural network is run, and the density matrix output by the quantum neural network is extracted. This density matrix can completely represent the information of the quantum state and is used as the input to the complex-valued neural network. The complex-valued neural network uses a composite activation function, preferably the CReLU activation function, which performs ReLU activation on both the real and imaginary parts of the complex number, as expressed by:
[0015] CReLU(x)=ReLU(R(x))+iReLU(I(x)),
[0016] R(x) and I(x) represent the real and imaginary parts of the complex input x, respectively. ReLU(x) = max(0,x) can effectively preserve complex information, improve feature extraction capabilities, and reduce information loss.
[0017] 2.4 Final prediction result output: After processing the input density matrix, the complex-valued neural network outputs a complex-valued result containing real and imaginary parts; trainable real and imaginary weights are set, and the real part and real weights, and imaginary weights of the complex-valued result are weighted and calculated separately. Then, the weighted real part result and imaginary part result are fused to obtain the final prediction result of the alloy phase structure.
[0018] Step 3: Model Optimization and Generalization Performance Verification
[0019] The validation set constructed in step 1 is input into the trained quantum-complex valued hybrid neural network model. The performance of the model is evaluated through the validation set, with the evaluation index being prediction accuracy. The model with the best performance is selected. The selected best model is applied to the test set constructed in step 1. The generalization performance of the model is verified by comparing the prediction results of the test set with the actual phase structure characteristics, proving the universality of the model in predicting phase structures of alloys from different sources and of different types.
[0020] During the model performance evaluation process, if the prediction accuracy of the validation set is lower than the preset threshold, return to step 2, adjust the number of qubits of the quantum neural network, the number of layers and neurons of the complex-valued neural network, the initial value of the trainable weights or the parameters of the composite activation function, and retrain the model until the model performance reaches the optimal level.
[0021] The beneficial effects of this invention are as follows:
[0022] 1. Comprehensive and scientific dataset construction: Data is collected from multiple sources, including experiments, simulations, literature, and databases, to ensure the comprehensiveness and diversity of the dataset. At the same time, the dataset is divided into training and validation sets, and a separate test set from independent sources is constructed to effectively avoid data overfitting, provide reliable support for model optimization and generalization performance verification, and improve the credibility of prediction results.
[0023] 2. High prediction accuracy: The model employs a quantum-complex-valued hybrid neural network structure, combining the parallel computing advantages of quantum neural networks with the information preservation capabilities of complex-valued neural networks. It achieves the conversion of classical data to quantum space through amplitude encoding, extracts the density matrix as the input of the complex-valued neural network, and uses a composite activation function to reduce information loss. Compared with traditional real-valued neural networks, it can more fully explore the complex correlation between alloying elements and phase structure, significantly improving the prediction accuracy of phase structure. Experimental verification shows that the prediction accuracy of this model can reach over 94%, which is better than existing machine learning models and single quantum models.
[0024] 3. Excellent generalization performance: The generalization performance of the model is verified through test sets from independent sources, ensuring that the model can be applied to the prediction of phase structures of alloys of different types and composition ranges. It has strong universality and solves the problem of insufficient generalization ability of existing models. It can be widely used in the research and development and design of various alloy materials.
[0025] 4. High efficiency: Compared with traditional prediction methods that rely on a large number of experiments, this invention achieves rapid prediction of alloy phase structure through model training, which greatly shortens the prediction cycle, reduces the time and experimental costs of alloy research and development, and provides an efficient technical means for alloy composition optimization. Attached Figure Description
[0026] Figure 1 This is a flowchart of the present invention;
[0027] Figure 2 In the prediction of three types of phase structures, the accuracy change of the prediction model proposed in this invention and two benchmark models during the training process is shown in the figure.
[0028] Figure 3 In the prediction of three types of phase structures, the loss change diagram of the prediction model and neural network proposed in this invention during the training process; Detailed Implementation
[0029] The technical solutions in 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.
[0030] This embodiment provides a method for predicting alloy phase structures based on a quantum-complex valued hybrid neural network, the flowchart of which is shown below. Figure 1 As shown, it includes the following steps:
[0031] Step 1: Construct a dataset related to alloy phase structure prediction.
[0032] Step 1.1: By collecting and organizing literature and databases, a dataset for alloy phase prediction was established. This dataset contains atomic percentages and phase structure information of alloys, totaling 1756 data points, involving 43 elements, and includes three types of phase structure information: body-centered cubic (FCC), face-centered cubic (BCC), and a hybrid structure that contains both of these structures (FCC+BCC).
[0033] Step 1.2: Rewrite the alloy atomic percentage data into a 43-dimensional vector according to the periodic table order. Use a random partitioning method to divide it into a training set (1317 data points) and a validation set (439 data points) at a ratio of 3:1; at the same time, use an additional 440 data points as the test set.
[0034] Step 2: Construct and train a quantum-complex hybrid neural network model.
[0035] Step 2.1: Construct trainable element weights: Input the percentage of alloy elements in the training set (the input feature dimension is the number of alloy element types) into a traditional 3-layer neural network (the number of neurons in the input layer = the number of element types (43), the number of neurons in the hidden layer are 64 and 128 respectively, and the number of neurons in the output layer = the number of element types (43). Train the network through the backpropagation algorithm to generate trainable weights corresponding to each alloy element, which are used to characterize the influence weights of different elements on the phase structure.
[0036] Step 2.2: Amplitude Encoding and Quantum Neural Network Input: Multiply the weights of the trainable elements by the corresponding element percentages to obtain the fused feature vector; using an amplitude encoding scheme, normalize the fused 43-dimensional feature vector and encode it into a quantum state, which is then input into a quantum neural network containing 6 qubits. The quantum neural network uses parameterized quantum circuits, including Hadamard gates, CNOT gates, and RY rotation gates, to mine complex correlations between features.
[0037] Step 2.3: Density matrix extraction and complex-valued neural network input: Run the quantum neural network, measure and extract the 64-dimensional density matrix of the output, and use it as the input of the complex-valued neural network; the complex-valued neural network adopts a 4-layer structure (the number of neurons in the input layer is 64, the number of neurons in the hidden layer is 128, 64, and 32 respectively, and the number of neurons in the output layer is the number of categories (3)). The CReLU composite activation function is used to activate the real and imaginary parts of the complex-valued input after the density matrix transformation, respectively, to reduce information loss.
[0038] Step 2.4: Final prediction result output: Set the real part trainable weights and imaginary part trainable weights (initial value is set to 0.5 for both). The real part and real part weights and imaginary part weights of the complex value result output by the complex value neural network are weighted and summed respectively. After fusion, the prediction result of the alloy phase structure is obtained (the predicted phase structure type is body-centered cubic, face-centered cubic or hybrid structure).
[0039] Step 3: Model selection and generalization performance verification.
[0040] 3.1 Model Optimization: The validation set is input into the trained hybrid neural network model, the prediction accuracy of the model is calculated, and finally the optimal model with a prediction accuracy of 93.17% is selected.
[0041] 3.2 Generalization performance verification: The test set was input into the optimal model. The prediction accuracy of the test set was 93.64%, which was close to that of the validation set. Moreover, the deviation between the prediction results and the actual phase structure characteristics of the test set was small, which proves that the model has good generalization performance and universality and can be applied to the prediction of phase structure of alloys from different sources and of different types.
[0042] Figure 2 The accuracy changes during training on the training set were compared between the proposed model (QCVNN) and neural networks (NN) and quantum neural networks (QNN). The proposed model achieves higher accuracy compared to the baseline model.
[0043] Figure 3 The presentation shows the loss variation of the proposed model and neural network on the training set during training. It can be observed that the proposed model's loss decreases faster and ultimately achieves a lower loss.
Claims
1. A system for predicting alloy phase structure based on a quantum-complex valued hybrid neural network, characterized in that, Includes the following steps: Step 1: Construct a dataset for predicting alloy phase structure: Collect information on the composition of alloys and their corresponding phase structure characteristics through experiments, simulations, literature reviews, and database searches to establish an alloy phase structure prediction dataset; randomly divide this dataset into a training set and a validation set, and simultaneously construct a phase structure test set from sources different from the aforementioned phase structure prediction dataset; Step 2: Construct and train a quantum-complex valued hybrid neural network model: using the training set as input, the input features of which are the percentages of each alloy element, the training process includes.
2. Constructing trainable element weights using neural networks: Input the percentage of alloy elements in the training set into a traditional neural network to train and generate trainable element weights; Amplitude encoding and quantum neural network input: The element weights are fused with the corresponding alloy element percentages, encoded into quantum states using an amplitude encoding scheme, and then input into the quantum neural network; Density matrix extraction and complex-valued neural network input: Extract the density matrix from the output of the quantum neural network and input it into a complex-valued neural network using a compound activation function; Final prediction output: The real and imaginary parts of the complex value output by the complex value neural network are combined with the corresponding trainable weights, and the final prediction result is output after fusion. Step 3: Model optimization and generalization performance verification: Use the validation set to select the model with the best performance, and use the test set to verify the generalization performance of the model to prove its universality. According to the method described in claim 1, in step 1, the alloy composition information includes only the type and percentage of each alloy element; the ratio of the training set to the validation set is 7:
3.
3. The method according to claim 1, characterized in that, In step 1, the experimental approach includes determining the alloy composition and phase structure through metallographic analysis, X-ray diffraction, and scanning electron microscopy; the simulation approach includes first-principles calculations and molecular dynamics simulations; the alloy types, composition ranges, or phase structure types in the test set do not overlap with the phase structure prediction dataset.
4. The method according to claim 1, characterized in that, In step 2.1, the traditional neural network has a 3-layer structure, with the number of neurons in the input layer equal to the number of alloy element types, one hidden layer, and the number of neurons in the output layer equal to the number of alloy element types. The trainable element weights are generated through backpropagation algorithm.
5. The method according to claim 1, characterized in that, In step 2.2, the fusion is a multiplication operation of the trainable element weights and the corresponding alloy element percentages. The encoding scheme is amplitude encoding, and the quantum neural network includes qubits, Hadamard gates, CNOT gates, and RY rotation gates.
6. The method according to claim 1, characterized in that, In step 2.3, the composite activation function is the CReLU activation function, and its expression is: CReLU(x)=ReLU(R(x))+iReLU(I(x)), Where R(x) and I(x) represent the real and imaginary parts of the complex input x, respectively, and ReLU(x) = max(0,x).
7. The method according to claim 1, characterized in that, In step 2.4, the trainable weights include real part trainable weights and imaginary part trainable weights, both with an initial value of 0.
5. The results of the real part and the imaginary part are fused by weighted summation.
8. The method according to claim 1, characterized in that, In step 3, the model performance evaluation metrics include prediction accuracy; if the prediction accuracy of the validation set is lower than a preset threshold, the number of qubits in the quantum neural network, the number of layers and neurons in the complex-valued neural network, the initial values of trainable weights, or the parameters of the composite activation function are adjusted, and the model is retrained.
