Method and apparatus for predicting molecular properties based on quantum quenching kinetic characteristic enhancement

Abstract

The application discloses a method for predicting molecular properties based on quantum quenching dynamics characteristic enhancement, comprising the following steps: obtaining parameters of a quantum many-body Hamiltonian by normalization and mapping based on a structural descriptor of a molecule, wherein the parameters are used for encoding structural characteristics of the molecule; performing time-ordered evolution by a quantum processor based on the parameters of the quantum many-body Hamiltonian, a preset initial driving Hamiltonian and a scheduling function, driving the system from an initial direct product state to an entangled state under non-adiabatic conditions, extracting monomer expectation values and two-body correlation function values based on the entangled state by local measurement, and splicing and extracting a quantized enhanced feature vector; and training and optimizing the generalization performance of a classical machine learning model under a supervised learning framework based on the quantum enhanced feature vector and a corresponding molecular property label, and outputting a prediction value of a molecular electrochemical property.

Description

Technical Field

[0001] This invention belongs to the field of quantum computing and machine learning technology, and in particular relates to a method and device for predicting molecular properties based on the enhancement of quantum quenching dynamics characteristics. Background Technology

[0002] In drug discovery and materials science, accurate prediction of molecular electrochemical properties is crucial for the screening and design of high-performance battery materials, catalysts, and drug molecules. However, the precise acquisition of these properties typically relies on first-principles methods such as solving the nucleus-electron many-body Schrödinger equation or density functional theory (DFT). As the complexity of molecular systems increases, the computational costs of these methods rise exponentially. Especially for strongly correlated electron systems, traditional quantum chemical methods face significant computational bottlenecks, resulting in a large number of candidate molecules not being adequately evaluated in virtual screening, severely limiting the efficiency of discovering new materials and drugs.

[0003] To overcome the aforementioned computational bottlenecks, machine learning methods have been widely applied in the prediction of molecular properties. In existing techniques, researchers have constructed molecular structure descriptors based on the types and coordination numbers of atoms in a molecule, combined with redox potentials calculated using DFT to establish a training set, and utilized Gaussian process regression models to predict the redox potential of electrolyte additives in lithium-ion batteries. In recent years, machine learning algorithms based on quantum computing principles have been considered to have potential advantages over classical methods in handling complex information and high-dimensional data correlation characteristics. For example, some studies have proposed a method to extract data patterns using the quenching dynamics of quantum spin glass, enhancing the performance of classical classifiers by using entangled states generated by rapid non-adiabatic evolution as feature mapping encoding on specific datasets.

[0004] However, existing technologies still have the following shortcomings: First, existing variable quantum algorithms typically require extensive parameter tuning, are prone to getting trapped in local optima, and are difficult to operate stably on noisy medium-scale quantum devices; second, existing quantum feature mapping methods are mostly limited to static ground-state encoding, failing to fully utilize the rich feature information contained in the superposition and entanglement structures of excited states generated during the non-adiabatic evolution of quantum systems; third, although pure classical machine learning methods have been explored in the prediction of molecular electrochemical properties, their prediction accuracy is limited by the expressive power of molecular descriptors, making it difficult to fully reflect the global electronic structure characteristics and quantum effects of molecules, resulting in systematic biases in the prediction results. Therefore, there is an urgent need for a method that can fully utilize the non-adiabatic dynamic evolution capabilities of quantum mechanics and significantly improve the prediction accuracy of molecular electrochemical properties while maintaining low computational costs. Summary of the Invention

[0005] To address the aforementioned issues, this application proposes a molecular property prediction method and apparatus based on quantum quenching dynamics feature enhancement. By using non-adiabatic quantum quenching, molecular descriptors are mapped to quantum-enhanced features carrying quantum correlation information, thereby improving the accuracy of electrochemical property predictions using classical machine learning models. Replacing traditional first-principles calculations with quantum-classical hybrid prediction reduces the cost of obtaining property parameters such as redox potential and solvation energy, improving the screening hit rate. Utilizing highly entangled states generated by non-adiabatic excitations, rather than static ground states, as the feature source overcomes the limitations of existing quantum feature mapping methods, generating richer features.

[0006] The first aspect of this application provides a method for predicting molecular properties based on enhanced quantum quenching dynamics features, including: Based on the molecular structure descriptor, the parameters of the quantum many-body Hamiltonian are obtained through normalization and mapping, and these parameters are used to encode the structural features of the molecule. Based on the parameters of the quantum many-body Hamiltonian, the preset initial driving Hamiltonian, and the scheduling function, the quantum processor executes time-dependent evolution in a time-sequential manner, driving the system from the initial direct product state to the entangled state under non-adiabatic conditions. The entangled state carries the structural information of the molecules. Based on the entangled state, the expected value of a single entity and the correlation function value of two entities are extracted by local measurement and then concatenated to extract a numerical quantum enhanced feature vector; Based on quantum-enhanced feature vectors and corresponding molecular property labels, a classical machine learning model is trained and optimized under a supervised learning framework to improve generalization performance, outputting predicted values ​​of molecular electrochemical properties, including at least one of redox potential, inner-shell recombination energy, solvation energy, and frontier orbital energy.

[0007] Preferred options also include: Calculate at least one of the correlation coefficient, mean square error, root mean square error, and mean absolute error between the predicted value and the true value, and use it as a performance evaluation index of the classical machine learning model. The expression for calculating the correlation coefficient is as follows:

[0008] In the formula, Let be the predicted value for the i-th sample. , These are the sample means of the predicted value and the actual value, respectively. To represent the true value of the i-th sample, n is the number of samples; The expression for calculating the mean squared error is:

[0009] In the formula, Mean square error; The formula for calculating the root mean square error is:

[0010] In the formula, This is the root mean square error; The expression for calculating the mean absolute error is: .

[0011] Preferably, the parameters of the quantum many-body Hamiltonian can be obtained using any one of the transverse field Ising model, Heisenberg model, Hubbard model, and XY model. If the transverse field Ising model is used, the calculation expression for the model structure is:

[0012] In the formula, , Here, represents the longitudinal magnetic field parameter acting on the i-th qubit, and the transverse coupling strength between the i-th and j-th qubits, respectively. For the Pauli-Z operator of the i-th qubit, , The Pauli-X operators for the i-th and j-th qubits are respectively; If the Heisenberg model is used, the calculation expression for the model structure is:

[0013] In the formula, , , These represent the spin coupling strength in the first direction, the spin coupling strength in the second direction, and the spin coupling strength in the third direction, respectively. For the Pauli-Z operator of the i-th qubit, , The Pauli-Y operators for the i-th and j-th qubits are respectively; If the Hubbard model is used, the calculation expression for the model structure is:

[0014] In the formula, For the electron jump integral, This represents the strength of the Coulomb interaction in its proper position. Chemical potential For an electron to jump from the j-th lattice point to the i-th lattice point, For Hermitian conjugate, , Let be the particle number operator that spins at the i-th lattice point in the first direction and the particle number operator that spins at the i-th lattice point in the opposite direction to the first direction, respectively. For the electron at the i-th lattice point with its spin pointing in the first direction Quantity; If the XY model is used, the calculation expression for the model structure is:

[0015] In the formula, The spin coupling strength is in the first plane.

[0016] Preferably, the step of extracting the expected value of a single entity and the correlation function value between the two entities further includes: When the quantum many-body Hamiltonian is a spin system model, the single-unit Pauli operator expectation value is measured for each qubit of the entangled state, and the two-qubit correlation function value is measured for adjacent qubits; When the quantum many-body Hamiltonian is a fermionic system model, the expected value of the single-lattice particle population is measured for each lattice point of the entangled state, and the value of the double-lattice correlation function is measured for adjacent lattice points. The measured single-body expected value is concatenated with the two-body correlation function value to form the quantum enhanced feature vector.

[0017] Preferably, the step of driving the system from the initial product state to the entangled state further includes: The quantum system is prepared with the ground state of the initial driving Hamiltonian set as an entangled direct product state; The normalized evolution time changes continuously from 0 to 1. The scheduling function controls the weight of the initial driving Hamiltonian to decrease from 1 to 0 and the weight of the problem Hamiltonian to increase from 0 to 1, and constructs a time-dependent Hamiltonian. The calculation expression is as follows:

[0018] In the formula, , Let be the scheduling functions for the initial driving Hamiltonian and the problem Hamiltonian, respectively, and s be the normalized evolution time. , These are the initial driving Hamiltonian and the problem Hamiltonian, respectively. Under rapid non-adiabatic conditions, the evolution from the initial product state to the entangled state is calculated using the following expression:

[0019] In the formula, It is a direct product state without entanglement.

[0020] Preferably, the expression for calculating the parameters of the quantum many-body Hamiltonian is:

[0021] In the formula, For the filtered feature data, Let be the coupling strength between the i-th qubit and the j-th qubit. , Let X be the Pauli operator for the i-th qubit and Z be the Pauli operator for the i-th qubit.

[0022] Preferably, the classical machine learning model is a Gaussian process regression model, and the kernel function of the classical machine learning model is calculated as follows:

[0023] In the formula, For constant kernel coefficients, The length scale is based on the Matern kernel. The scale is the dot product kernel.

[0024] Preferably, the structural descriptor of the molecule is obtained by at least one of the following methods: statistical calculation based on the molecule's SMILES string or chemical formula; or calculation of the molecular structure using a cheminformatics toolkit.

[0025] Preferably, the molecular-based structural descriptor specifically includes: The structured descriptors of the molecules are screened for mutual information features, and descriptors whose mutual information values ​​meet preset conditions are selected as the structured descriptors input to the normalization and mapping steps; The expression for calculating the mutual information value is:

[0026] In the formula, X represents a single molecule descriptor feature, and Y represents the target value of the molecular electrochemical property to be predicted. Let X and Y be the joint probability density function. , Let X and Y be the marginal probability density functions, respectively.

[0027] A second aspect of this application provides a molecular property prediction device based on enhanced quantum quenching dynamics features, comprising: The molecular feature encoding module is used to obtain the parameters of the quantum many-body Hamiltonian based on the molecular structure descriptor through normalization and mapping. These parameters are used to encode the structural features of the molecule. The quantum nonadiabatic evolution module is used to perform time-dependent evolution in a time-sequential manner through a quantum processor based on the parameters of the quantum many-body Hamiltonian, a preset initial driving Hamiltonian, and a scheduling function, thereby driving the system from an initial direct product state to an entangled state under nonadiabatic conditions. The entangled state carries the structural information of the molecules. The quantum feature extraction module is used to extract the single-unit expected value and the two-unit correlation function value based on the entangled state through local measurement, and then concatenate them to extract a numerical quantum enhanced feature vector; The machine learning prediction module is used to train and optimize the generalization performance of a classical machine learning model under a supervised learning framework based on quantum-enhanced feature vectors and corresponding molecular property labels, and output predicted values ​​of molecular electrochemical properties, including at least one of redox potential, inner-shell recombination energy, solvation energy and frontier orbital energy.

[0028] Because this application adopts the above technical solution, it has the following advantages and positive effects compared with the prior art: By mapping the molecular structure descriptor to quantum many-body Hamiltonian parameters and performing non-adiabatic evolution, molecular structure information is encoded in the final quantum state, realizing the quantized representation of molecular structure features and providing a feature source rich in quantum correlation information for subsequent machine learning predictions.

[0029] By employing any one of the transverse field Ising model, Heisenberg model, Hubbard model, or XY model to construct the problem Hamiltonian, we can flexibly select the most suitable quantum many-body model for different molecular system characteristics, enabling molecular descriptors to be encoded into quantum system parameters in an optimal manner, thereby improving the accuracy and versatility of feature mapping.

[0030] By extracting the expected value of a single cell and the correlation function value of two cells through local measurement of entangled states and concatenating them into a quantum-enhanced feature vector, quantum correlation information that is difficult to simulate by classical computers is transformed into numerical features, enabling the original molecular descriptor to obtain a high-dimensional nonlinear mapping and enhancing the feature's ability to characterize the electrochemical properties of molecules.

[0031] By inputting quantum-enhanced feature vectors into a classical machine learning model and training it under a supervised learning framework, and combining cross-validation and early stopping strategies to optimize generalization performance, high-precision prediction of electrochemical properties such as molecular redox potential, inner-shell recombination energy, solvation energy, and frontier orbital energy is achieved. This significantly reduces the computational cost of traditional quantum chemical methods and improves the virtual screening hit rate.

[0032] By using entangled states generated by non-adiabatic evolution instead of adiabatic ground states as the source of features, we can fully utilize the dynamic evolution capabilities of quantum systems, overcome the limitations of existing static ground state encoding methods, and generate richer and more beneficial quantum features for classification / regression tasks. Attached Figure Description

[0033] The specific embodiments of this application will be further described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is the main flowchart of this application; Figure 2 This is a comparative diagram of the present application and the conventional solution; Figure 3 This is a schematic diagram of the paths for the initial driving Hamiltonian and the problem Hamiltonian. Detailed Implementation

[0034] The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. The advantages and features of the present application will become clearer from the following description and claims. It should be noted that the drawings are all in a very simplified form and use non-precise ratios, and are only used to facilitate and clarify the illustration of the embodiments of the present application.

[0035] It should be noted that all directional indicators (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicator will also change accordingly.

[0036] First Embodiment See Figure 1 and Figure 2 The first aspect of this application provides a method for predicting molecular properties based on enhanced quantum quenching dynamics, comprising: Based on the molecular structure descriptor, the parameters of the quantum many-body Hamiltonian are obtained through normalization and mapping. The parameters are used to encode the structural features of the molecule. Based on the parameters of the quantum many-body Hamiltonian, the preset initial driving Hamiltonian, and the scheduling function, the quantum processor executes the time-dependent evolution in a time-sequential manner, driving the system from the initial direct product state to the entangled state under non-adiabatic conditions. The entangled state carries the structural information of the molecules. Based on entangled states, the expected value of a single entity and the correlation function value of two entities are extracted through local measurement, and then concatenated to extract a numerical quantum enhanced feature vector; Based on quantum-enhanced feature vectors and corresponding molecular property labels, a classical machine learning model is trained and optimized under a supervised learning framework to improve generalization performance, outputting predicted values ​​of molecular electrochemical properties, including at least one of redox potential, inner-shell recombination energy, solvation energy, and frontier orbital energy.

[0037] This application utilizes quantum quenching dynamics to map classical molecular descriptors into quantum-enhanced features. Through non-adiabatic evolution, it generates quantum correlation information that is difficult for classical computers to simulate. This information is then combined with a classical machine learning model to achieve high-precision prediction of molecular electrochemical properties. The structural descriptor of the molecule is normalized and mapped to parameters of a quantum many-body Hamiltonian, encoding molecular structural features into the quantum system as physical parameters. Subsequently, a scheduling function controls the time-dependent Hamiltonian to perform rapid non-adiabatic evolution, causing the system to transition from an unentangled direct product state to a highly entangled superposition of excited states. During this process, molecular structural information is transformed into entangled structural features of the quantum state. By performing local measurements on the final quantum state, the expected value of the monomer and the correlation function value of the two-body state are extracted and concatenated into a numerical quantum-enhanced feature vector. Finally, this feature vector is input into a classical machine learning model for supervised learning training, completing end-to-end prediction from molecular structure to electrochemical properties. This application significantly improves the accuracy of electrochemical property prediction by mapping molecular descriptors into enhanced features containing quantum correlation information through non-adiabatic quantum quenching, thus enhancing the accuracy of the classical machine learning model. This application replaces traditional DFT first-principles calculations with quantum-classical hybrid prediction, rapidly obtaining property parameters such as redox potential, recombination energy, and solvation energy without requiring expensive molecule-by-molecule quantum chemical calculations, significantly reducing computational costs and improving virtual screening efficiency. This application utilizes entangled states generated by non-adiabatic excitations, rather than static ground states, as feature sources, overcoming the limitation of existing quantum feature mappings being restricted to static ground state encoding, thus enabling feature vectors to contain richer quantum correlation information. This application supports flexible selection of four quantum many-body Hamiltonians: the transverse field Ising model, the Heisenberg model, the Hubbard model, and the XY model, adaptable to the electronic structure characteristics of different molecular systems, and suitable for predicting various electrochemical properties such as redox potential, inner-shell recombination energy, solvation energy, and frontier orbital energy.

[0038] In this embodiment, the dataset preprocessing process is as follows: The input data is in Excel format, containing the target column Vox / Vred (continuous numerical values ​​representing redox potentials), and the remaining 22 columns are numerical molecular descriptors. Based on mutual information regression, the Top 10 features with the strongest correlation to the target variable are selected to reduce the number of qubits while retaining key information; the test set is randomly divided at a ratio of 10%, and the training set at 90%, ensuring consistent data distribution.

[0039] Taking the redox potential prediction task of lithium-ion battery electrolyte additives as an example, the original features contain 22 molecular descriptors: ['B4','B3','C4','C3','C2','N3','N2','N1','O2','O1','F1','Si4','P4','P3','S4','S3','S2','Cl1','BR1','R5','R6','rad'] The top 10 features are selected using mutual information, corresponding to 10 qubits. The sample size in this embodiment is 149, but other parameters can also be used. This embodiment does not impose any restrictions.

[0040] Figure 2 Figure 'a' in the diagram shows a performance comparison between quantum-enhanced features and classical features in the redox potential prediction task. In terms of correlation coefficient (CC), the quantum-enhanced features achieved 74.8%, a 14.7 percentage point improvement over the classical baseline of 60.1%. In terms of mean square error (MSE), the quantum-enhanced features achieved 0.70, a 16.7% reduction compared to the classical baseline of 0.84. In terms of root mean square error (RMSE), the quantum-enhanced features achieved 0.84, an 8.7% reduction compared to the classical baseline of 0.92. In terms of mean absolute error (MAE), the quantum-enhanced features achieved 0.60, a 10.4% reduction compared to the classical baseline of 0.67. All four metrics comprehensively outperform the classical feature baseline, indicating that the quantum-enhanced features extracted through quantum quenching dynamics in this application can effectively capture molecular structure information related to redox potential, significantly improving prediction accuracy and stability. Figure 2 Figure b shows the performance comparison between quantum-enhanced features and classical features in the inner-layer recombination energy prediction task. In terms of correlation coefficient (CC), the quantum-enhanced feature achieved 70.0%, a 20.3 percentage point improvement over the classical baseline of 49.7%. In terms of mean squared error (MSE), the quantum-enhanced feature was 0.51, a 42.0% reduction from the classical baseline of 0.88. In terms of root mean square error (RMSE), the quantum-enhanced feature was 0.72, a 23.4% reduction from the classical baseline of 0.94. In terms of mean absolute error (MAE), the quantum-enhanced feature was 0.49, a 36.4% reduction from the classical baseline of 0.77. All four metrics comprehensively outperformed the classical feature baseline, with a 42.0% reduction in MSE, further validating the significant advantage of the quantum-enhanced feature in molecular property prediction. The mean squared error (MSE), root mean square error (RMSE), and mean absolute error (MAE) are used to measure model stability; closer to 0 is better. Monomer expected value. Measuring each quantum bit This yields a 10-dimensional vector. .

[0041] See Figure 2The control group used classical features (selected Top-K features) and Gaussian process regression; the experimental group used quantum features (single-qubit expectation value, two-body correlation) and Gaussian process regression. Fairness was guaranteed: except for the quantum component, all other components strictly maintained control variables. This invention, through quantum quenching dynamics-extracted quantum enhancement features, can provide classical machine learning models with more effective and richer molecular structure information, thereby significantly improving the predictive performance of molecular electrochemical properties.

[0042] Preferred options also include: Calculate at least one of the following: correlation coefficient, mean square error, root mean square error, and mean absolute error between predicted and true values, and use it as a performance evaluation metric for classic machine learning models. The formula for calculating the correlation coefficient is as follows:

[0043] In the formula, Let be the predicted value for the i-th sample. , These are the sample means of the predicted value and the actual value, respectively. To represent the true value of the i-th sample, n is the number of samples; The expression for calculating the mean squared error is:

[0044] In the formula, Mean squared error; The formula for calculating the root mean square error is:

[0045] In the formula, This is the root mean square error; The expression for calculating the mean absolute error is: .

[0046] Preferably, the parameters of the quantum many-body Hamiltonian can be obtained using any one of the transverse field Ising model, Heisenberg model, Hubbard model, and XY model. If the transverse field Ising model is used, the calculation expression for the model structure is:

[0047] In the formula, , Here, represents the longitudinal magnetic field parameter acting on the i-th qubit, and the transverse coupling strength between the i-th and j-th qubits, respectively. For the Pauli-Z operator of the i-th qubit, , The Pauli-X operators for the i-th and j-th qubits are respectively; If the Heisenberg model is used, the calculation expression for the model structure is:

[0048] In the formula, , , These represent the spin coupling strength in the first direction, the spin coupling strength in the second direction, and the spin coupling strength in the third direction, respectively. For the Pauli-Z operator of the i-th qubit, , The Pauli-Y operators for the i-th and j-th qubits are respectively; If the Hubbard model is used, the calculation expression for the model structure is:

[0049] In the formula, For the electron jump integral, This represents the strength of the Coulomb interaction in its proper position. Chemical potential, For an electron to jump from the j-th lattice point to the i-th lattice point, For Hermitian conjugate, , Let be the particle number operator that spins at the i-th lattice point in the first direction and the particle number operator that spins at the i-th lattice point in the opposite direction to the first direction, respectively. For the electron at the i-th lattice point with its spin pointing in the first direction Quantity; If the XY model is used, the calculation expression for the model structure is:

[0050] In the formula, The spin coupling strength is in the first plane.

[0051] This application supports any one of four quantum many-body Hamiltonian models: the transverse field Ising model, the Heisenberg model, the Hubbard model, and the XY model. By mapping the components of the normalized molecular descriptor to physical parameters (such as coupling strength, magnetic field strength, electron hopping integral, chemical potential, etc.) in the selected model, molecular structural features can be encoded in the quantum system through different physical mechanisms, thereby adaptively selecting the optimal encoding model based on the electronic structure characteristics of the target molecular system. This application quantitatively evaluates the predictive performance of the classical machine learning model by calculating at least one of the following evaluation metrics: correlation coefficient between predicted and true values, mean square error, root mean square error, and mean absolute error, thereby verifying the effectiveness and generalization ability of the quantum-enhanced feature extraction method.

[0052] Preferably, the step of extracting the expected value of a single entity and the correlation function value between the two entities further includes: When the quantum many-body Hamiltonian is a spin system model, the single-bit Pauli operator expectation value is measured for each qubit of the entangled state, and the two-bit correlation function value is measured for adjacent qubits. When the quantum many-body Hamiltonian is a fermionic system model, the expected value of the single-lattice particle population is measured for each lattice point of the entangled state, and the value of the double-lattice correlation function is measured for adjacent lattice points. The measured single-body expected value is concatenated with the two-body correlation function value to form a quantum-enhanced eigenvector.

[0053] For two common physical modeling forms of quantum many-body systems—spin-based and fermionic-based models—different local observation extraction schemes are adapted for each. The extracted information is then uniformly encoded into a quantum-enhanced feature vector with the same structure for subsequent quantum machine learning or quantum many-body tasks. Specifically, when the quantum many-body Hamiltonian to be processed is a spin-based model, this method measures the expectation values ​​of three Pauli operators for each qubit in the entangled state to obtain the polarization information of a single qubit. Simultaneously, it measures the two-qubit correlation function value for each pair of adjacent qubits to obtain the quantum correlation information between adjacent spins. When the quantum many-body Hamiltonian is a fermionic-based model, this method measures the expectation value of the single-lattice particle population for each lattice point in the entangled state to obtain the occupancy probability of that lattice point. Simultaneously, it measures the two-lattice correlation function value for each pair of adjacent lattice points to obtain the density correlation or transition correlation information between adjacent lattice points. After completing the above measurements, the method concatenates all the single-unit expectation values ​​and two-unit correlation function values ​​in a fixed order into a complete feature vector, which is the quantum-enhanced feature vector. From a physical perspective, the local physical information of quantum many-body systems is mainly distributed at the individual and two-body levels. Individual quantities reflect the state of each local degree of freedom, while two-body quantities reflect the interaction and correlation between adjacent degrees of freedom. Correlation functions of three-body systems and higher orders typically contribute little to macroscopically observable quantities and can be reasonably ignored with limited measurement resources. From an efficiency perspective, the total number of individual and adjacent two-body measurements is linearly related to the system size; that is, the number of measurements increases linearly only with the number of qubits or lattice points, lower than the exponential number of measurements required for complete quantum state tomography. This makes the method highly scalable in quantum many-body systems. Furthermore, by uniformly mapping the observables of two different physical systems to isomorphic feature vectors, downstream quantum machine learning models do not need to design different input layers for spin and fermionic systems, thus achieving cross-physical system data format standardization. It achieves a unified cross-system representation of spin and fermionic systems, mapping models of two different underlying physical variables to isomorphic quantum enhancement feature vectors, eliminating the need for downstream models to redesign input layers for different systems. The measurement complexity is reduced, with the total number of single-body measurements and adjacent-body measurements linearly related to the system size. Compared to the exponential number of measurements in complete quantum state tomography, this represents exponential compression and offers scalability. By introducing adjacent-body correlation functions, local quantum correlation information at the mean-field level is captured, carrying richer physical information than traditional schemes that only use single-body expectation values. Physical interpretability is maintained; each component in the eigenvector has a clear physical meaning, facilitating the comparison and verification of model decisions with underlying physical laws. The measurement operations involved are all mature and implementable basic measurement schemes on mainstream quantum platforms, requiring no complex multi-body joint measurements and demonstrating good engineering feasibility.

[0054] Among them, the two-body relationship This yields another 10-dimensional vector. The expected value of each individual sample is then concatenated with the correlation between the two samples to obtain a 20-dimensional quantum feature vector for each sample. This vector serves as a nonlinear high-dimensional mapping of the original molecular descriptor, which is used in subsequent regression models.

[0055] See Figure 3 Preferably, the step of driving the system from the initial product state to the entangled state further includes: The quantum system is prepared with the initial ground state of the driving Hamiltonian set as an unentangled product state; The normalized evolution time varies continuously from 0 to 1. The weights of the initial driving Hamiltonian are controlled by the scheduling function, decreasing from 1 to 0, while the weights of the problem Hamiltonian are increased from 0 to 1. A time-dependent Hamiltonian is constructed, and its calculation expression is as follows:

[0056] In the formula, , Here, are the scheduling functions for the initial driving Hamiltonian and the problem Hamiltonian, respectively, and s is the normalized evolution time. , These are the initial driving Hamiltonian and the problem Hamiltonian, respectively. Under rapid nonadiabatic conditions, the evolution from the initial product state to the entangled state is calculated using the following expression:

[0057] In the formula, It is a direct product state without entanglement.

[0058] This method first prepares the quantum system in the ground state of the initial driving Hamiltonian H_initial, which is an entangled direct product state |ψ>_i, easy to prepare and physically well-defined and controllable. Then, a normalized evolution time s is introduced, continuously varying from 0 to 1. Scheduling functions A(s) and B(s) control the weights of the initial driving Hamiltonian and the problem Hamiltonian, respectively, decreasing A(s) from 1 to 0 and increasing B(s) from 0 to 1, thereby constructing a time-dependent Hamiltonian. During the evolution, the system undergoes unitary evolution driven by a time-dependent Hamiltonian. Finally, under rapid nonadiabatic conditions, the system transitions from the initial product state. Starting from a point, it evolves along a time-dependent path to the target entangled state; its evolution process is based on an integral formula. The description expresses the superposition and cumulative evolution effect of continuous quantum states along the entire normalized evolution path of the system. By designing a continuous path connecting the initial product state and the target entangled state, the quantum state is driven to evolve along this path using a time-dependent Hamiltonian, and the deterministic preparation of the entangled state is achieved using unitary evolution in quantum mechanics. The introduction of scheduling functions A(s) and B(s) provides additional degrees of freedom for control, which can be used to optimize the shape of the evolution path, suppress non-adiabatic transitions, and achieve high-fidelity state transfers under non-adiabatic conditions. The transverse field strength decreases linearly from 1 to 0, corresponding to the annealing process in quantum annealing. The problem is that the Hamiltonian strength increases linearly from 0 to 1, gradually encoding classical features into the quantum state.

[0059] In this embodiment, the total evolution time is set to τ = 30 ns, the number of steps is set to steps = 300, and the time step size is... The time is 0.1 ns. The initial state of the system is prepared as follows: This means that all qubits are in the computational ground state. It should be noted that the specific values ​​of the total evolution time and the number of off-steps can be adaptively adjusted according to the coherence time of the quantum processor, the required simulation accuracy, and computational resources. For example, when the number of qubits is large or higher simulation accuracy is required, the total evolution time and the number of off-steps can be appropriately increased; when the coherence time of the quantum processor is short, the total evolution time can be shortened accordingly to ensure that the evolution is completed before decoherence.

[0060] Preferably, the expression for calculating the parameters of the quantum many-body Hamiltonian is:

[0061] In the formula, For the filtered feature data, Let be the coupling strength between the i-th qubit and the j-th qubit. , Let X be the Pauli operator for the i-th qubit and Z be the Pauli operator for the i-th qubit.

[0062] The first term is the interaction term. This represents the coupling energy between adjacent qubits. The coupling strength between the i-th qubit and the j-th qubit is determined by the filtered feature data x. and Let X be the Pauli operator for the i-th and j-th qubits, respectively. Let be the Pauli Z operator for the i-th qubit. This term describes the tendency for adjacent spins to align with or out of alignment, reflecting the interaction energy within the system. (Summation notation) This indicates that the summation is performed only on adjacent pairs of qubits, reflecting the locality of this interaction. The second term is the transverse magnetic field term. This represents the energy contribution of each qubit under the influence of an external field in the Z direction. Let be the local magnetic field strength at the i-th qubit. Let be the Pauli Z operator for the i-th qubit. This term describes the orientation preference of each spin under the influence of an external field in the Z direction. The negative sign indicates that the system tends to align the spins along the positive Z direction to reduce energy. This expression is expressed by adjusting... and The relative magnitudes of these terms can be used to control the quantum phase transition behavior of a quantum system: when the transverse magnetic field term dominates, the system is in a paramagnetic phase, i.e., the spins are aligned along the Z direction; when the interaction term dominates, the system enters a ferromagnetic phase, i.e., the spins are ordered in the X direction; when the two terms compete, quantum entanglement may occur. The parameters in the formula... and All of these originate from the filtered feature data x, which means that the problem Hamiltonian can dynamically adjust its parameter configuration according to the input data, enabling the system to adaptively construct the corresponding quantum many-body model for different specific problems.

[0063] Optionally, the initial driving Hamiltonian is the transverse field driving Hamiltonian, and its calculation expression is as follows:

[0064] In the formula, Let X be the Pauli X operator for the i-th qubit, and N represent the total number of qubits. The initial driving Hamiltonian is used to provide the initial transverse field for quantum quenching, and its ground state is an unentangled direct product state.

[0065] Optionally, the operators in the time-dependent Hamiltonian are constructed into a high-dimensional matrix through the Kronecker product, specifically including: the Pauli X operator of the i-th qubit. The Kronecker product is constructed as follows:

[0066] In the formula, The X operator for the i-th qubit, Let be the identity matrix acting on all n qubits, where n is the total number of qubits. The Kronecker product of the two-body correlation function is constructed as follows:

[0067] In the formula, i <j, For Pauli Z operator, Let Kronecker product be the product of k identity matrices.

[0068] This application sets the initial driving Hamiltonian as a transverse field-driven Hamiltonian with an unentangled ground state. Starting from an easily prepared initial state, the system undergoes time-dependent evolution, during which quantum fluctuations are generated by the transverse field, promoting the transition from a product state to a highly entangled state. This provides controllable initial conditions and a physical driving mechanism for non-adiabatic evolution. By extending single-qubit and two-body correlation operators to high-dimensional matrices through the Kronecker product, a rigorous mathematical mapping of operators from single-qubit space to the full Hilbert space is achieved, ensuring the accuracy and computability of the Hamiltonian construction and providing a complete mathematical foundation for time-dependent evolution.

[0069] Preferably, the classic machine learning model is the Gaussian process regression model, and the kernel function of the classic machine learning model is calculated as follows:

[0070] In the formula, For constant kernel coefficients, The length scale is based on the Matern kernel. The scale is the dot product kernel.

[0071] Quantum-enhanced eigenvectors integrate the expected value of a single entity with the correlation function value between two entities. These eigencomponents often exhibit complex nonlinear correlations, with different eigencomponents contributing differently to the target task and showing varying degrees of correlation. Traditional linear models or simple kernel methods struggle to effectively capture the complex correlation structures in such high-dimensional feature spaces. Quantum state preparation and measurement require repeated execution on real quantum hardware, making the acquisition of a large number of labeled samples far more costly than classical data acquisition. Therefore, machine learning models suitable for small-sample scenarios are essential in this field. However, common methods such as deep neural networks are prone to overfitting under small-sample conditions, resulting in severely inadequate generalization performance. Quantum measurement itself exhibits statistical fluctuations and readout errors, leading to inherent uncertainties and noise in the extracted eigenvectors. Traditional machine learning models typically only output point estimates and cannot quantify the reliability of predictions, which is a significant limitation for tasks requiring confidence assessment, such as quantum state classification and phase transition boundary identification. In Gaussian process regression models, the choice of kernel function directly determines the model's ability to characterize prior assumptions such as function smoothness, periodicity, and correlation scale. Commonly used single kernel functions, such as the RBF kernel, Matern kernel, or linear kernel, are not limited in this embodiment. However, any single kernel function, such as the RBF kernel, Matern kernel, or linear kernel, can only capture specific types of data correlation structures and is difficult to simultaneously adapt to the complex multi-scale and multi-modal correlation characteristics that may exist in quantum enhanced feature vectors. As a Bayesian nonparametric model, Gaussian process regression can constrain the complexity of the model by performing posterior inference on the prior distribution defined by the kernel function when the training sample is limited, thereby avoiding overfitting under small sample conditions. This allows for the training of a predictive model with generalization ability with a small number of labeled samples on quantum hardware, reducing the data acquisition cost of quantum machine learning. Gaussian process regression provides uncertainty quantification capabilities. The predicted output includes not only mean estimation but also variance estimation, which can quantify the impact of quantum measurement noise and the uncertainty of the model itself on the prediction results. Composite kernel functions achieve joint capture of multi-scale features. Matern kernels can characterize locally smooth correlation structures in feature space. A smoothness level of ν=1.5 falls between that of RBF kernels and exponential kernels, retaining flexibility to adapt to the nonlinear characteristics of quantum data while providing certain smoothing constraints to avoid oscillations. Dot product kernels This allows the model to capture the linear trends and global correlation structures between feature components. When combined additively, the model adapts to data correlation patterns from different sources and scales in the quantum-enhanced feature vectors. The parameters of the kernel function all have explicit physical or statistical interpretability. Constant kernel coefficients. Control the overall signal amplitude and length scale Controlling the model's sensitivity to feature changes, dot product kernel size The strength of linear trends is controlled. These parameters are not only highly interpretable but also adaptively optimized using methods such as marginal likelihood maximization, eliminating the need for manual parameter tuning based on experience. They automatically adapt to optimal configurations across different quantum systems and target tasks. The choice of a Matern kernel ν=1.5 achieves a good balance between smoothness and computational efficiency. A Matern kernel with ν=1.5 requires only one differentiation to define. Compared to the higher-order smoothness requirements of ν=2.5 or other values, ν=1.5 is more computationally efficient and has better numerical stability, making it suitable for use with quantum-enhanced eigenvectors. It balances model expressiveness and computational efficiency when processing medium-scale quantum data.

[0072] Preferably, the structural descriptor of the molecule is obtained by at least one of the following methods: statistical calculation based on the molecule's SMILES string or chemical formula; or calculation of the molecular structure using a cheminformatics toolkit.

[0073] The first approach uses statistical calculations based on SMILES strings or chemical formulas. It rapidly extracts statistical features such as atomic composition, molecular weight, and number of functional groups by parsing and matching the one-dimensional sequence representation of molecules, without relying on external tool libraries. The second approach uses cheminformatics toolkits to calculate molecular structures. These toolkits analyze molecular graph structures or three-dimensional conformations, systematically calculating structured features such as topological descriptors, physicochemical property descriptors, and molecular fingerprints. The requirement of at least one of these approaches means that a single path or a combination can be flexibly chosen depending on the specific application scenario. This feature uses the molecular structured descriptor as the filtered feature data x, inputting the quantum many-body Hamiltonian. The method performs subsequent state evolution and feature extraction, providing a standardized classical-quantum interface. The direct SMILES-based statistical computation method requires no third-party libraries, is simple to deploy, and has a fast computation speed, making it suitable for rapid screening and lightweight applications. The method using cheminformatics toolkits, on the other hand, can obtain richer and more accurate structural features, suitable for professional applications with high accuracy requirements. These two optional methods make the method compatible with molecular data sources from different input formats and toolchains. Users can flexibly choose the optimal solution based on data availability and computational resources, demonstrating good practicality and scalability in fields such as drug molecule property prediction and material performance screening.

[0074] Preferred molecular-based structural descriptors specifically include: Mutual information features are used to filter the structural descriptors of molecules, and descriptors whose mutual information values ​​meet preset conditions are selected as input structural descriptors for normalization and mapping steps. The expression for calculating mutual information value is:

[0075] In the formula, X represents a single molecule descriptor feature, and Y represents the target value of the molecular electrochemical property to be predicted. Let X and Y be the joint probability density function. , Let X and Y be the marginal probability density functions, respectively.

[0076] The core of this technique lies in using mutual information to screen molecular structural descriptors, selecting a subset of descriptors with the highest correlation to the electrochemical properties of the target molecule for subsequent normalization and mapping steps. Mutual information is a measure in information theory of the degree of interdependence between two random variables, and its calculation expression is: In this expression, X represents a single molecule descriptor feature, and Y represents the target value of the molecular electrochemical property to be predicted. X is a single feature, and Y is a continuous target variable. A higher mutual information value indicates a greater contribution of the feature to the prediction of the target value. Target values ​​of molecular electrochemical properties include, for example, redox potential, band gap energy, and electron affinity. Let X and Y be the joint probability density function. and Let X and Y be the marginal probability density functions, respectively. The mutual information value reflects the contribution of the descriptor to the predicted target property: when the mutual information value is 0, it means that the descriptor is completely independent of the target property, i.e., it has no predictive value; the larger the mutual information value, the greater the contribution of the descriptor to the prediction of the target property. This method first calculates the mutual information value between each feature in all molecular descriptors and the target property Y, and then filters according to preset conditions, using only the filtered high-correlation descriptors as input to the subsequent normalization and quantum many-body Hamiltonian mapping steps. The essence of this filtering process is to construct a feature-target correlation ranking through information theory metrics, automatically identifying the most predictive key feature subset from the high-dimensional candidate feature set, and achieving effective dimensionality reduction of the input feature space without losing key information. The filtered descriptors are used as the filtered feature data x in the previous technical solution, and further input into the quantum many-body Hamiltonian. The process involves parameter mapping and quantum state evolution, transforming limited high-quality chemical features into input for a quantum machine learning model. Mutual information (MIS) offers an advantage over linear metrics like the Pearson correlation coefficient in capturing the nonlinear correlation between descriptors and target properties. Even with complex nonlinear mappings, MIS accurately assesses the correlation strength, preventing the omission of important features. By selecting high-contribution features with MIS values ​​that meet preset conditions and eliminating redundant descriptors irrelevant to the target property or with extremely weak correlations, the input dimensionality for subsequent normalization and quantum Hamiltonian mapping is effectively reduced, minimizing quantum resource consumption. After removing noisy features irrelevant to the target property, the machine learning model can focus on truly relevant chemical information, avoiding overfitting caused by redundant features and improving the model's prediction accuracy and generalization performance on unknown molecules. The calculation of MIS values ​​is entirely based on the distribution characteristics of the data itself, requiring no pre-set empirical rules. The selection results have clear information theory implications, facilitating researchers' understanding of which molecular structural features play a decisive role in the target electrochemical properties. In this embodiment, mutual information is used to filter features of the original molecular descriptors. The specific steps are as follows: First, input preparation: obtain the original feature matrix X and the target vector Y, where the dimension of X is n. samples ×22, representing n samples 22 molecular descriptor features from 22 samples; the dimension of Y is n. samples represents the target redox potential value for each sample. X includes X1...X... 22 , where X1...X 22 The value in the table represents the molecular descriptor feature with index j, where j is any value from 1 to 22. Second, mutual information is calculated. and They are X (X1...X) 22 The marginal probability density functions of Y and Y are used to determine the mutual information value. A higher mutual information value indicates a greater contribution of the feature to the prediction of the target value. Third, the mutual information score between each feature and the target vector Y is calculated, and the features are sorted in descending order of score. The top 10 features are selected as retained features. A fixed random seed of 42 is used to ensure the reproducibility of the selection results. Fourth, the selected feature matrix with dimension n is output. samples ×10, and the index of the selected feature in the original feature matrix. Fifth, after performing mutual information feature filtering (Top10), the dimension is n. samples ×10, and the index of the selected feature in the original feature matrix. In this embodiment, the selected feature index is: [3, 9, 2, 19, 5, 20, 1, 8, 12, 4], and the selected feature name is: ['C3', 'O1', 'C4', 'R5', 'N3', 'R6', 'S2', 'O2', 'P4', 'C2'].

[0077] Second Embodiment A second aspect of this application provides a molecular property prediction device based on enhanced quantum quenching dynamics features, comprising: The molecular feature encoding module is used to obtain the parameters of the quantum many-body Hamiltonian based on the molecular structure descriptor through normalization and mapping. The parameters are used to encode the structural features of the molecule. The quantum nonadiabatic evolution module is used to execute time-dependent evolution in a time-sequential manner through a quantum processor based on the parameters of the quantum many-body Hamiltonian, the preset initial driving Hamiltonian, and the scheduling function. Under nonadiabatic conditions, it drives the system from the initial product state to the entangled state, and the entangled state carries the structural information of the molecules. The quantum feature extraction module is used to extract the single-body expected value and the two-body correlation function value based on entangled state through local measurement, and then concatenate them to extract a numerical quantum enhanced feature vector. The machine learning prediction module is used to train and optimize the generalization performance of a classical machine learning model under a supervised learning framework based on quantum-enhanced feature vectors and corresponding molecular property labels, and outputs predicted values ​​of molecular electrochemical properties, including at least one of redox potential, inner-shell recombination energy, solvation energy and frontier orbital energy.

[0078] The molecular feature encoding module normalizes the structural descriptors of molecules and maps them to specific parameters of the quantum many-body Hamiltonian, thus encoding the molecular structural features into the Hamiltonian of the quantum system. The quantum non-adiabatic evolution module performs time-dependent evolution on the quantum processor: starting from an initial product state, the system smoothly transitions from the initial driving Hamiltonian to the problem Hamiltonian under the control of a scheduling function, preparing the target entangled state carrying molecular structural information under rapid non-adiabatic conditions. The quantum feature extraction module measures the entangled state using a local measurement scheme, adaptively selecting either the monomer Pauli expectation or the lattice occupancy number measurement based on the model type, and combining this with the correlation function between adjacent bodies to concatenate all measured values ​​into a numerically enhanced quantum feature vector. The machine learning prediction module uses the enhanced quantum feature vector as input and molecular property labels as supervision signals to train a Gaussian process regression model, ultimately outputting a predicted value for at least one electrochemical property among redox potential, inner-shell recombination energy, solvation energy, and frontier orbital energy. An end-to-end automated prediction pipeline is constructed. The entire process, from molecular structure input to property prediction output, requires no human intervention, reducing operational complexity and human error. A single entangled state synchronously encodes the complex coupling relationships between multiple molecular parameters through quantum correlation structures, providing rich information beyond classical descriptions for subsequent predictions. The linear measurement scheme of single-state plus two-state correlations avoids the exponential overhead of complete quantum state tomography, enabling efficient operation even with limited quantum resources. Rapid non-adiabatic evolution shortens the entanglement preparation time, effectively suppresses decoherence effects, and improves the fidelity of quantum feature extraction. The Gaussian process regression model adapts to small sample scenarios, effectively avoiding overfitting under limited labeled data, while outputting prediction confidence intervals to quantify prediction reliability. It also adapts to the complex multi-scale correlation structures in quantum-enhanced features through composite kernel functions. It has broad applicability in fields such as battery material screening, organic optoelectronic material design, and catalyst development, helping to narrow down the candidate molecule range and reduce wet experiment costs. The quantum processor handles entanglement preparation and feature extraction, while the classical processor handles preprocessing and model training, avoiding the limitations of a single computational paradigm. The interfaces of each module are highly standardized, which facilitates flexible migration and adaptation between different quantum hardware platforms and machine learning models.

[0079] In the description of this application, it should be noted that the terms "inner" and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product is in use. They are used only for the convenience of describing this application and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. Furthermore, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0080] It should also be noted that, unless otherwise explicitly specified and limited, the terms "setup" and "connection" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0081] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific identification content executed by the system and device described above can be referred to the corresponding process in the foregoing method embodiments.

[0082] The embodiments of this application have been described in detail above with reference to the accompanying drawings, but this application is not limited to the above embodiments. Even if various changes are made to this application, if these changes fall within the scope of the claims of this application and their equivalents, they shall still fall within the protection scope of this application.

Claims

1. A method for predicting molecular properties based on enhanced quantum quenching dynamics, characterized in that, include: Based on the molecular structure descriptor, the parameters of the quantum many-body Hamiltonian are obtained through normalization and mapping, and these parameters are used to encode the structural features of the molecule. Based on the parameters of the quantum many-body Hamiltonian, the preset initial driving Hamiltonian, and the scheduling function, the quantum processor executes time-dependent evolution in a time-sequential manner, driving the system from the initial direct product state to the entangled state under non-adiabatic conditions. The entangled state carries the structural information of the molecules. Based on the entangled state, the expected value of a single entity and the correlation function value of two entities are extracted by local measurement and then concatenated to extract a numerically enhanced quantum feature vector. Based on quantum-enhanced feature vectors and corresponding molecular property labels, a classical machine learning model is trained and optimized under a supervised learning framework to improve generalization performance, outputting predicted values ​​of molecular electrochemical properties, including at least one of redox potential, inner-shell recombination energy, solvation energy, and frontier orbital energy.

2. The molecular property prediction method based on enhanced quantum quenching dynamics characteristics according to claim 1, characterized in that, Also includes: Calculate at least one of the correlation coefficient, mean square error, root mean square error, and mean absolute error between the predicted value and the true value, and use it as a performance evaluation index of the classical machine learning model. The expression for calculating the correlation coefficient is as follows: In the formula, Let be the predicted value for the i-th sample. , These are the sample means of the predicted value and the actual value, respectively. To represent the true value of the i-th sample, n is the number of samples; The expression for calculating the mean squared error is: In the formula, Mean squared error; The formula for calculating the root mean square error is: In the formula, This is the root mean square error; The expression for calculating the mean absolute error is: 。 3. The molecular property prediction method based on enhanced quantum quenching dynamics characteristics according to claim 1, characterized in that, The parameters of the quantum many-body Hamiltonian can be obtained using any of the following models: the transverse field Ising model, the Heisenberg model, the Hubbard model, and the XY model. If the transverse field Ising model is used, the calculation expression for the model structure is: In the formula, , Here, represents the longitudinal magnetic field parameter acting on the i-th qubit, and the transverse coupling strength between the i-th and j-th qubits, respectively. For the Pauli-Z operator of the i-th qubit, , The Pauli-X operators for the i-th and j-th qubits are respectively; If the Heisenberg model is used, the calculation expression for the model structure is: In the formula, , , These represent the spin coupling strength in the first direction, the spin coupling strength in the second direction, and the spin coupling strength in the third direction, respectively. For the Pauli-Z operator of the i-th qubit, , The Pauli-Y operators for the i-th and j-th qubits are respectively; If the Hubbard model is used, the calculation expression for the model structure is: In the formula, For the electron jump integral, This represents the strength of the Coulomb interaction in its proper position. Chemical potential For an electron to jump from the j-th lattice point to the i-th lattice point, For Hermitian conjugate, , Let be the particle number operator that spins at the i-th lattice point in the first direction and the particle number operator that spins at the i-th lattice point in the opposite direction to the first direction, respectively. For the electron at the i-th lattice point with its spin pointing in the first direction Quantity; If the XY model is used, the calculation expression for the model structure is: In the formula, The spin coupling strength is in the first plane.

4. The molecular property prediction method based on enhanced quantum quenching dynamics characteristics according to claim 1, characterized in that, The steps for extracting the expected value of a single entity and the correlation function value between two entities further include: When the quantum many-body Hamiltonian is a spin system model, the single-bit Pauli operator expectation value is measured for each qubit of the entangled state, and the two-bit correlation function value is measured for adjacent qubits; When the quantum many-body Hamiltonian is a fermionic system model, the expected value of the single-lattice particle population is measured for each lattice point of the entangled state, and the value of the double-lattice correlation function is measured for adjacent lattice points. The measured single-body expected value is concatenated with the two-body correlation function value to form the quantum enhanced feature vector.

5. The molecular property prediction method based on enhanced quantum quenching dynamics characteristics according to claim 1, characterized in that, The steps to drive the system from the initial product state to the entangled state further include: The quantum system is prepared with the ground state of the initial driving Hamiltonian set as an entangled direct product state; The normalized evolution time changes continuously from 0 to 1. The scheduling function controls the weight of the initial driving Hamiltonian to decrease from 1 to 0 and the weight of the problem Hamiltonian to increase from 0 to 1, and constructs a time-dependent Hamiltonian. The calculation expression is as follows: In the formula, , Let be the scheduling functions for the initial driving Hamiltonian and the problem Hamiltonian, respectively, and s be the normalized evolution time. , These are the initial driving Hamiltonian and the problem Hamiltonian, respectively. Under rapid non-adiabatic conditions, the evolution from the initial direct product state to the entangled state is calculated using the following expression: In the formula, It is an unentangled direct product state.

6. The molecular property prediction method based on enhanced quantum quenching dynamics characteristics according to claim 5, characterized in that, The expression for calculating the parameters of the quantum many-body Hamiltonian is as follows: In the formula, For the filtered feature data, Let be the coupling strength between the i-th qubit and the j-th qubit. , Let X be the Pauli operator for the i-th qubit and Z be the Pauli operator for the i-th qubit.

7. The molecular property prediction method based on enhanced quantum quenching dynamics characteristics according to claim 1, characterized in that, The classic machine learning model is a Gaussian process regression model, and the kernel function of the classic machine learning model is calculated as follows: In the formula, For constant kernel coefficients, The length scale is based on the Matern kernel. The scale is the dot product kernel.

8. The molecular property prediction method based on enhanced quantum quenching dynamics characteristics according to claim 1, characterized in that, The structured descriptor of the molecule is obtained by at least one of the following methods: statistical calculation based on the molecule's SMILES string or chemical formula; or calculation of the molecular structure using a cheminformatics toolkit.

9. The molecular property prediction method based on enhanced quantum quenching dynamics characteristics according to claim 1, characterized in that, The molecular-based structured descriptor specifically includes: The structured descriptors of the molecules are screened for mutual information features, and descriptors whose mutual information values ​​meet preset conditions are selected as the structured descriptors input to the normalization and mapping steps; The expression for calculating the mutual information value is: In the formula, X represents a single molecule descriptor feature, and Y represents the target value of the molecular electrochemical property to be predicted. Let X and Y be the joint probability density function. , Let X and Y be the marginal probability density functions, respectively.

10. A molecular property prediction device based on enhanced quantum quenching dynamics, characterized in that, include: The molecular feature encoding module is used to obtain the parameters of the quantum many-body Hamiltonian based on the molecular structure descriptor through normalization and mapping. These parameters are used to encode the structural features of the molecule. The quantum nonadiabatic evolution module is used to perform time-dependent evolution in a time-sequential manner through a quantum processor based on the parameters of the quantum many-body Hamiltonian, a preset initial driving Hamiltonian, and a scheduling function, thereby driving the system from an initial direct product state to an entangled state under nonadiabatic conditions. The entangled state carries the structural information of the molecules. The quantum feature extraction module is used to extract the single-unit expected value and the two-unit correlation function value based on the entangled state through local measurement, and then concatenate them to extract a numerical quantum enhanced feature vector; The machine learning prediction module is used to train and optimize the generalization performance of a classical machine learning model under a supervised learning framework based on quantum-enhanced feature vectors and corresponding molecular property labels, and output predicted values ​​of molecular electrochemical properties, including at least one of redox potential, inner-shell recombination energy, solvation energy and frontier orbital energy.

Images