CN-121983203-A - Quantum transport simulation-oriented defect device tight binding modeling method

CN121983203ACN 121983203 ACN121983203 ACN 121983203ACN-121983203-A

Abstract

The application discloses a method for constructing a channel binding model of a semiconductor device with defects and non-period. The method takes a supercell atomic structure as a characteristic and a density functional theory calculation energy band as a label to construct a physical perception graph neural network integrating a symmetric graph attention mechanism. During feature learning and band fitting, the model imposes an exponentially decaying physical constraint on the transition parameters. Through co-training of defect-free supercells and defect-containing supercells, the network can automatically analyze local perturbation caused by defects and dynamically adjust local parameters related to defect positions in a defect-free model so as to accurately fit the energy band structure of a defect system. The mechanism strictly guarantees the consistency of supercell transition matrixes of systems containing defects and no defects while constructing a model, so that a complete set of model parameters are output. The application effectively solves the technical bottleneck that the coupling matrix between complex systems is difficult to obtain in the modeling of the semiconductor material.

Inventors

LV YAWEI
TANG WEIMIN
JIANG CHANGZHONG

Assignees

湖南大学

Dates

Publication Date: 20260505
Application Date: 20260402

Claims (10)

1. A method for constructing a tight binding model of a semiconductor device with defects and non-periods is characterized by comprising the following steps: Step S1, data acquisition and graph structure construction are carried out, the atomic structures of flawless supercells and flawless supercells of materials to be simulated are obtained, DFT is utilized to calculate respective energy band structures to be used as training labels, the atomic structures of the supercells are converted into graph structures, atoms are nodes, chemical bonds are edges, and node features, edge features and global features are extracted; Step S2, constructing a physical perception diagram neural network, constructing a GNN framework comprising a feature encoder, a symmetrical diagram attention layer and a parameter predictor, updating node states through symmetrical message transmission by the network, outputting parameters of a TB model, wherein the parameters comprise track potential energy and transition parameters, and applying physical constraint of exponentially decaying along with distance to the output transition parameters; S3, defect perception co-training, namely inputting data of defect-free supercells and defect-containing supercells into the GNN framework at the same time, performing co-training by adopting a shared weight strategy, calculating a weighted error between a predicted energy band and a DFT calculated energy band, and enabling a network to establish a mapping relation between local atomic environment change caused by defects and TB model parameter change; S4, assembling a device channel Hamiltonian matrix, constructing a device channel structure consisting of defect-free supercell chains, embedding defect-containing supercells, predicting Hamiltonian matrixes inside the supercells and between nearest neighbor supercells based on a trained model, and assembling the transition matrixes among the defect-free supercells as transition matrixes among the defect-free supercells and the defect-containing supercells to form a complete defect-containing and non-periodic semiconductor device channel Hamiltonian matrix for quantum transport simulation.
2. The method according to claim 1, wherein in step S1, the connection determination condition of the atomic graph edges of the graph structure is: Interatomic distance d ij ＜ R i + R j +Δr Wherein R i , R j is the covalent radius of the atom, and Deltar is the delocalization offset; For the positive ions in the oxide semiconductor material, the delocalization offset deltar is set to be a positive value so as to cover the delocalization effect of the electron orbitals of the outer layer of the metal atoms, and ensure that the graph structure contains long-range electron interactions.
3. The method according to claim 1, wherein in step S2, the symmetrical graph attention layer employs a symmetrical messaging mechanism, the calculation of the attention coefficient α ij of which is based on the mean operation of node features: wherein x i , x j is the feature vector of the node i, j, e ij is the edge feature vector, alpha T is the weight vector, |indicates the splicing operation, leakyReLU is the nonlinear activation function, and Softmax is the normalization function; The attention coefficient is forced to meet alpha ij = α ji through the averaging operation, so that the predicted interatomic interaction is ensured to accord with physical symmetry.
4. The method according to claim 1, characterized in that in step S2, the physical constraint is that the following forced correction is performed on the transition parameter t ij ' of the original output of the network: wherein r ij is interatomic distance, and r cut is a truncated radius parameter; the physical constraints force the transition strength to decay exponentially with increasing interatomic distance to ensure sparsity and physical interpretability of the generated TB model.
5. The method according to claim 1, wherein in step S3, the defect-aware co-training implements parameter prediction by sharing network weights, using local environment alike variability of GNN architecture: For the region far from the defect center in the supercell containing the defect, predicting Hamiltonian quantity parameters consistent with the numerical value of the supercell without the defect by the network based on a local atomic coordination environment consistent with the supercell without the defect; for the area near the defect center, the network automatically adjusts Hamiltonian parameters to fit the defect state by sensing the geometric distortion of bond length, bond angle and coordination number.
6. The method according to claim 1, characterized in that in step S3, the loss function employed by the co-training comprises fermi-dirac-like weighting factors: wherein E is an energy band eigenvalue, and alpha is an attenuation coefficient; The weighting factors endow the energy band eigenvalues near the conduction band bottom and the valence band top with higher weights, and the weights decay exponentially along with the energy away from the band edge, so that the simulation precision in the key energy window of the on-off characteristic of the device is improved.
7. The method according to claim 1, wherein in step S3, the training process uses a fully micromanipulation energy band eigenvalue solver that constructs a hamiltonian matrix using vectorized matrix operations and incorporates hermite forced operations to support end-to-end gradient back propagation of the energy band eigenvalue solver process.
8. The method according to claim 1, characterized in that in step S4, the approximate characterization of the interface coupling matrix is based on the following strategy: Based on the physical fact that the atomic environment containing the edge of the defect supercell is converged to a non-defective state, a transition matrix H NN DF-DF between the non-defective supercells predicted by the model is directly selected as a coupling transition matrix H NN DF-DC at the joint of the non-defective supercells and the defective supercells, so that the problem that interface coupling parameters cannot be directly calculated under non-periodic boundary conditions is solved.
9. The method according to claim 1, wherein in step S4, the full-channel hamiltonian assembly is performed by assembling a predicted defect-free supercell intracellular matrix, a defect-containing supercell intracellular matrix, and an interface coupling matrix into a block tri-diagonal matrix according to a geometric design of a device channel, and inputting the matrix into an unbalanced green function solver for current-voltage characteristics and electron transmission spectrum simulation.
10. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor is operable to implement the method of any one of claims 1-9 when the program is executed.

Description

Quantum transport simulation-oriented defect device tight binding modeling method Technical Field The application belongs to the field of physics and quantum transport simulation of semiconductor devices, and particularly relates to an automatic construction of a tightly bound (Tight-Binding, TB) model, in particular to a construction technology applied to a channel quantum transport tightly bound model of a semiconductor device with defects and non-periodicity. Background With the continuous reduction of the channel size of semiconductor devices, the effect of quantum effects in carrier transport is increasingly significant. However, the actual device channel has defects, amorphous structures or non-ideal factors such as interface disorder, so that the conventional drift-diffusion theory faces a great challenge for describing the carrier behavior. The existing TB model construction method has the technical defects that on one hand, the empirical TB model parameters are extracted from ideal and periodical lattices, the systematic portability is lacked, the response of an electronic structure to local structure disturbance (such as defects) cannot be effectively captured, and the complex or non-ideal system is difficult to adapt, and on the other hand, the TB model construction method combining the first principle density functional theory (Density Functional Theory, DFT) and the maximum localized ten-denier function (Maximally Localized Wannier Functions, MLWFs) is capable of obtaining the TB model, the basis of the TB model is still dependent on the assumption of periodical primitive cells, and the applicability of the TB model construction method to a real device channel containing defects or amorphous regions is fundamentally limited. It is particularly critical that the lack of translational symmetry of the non-periodic system results in a failure of the parameters to shift naturally between different regions, making the electronic coupling parameters between defective and non-defective supercells difficult to determine, which is a long standing challenge in modeling from computational quantum transport. Disclosure of Invention The embodiment of the application aims to provide a method and a device for constructing a tightly-bound model of channel quantum transportation of a semiconductor device with defects and non-periodicity, a storage medium and electronic equipment. Through the TB model automatic construction framework based on the physical perception map neural network (Graph Neural Network, GNN) provided by the embodiment of the application, local structure disturbance can be automatically encoded into parameter change by using a Defect perception Co-training (Defect-aware Co-training) mechanism, so that a coupling transition matrix between Defect-containing supercells and Defect-free supercells is accurately constructed, the problem of constructing a channel Hamiltonian matrix of a semiconductor device containing defects and non-periods in the head calculation quantum transport simulation is solved, and the physical precision and the calculation efficiency of the simulation are improved. According to a first aspect of the present application, an embodiment of the present application provides a method for constructing a tightly-bound model of channel quantum transport of a semiconductor device including defects and non-periods, the method including the steps of: (1) The method comprises the steps of data acquisition and graph structure construction, namely acquiring the atomic structures of Defect-Free (DF) and Defect-containing (Defect-Containing, DC) of a material to be simulated, calculating the corresponding DFT energy band structures as training labels, converting the atomic structures of the supercells into graph structures, wherein atoms are nodes, chemical bonds are edges, and extracting node characteristics, edge characteristics and global characteristics; (2) The construction method comprises the steps of constructing a GNN framework comprising a feature encoder, a symmetrical graph attention layer and a parameter predictor, updating node states through symmetrical message transmission, and outputting TB Hamiltonian quantity parameters, wherein the parameters comprise track potential energy and transition parameters; (3) In the defect perception co-training stage, data of the defect-free supercells and the defect-containing supercells are simultaneously input into the GNN framework, and a shared weight strategy is adopted for co-training; (4) And in the full-channel Hamiltonian amount assembling and simulating stage, the intracellular matrixes of all areas of the channel of the device are respectively predicted based on a trained model, and the coupling transition matrixes of the interfaces of the defect-free supercells and the defect-containing supercells are approximately represented by utilizing the interlayer transition matrixes of the defect-free supercells, so that the complete Ha