Defective device tight constraint modeling method for quantum transport simulation

Abstract

The invention discloses a construction method of a channel tight constraint model of an aperiodic semiconductor device with defects. According to the method, a supercell atomic structure is used as a feature, a density functional theory calculation energy band is used as a label, and a physical perception graph neural network integrated with a symmetric graph attention mechanism is constructed. In the feature learning and energy band fitting process, the model applies physical constraint of exponential decay to transition parameters. Through co-training defect-free 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 precisely fit an energy band structure of a defect system. According to the mechanism, the consistency of the super-intercellular transition matrixes of defect-containing and defect-free systems is strictly guaranteed while the model is constructed, so that a whole set of model parameters are output. According to the method, the technical bottleneck that a coupling matrix between complex systems is difficult to obtain in semiconductor material modeling is effectively solved.

Description

Technical Field

[0001] This application belongs to the field of semiconductor device physics and quantum transport simulation, specifically involving the automated construction of tight-binding (TB) models, and in particular a technique for constructing tight-binding models of quantum transport in defective, non-periodic semiconductor device channels. Background Technology

[0002] As the channel size of semiconductor devices continues to shrink, the influence of quantum effects on carrier transport becomes increasingly significant. However, practical device channels are generally subject to non-ideal factors such as defects, amorphous structures, or interface disorder, posing a significant challenge to the traditional drift-diffusion theory in describing carrier behavior.

[0003] Existing methods for constructing quantum transport (TB) models suffer from the following technical limitations: Firstly, empirical TB model parameters are extracted from ideal, periodic lattices, lacking systematic portability and failing to effectively capture the electronic structure's response to local structural perturbations (such as defects), making them unsuitable for complex or non-ideal systems. Secondly, while methods combining first-principles density functional theory (DFT) with maximum localized Wannier functions (MLWFs) can yield TB models, their foundation still relies on the assumption of periodic units, fundamentally limiting their applicability to real device channels containing defects or amorphous regions. Crucially, aperiodic systems lack translational symmetry, preventing parameters from naturally transferring between different regions, making it difficult to determine the electronic coupling parameters between defective and defect-free supercells—a long-standing challenge in ab initio quantum transport modeling. Summary of the Invention

[0004] The purpose of this application is to provide a method, apparatus, storage medium, and electronic device for constructing a tight-binding model of quantum transport in a defective, aperiodic semiconductor device channel. Through the automated TB model construction architecture based on a Graph Neural Network (GNN) proposed in this application, and utilizing a defect-aware co-training mechanism, local structural perturbations can be automatically encoded as parameter changes, thereby accurately constructing the coupling transition matrix between defective and defect-free supercells. This solves the problem of constructing the Hamiltonian matrix of a defective, aperiodic semiconductor device channel in ab initio quantum transport simulations, improving the physical accuracy and computational efficiency of the simulation.

[0005] According to a first aspect of this application, embodiments of this application provide a method for constructing a tight-binding model of quantum transport in a defective, aperiodic semiconductor device channel, the method comprising the following steps:

[0006] (1) In the data acquisition and graph structure construction stage, the atomic structures of the defect-free (DF) and defect-containing (DC) supercells of the material to be simulated are acquired, and the corresponding DFT band structures are calculated as training labels. The atomic structure of the supercell is then converted into a graph structure, where atoms are nodes and chemical bonds are edges. Node features, edge features, and global features are extracted.

[0007] (2) In the physical perception graph neural network construction stage, a GNN architecture including a feature encoder, a symmetric graph attention layer and a parameter predictor is built; the architecture updates the node state through symmetric message passing and outputs TB Hamiltonian parameters, which include orbital in-potential energy and transition parameters.

[0008] (3) In the defect-aware co-training stage, the data of defect-free supercells and defective supercells are simultaneously input into the GNN architecture, and co-training is performed using a shared weight strategy; the weighted error between the predicted band and the DFT calculated band is calculated, so that the network establishes a mapping relationship between the local atomic environment and the Hamiltonian parameters.

[0009] (4) In the full-channel Hamiltonian assembly and simulation stage, the intracellular matrix of each region of the device channel is predicted based on the trained model, and the interlayer transition matrix of the defect-free supercell is used to approximate the coupling transition matrix of the interface between the defect-free supercell and the defective supercell. The complete Hamiltonian matrix of the defective, non-periodic semiconductor device channel is assembled for quantum transport simulation.

[0010] In some embodiments, in step (1), the connection condition for graph edges is defined as the interatomic distance d during graph structure construction. ij < R i + R j + Δr; where R i , R j denoted as the covalent radius of the atom, and Δr as the delocalization offset.

[0011] For cations in oxide semiconductor materials, the delocalization offset Δr is set to a positive value to cover the delocalization effect of the outer electron orbitals of metal atoms, ensuring that the graph structure contains long-range electronic interactions.

[0012] In some embodiments, in step (2), the symmetric graph attention layer employs a symmetric message passing mechanism, and its attention coefficient α ij The calculation uses a mean operation based on node features:

[0013]

[0014] In the formula, x i , x j These are node features, e ij The edge features are defined by ||, which represents the concatenation operation. LeakyReLU is a non-linear activation function, and Softmax is a normalization function used to ensure that the sum of the attention coefficients of all neighbors of a node is 1. The mean operation is used to force α to be equal to 1. ij = α ji This ensures that the predicted interatomic interactions conform to physical symmetry.

[0015] In some embodiments, in step (2), the GNN architecture incorporates physical constraints, which refer to the transition parameters t of the original network output. ij Perform the following forced correction:

[0016]

[0017] In the formula, t ij ' For the original output, r ij r is the interatomic distance. cut The cutoff radius parameter. This constraint forces the transition intensity to decay exponentially with increasing interatomic distance, ensuring that the generated TB model is sparsity and physically interpretable.

[0018] In some embodiments, step (3) employs a defect-aware co-training strategy, simultaneously inputting data from defect-free and defective supercells into the GNN architecture, and forcing the network to establish a mapping relationship between changes in the local atomic environment and changes in Hamiltonian parameters by sharing network weights; the GNN utilizes the isovariability of the local environment to make the parameters of the region far from the defect center automatically converge to the defect-free state, while for the region near the defect center, the parameters are automatically adjusted according to the geometric distortions of bond length, bond angle, and coordination number.

[0019] During this co-training process, the error between the predicted band and the DFT band is calculated using a weighted loss function, which includes a Fermi-Dirac weighting factor:

[0020]

[0021] In the formula, E is the band eigenvalue and α is the attenuation coefficient.

[0022] The weighting factor assigns higher weights to the band eigenvalues ​​near the conduction band bottom and valence band top, and the weights decrease exponentially as the energy moves away from the band edge, thereby improving the simulation accuracy within the critical energy window of the device's turn-on and turn-off characteristics.

[0023] In some embodiments, step (4), the approximate representation of the interface coupling matrix is ​​based on the following strategy:

[0024] Based on the physical fact that the atomic environment at the edge of a defective supercell has converged to a defect-free state, the transition matrix H between defect-free supercells predicted by the model is directly selected. NN DF-DF H is the interface coupling matrix at the junction of defect-free and defective supercells. NN DF-DC This addresses the problem of not being able to directly calculate interface coupling parameters under aperiodic boundary conditions.

[0025] According to a second aspect of this application, some embodiments of this application utilize a method for constructing a defective, aperiodic semiconductor device channel quantum transport tight-binding model, specifically for modeling amorphous indium gallium zinc oxide (a-IGZO) thin-film transistors, including:

[0026] Obtain defect-free supercells and oxygen-containing vacancies (V0) of amorphous α-IGZO generated through a molecular dynamics (MD) melt-quench process. O The defective supercell structure of the oxide semiconductor was constructed, and its electronic band structure was calculated using the DFT+U method as training data. To address the characteristics of oxide semiconductors, a delocalization offset (Δr) was introduced into the cations during graph feature construction, and a cutoff radius (rc) was applied to the output transition parameters. cut The exponential decay constraint is applied to output TB parameters that are highly sparsity and physically definite.

[0027] Through co-training, defect states deep in the bandgap and conduction band bottom tailing effects are captured, and local structural perturbations (such as lattice distortions caused by metal-metal bonding) are automatically encoded as changes in Hamiltonian matrix elements. By approximating interface coupling with the interlayer transition matrix of a defect-free supercell, Hamiltonian matrices of defective, non-periodic semiconductor device channels with a specific length are assembled. Combined with non-equilibrium Green's function (NEGF) to simulate device quantum transport, current-voltage (IV) characteristics are obtained to characterize defect-induced turn-on current decay and negative differential resistance (NDR) effects.

[0028] According to a third aspect of this application, some embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can implement the methods described in any embodiment of the first aspect.

[0029] According to a fourth aspect of this application, some embodiments of this application provide an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, can implement the method as described in any embodiment of the first aspect.

[0030] Compared with the prior art, this application has the following beneficial effects:

[0031] This application leverages the perceptual characteristics of GNN architecture regarding local atomic environments and employs a co-training strategy to enable the model to automatically match transition parameters based on the similarity of local structures. This breakthrough solves the problem that traditional methods cannot determine the interface coupling matrix (H) between defect-free and defective supercells. NN DF-DC This addresses the problem of [missing information - likely related to quantum transport], thus enabling the accurate construction of a tight-binding model for channel quantum transport in defective, aperiodic semiconductor devices. Furthermore, the construction of the Hamiltonian matrix strictly adheres to atomic bonding relationships, exhibiting high sparsity and distance-dependent exponential decay characteristics, with clear physical meaning. By introducing a Fermi-Dirac-like weighted loss function, the model's prediction accuracy in key energy windows such as the conduction band bottom and valence band top is significantly improved, allowing for more accurate simulation of key electrical characteristics of semiconductor devices, such as turn-on voltage and on-state current. Attached Figure Description

[0032] Figure 1 This illustration shows an overall flowchart of a method for constructing a tight-binding model of quantum transport in a defective, non-periodic semiconductor device channel, as provided in an embodiment of this application.

[0033] Figure 2 A detailed schematic diagram of the physical perception graph neural network provided in an embodiment of this application is shown.

[0034] Figure 3 A schematic diagram of the defect-aware co-training strategy provided in an embodiment of this application is shown.

[0035] Figure 4 This illustration shows a schematic diagram of the assembly and application process of a defective, non-periodic semiconductor device channel Hamiltonian provided in an embodiment of this application.

[0036] Figure 5 The diagram shows a comparison of the physical interpretability verification of the tightness parameters provided in the embodiments of this application.

[0037] Figure 6 The diagram shows the verification results of the electronic band structure training for the defective indium gallium zinc oxide system provided in the embodiments of this application.

[0038] Figure 7 The diagram illustrates the effectiveness verification of the interface coupling matrix approximation strategy provided in the embodiments of this application.

[0039] Figure 8 An analytical diagram illustrating the influence of metal-metal bond-induced defects on the band structure provided in an embodiment of this application is shown.

[0040] Figure 9 The figure shows the quantum transport simulation results based on the amorphous indium gallium zinc oxide defect device provided in the embodiments of this application. Detailed Implementation

[0041] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0042] Implementation Example 1

[0043] This embodiment provides a method for constructing a tight-binding model of channel quantum transport in defective, aperiodic semiconductor devices, aiming to solve the problem that existing technologies cannot accurately construct the electronic coupling transition matrix between defective and defect-free supercells. Figure 1 As shown, the method specifically includes the following steps:

[0044] Step S1: Obtain DFT band structure data of defect-free and defective supercells, introduce delocalization offsets to construct graph structure, and extract nodes, edges and global features as input;

[0045] Step S2: Construct a physical perception graph neural network and use a symmetric graph attention mechanism based on the mean of node features to perform feature extraction and symmetric message passing;

[0046] Step S3: Train dual-path data using a defect-aware co-training strategy with shared weights, and establish the mapping relationship between the local environment and parameters using the Fermi-Dirac weighted loss function;

[0047] Step S4: Output tight-binding parameters to apply physical constraints to the transition parameters that decay exponentially with distance, ensuring that the model has sparsity and physical interpretability;

[0048] Step S5: Using the interlayer matrix of the defect-free supercell to approximate the interface coupling, a full-channel Hamiltonian matrix is ​​assembled and combined with NEGF to perform quantum transport simulation.

[0049] In step S1, the implementation process includes:

[0050] For the semiconductor material system to be simulated, two types of supercell structures are constructed: defect-free supercells (DF) and defective supercells (DC). The electronic structure of the supercells is calculated using DFT software (Quantum ESPRESSO), and the converged band structure eigenvalues ​​are used as training labels. For amorphous materials, the amorphous structure can be generated by simulating the "melting-quenching" process through molecular dynamics. For crystalline materials, the defective structure can be generated by expanding the cell and introducing point defects (such as interstitial atoms or vacancies).

[0051] The atomic structure is transformed into a graph structure G = (V, E). Here, nodes V represent atoms, and edges E represent chemical bonds; the connection criterion for graph edges is defined as the interatomic distance d. ij < R i + R j + Δr. Where R i , R j The covalent radius of the atom is given. Specifically, for cations in oxide semiconductors, a delocalization offset Δr is introduced to cover the delocalization effect of the outer electron orbitals of metal atoms, ensuring that the graph structure correctly reflects long-range electron interactions. The physical properties of each atom are extracted as node features, including electronegativity, number of valence electrons, covalent radius, ionic charge, number of bonds, and atom type. Bond vectors and bond types are extracted as edge features. The supercell volume is extracted as a global feature.

[0052] Based on the above description, two sets of input feature data—one for defect-free supercells (DF) and one for defective supercells (DC)—are finally constructed. Based on this, as follows: Figure 2 As shown, in step S2, a multi-layer perceptron (MLP) is used to map the input atomic features, bond features, and volume features into high-dimensional embedding vectors, respectively. The process of constructing the physical perception GNN architecture includes:

[0053] GNN contains multiple layers of symmetric graph attention convolutional layers. In each layer, the current node state is updated by aggregating neighbor node information. To ensure physical symmetry for any connected node pair (i, j), the symmetric attention coefficient α is calculated based on the concatenation of the node feature mean and edge features. ij ;

[0054] Based on the symmetric attention coefficient and the mean of node features, a symmetric message m is constructed from node j to node i. ij The messages from neighboring nodes are aggregated on average, and the feature state of the current node is updated through a feedforward neural network.

[0055] In step S3, the following is adopted: Figure 3The core of the defect-aware co-training strategy shown in this step lies in establishing a unified mapping mechanism that can automatically encode local structural perturbations into changes in electronic parameters through a "weight sharing" strategy. The specific implementation process is as follows:

[0056] The system will use defect-free supercell data (including atomic feature X) DF Key features E DF Volume V DF ) and defective supercell data (X DC E DC V DC The two parallel inputs are fed into the same Physical Aware GNN. When processing these two very different systems, the GNN forces the sharing of the same set of network weight parameters. This forces the network to learn a set of universal physical rules that do not depend on any specific system, namely, to establish a direct mapping function from the "local atomic environment" to the "TB parameters": f(LocalEnvironment) → TB Parameters.

[0057] Figure 3 The "key co-training mechanism" on the right utilizes the inherent local environmental equivariance of GNNs to automatically adjust prediction behavior based on the similarity of the local atomic environments: In the remote region, the local coordination environment of atoms is highly similar to that of the defect-free supercell. Therefore, the network with shared weights automatically predicts Hamiltonian parameters that are almost identical to those of the defect-free supercell, achieving natural convergence of parameters in the bulk material region without manual intervention. In the defect neighborhood, near the defect center, the network keenly perceives geometric distortions in bond lengths, bond angles, and coordination numbers, and automatically adjusts the Hamiltonian parameters in this region through backpropagation to accurately fit deep-level defect states or conduction band bottom reconstructions appearing in DFT calculations.

[0058] In step S4, as Figure 3 As shown, the TB Hamiltonian parameters output by the GNN after training are: the potential energy of the atomic orbitals and the transition parameters t between orbitals. ij To ensure the physical interpretability and sparsity of the generated TB model, hard physical constraints are imposed on the predicted transition parameters:

[0059]

[0060] In the formula, r ij r is the interatomic distance. cut t is the cutoff radius parameter. ij ' This is the original output from the network.

[0061] This constraint forces the transition strength to decay exponentially with increasing distance. Furthermore, the parameters are fed into a differentiable band eigenvalue solver, which incorporates Hermitian enforcement operations and uses vectorized matrix operations to quickly construct the Hamiltonian and solve for the band eigenvalues, supporting end-to-end gradient backpropagation from band error to network weights.

[0062] To address the issue that traditional mean square error methods do not adequately address deep energy levels. Figure 3 The loss function is defined at the bottom as the predicted band. pred DF ) and DFT band DFT DF The weighted loss between ( ). Introduce a Fermi-Dirac weighting factor:

[0063]

[0064] In the formula, E is the band eigenvalue and α is the attenuation coefficient;

[0065] This weighting function assigns extremely high importance to energy bands near the band edge, with the weights decreasing exponentially as the energy moves away from the band edge. This ensures that the generated TB model has extremely high accuracy within the critical energy window that determines the device's turn-on / turn-off characteristics and the effective mass of charge carriers.

[0066] In step S5, the trained model is used to solve the problem that traditional methods cannot define the non-periodic interface coupling transition matrix, such as... Figure 4 As shown, a device TB Hamiltonian for quantum transport simulation is constructed, including:

[0067] The device channel is divided into a series of supercells, and the intracellular matrix H of the defect-free supercell is predicted using a trained model. DF and the intracellular matrix H of the defective supercell DC Based on the "defect-aware" characteristic verified in step S3: because the selected defective supercell has a sufficiently large size, the lattice distortion and electronic structure perturbation caused by defects are confined to the supercell center and do not propagate to the supercell boundary. Therefore, the atomic environment at the edge of the defective supercell has converged to the defect-free state. Based on this physical fact, the interlayer transition matrix H at the junction of the defect-free supercell and the defective supercell is... NN DF-DC Numerical transition matrix H between the defect-free supercell and the defect-free supercell NN DF-DF The heights are similar, so H can be used directly. NN DF-DF As the interface coupling transition matrix.

[0068] This strategy cleverly avoids the limitation that H cannot be directly obtained through Wannier interpolation under aperiodic boundary conditions. NNDF-DC Overcoming the theoretical dilemma, seamless splicing of defect-free and defective supercells was achieved.

[0069] Based on the device's geometric design (source-defect-free channel-defect-defect-free channel-drain, i.e., ... -DF-DC-DF-...), the predicted sub-matrices are constructed into a huge block tri-diagonal Hamiltonian matrix H. Device The matrix is ​​input into the NEGF solver, and combined with the self-consistent solution of the Poisson equation, current-voltage (IV) characteristics and electron transmission spectrum simulations are performed, thereby revealing the specific impact of defects on the macroscopic electrical performance of the device at the atomic scale.

[0070] Implementation Example 2

[0071] This embodiment aims to illustrate and verify in detail the effectiveness and universality of the method for constructing a tight-binding quantum transport model for defective, non-periodic semiconductor device channels described in this application, using amorphous indium gallium zinc oxide (a-IGZO) thin-film transistor device channel modeling as an example. Specifically, it includes the following steps:

[0072] Step S1, Data Preparation and Parameter Setting:

[0073] 84-atom α-IGZO was selected as the research object. Molecular dynamics (MD) simulations of the "crystal melting-quenching" process were used, and after structural relaxation optimization, an amorphous defect-free supercell was generated. Based on this, oxygen atoms were randomly removed to form V1... O This leads to the relaxation of adjacent metal atoms, forming metal-metal (MM) bonds, thus constructing a defective supercell. The electronic band structure of this supercell is calculated using the DFT+U method as a training label. During graph structure construction, for In, Ga, and Zn cations, a delocalization offset Δr of 0.5 Å is introduced into the bonding radius to accurately capture the delocalization effect of the outer electron orbitals of metal atoms, exponentially decaying the cutoff radius r in the physical constraints. cut Set it to 3 angstroms.

[0074] Step S2, Model Training and Physical Interpretability Validation:

[0075] The GNN's defect-aware co-training strategy was used to train the supercell data. The trained model exhibited high physical interpretability and accurate capture of local defect environments, such as... Figure 5 As shown, the model successfully learned the physical nature of the TB Hamiltonian. Figure 5 a and 5b respectively show the a-IGZO supercell TB Hamiltonian matrix obtained through the physical perception GNN architecture and MLWFs transformation, and the matrix obtained through MLWFs ( Figure 5 (b) In comparison, the Hamiltonian matrix predicted by the physical perception GNN architecture exhibits significantly enhanced sparsity, indicating that this method improves the locality of the TB model, and the degree of localization can be further tuned by adjusting the bonding radius of atoms.

[0076] Furthermore, the physical constraint that the transition integral value decays exponentially with increasing orbital distance, in Figure 5 This is clearly demonstrated in c. In comparison, the result obtained through MLWFs transformation ( Figure 5 In d), the exponential decay behavior is significantly weaker. This proves that the hard physical constraint t introduced in this application... ij ∝ exp(-r ij / r cut This forces neural networks to strictly follow the distance decay law of quantum mechanics, rather than simply fitting data.

[0077] Figure 6 The image shows a comparison of the band structure under defect-free and defect-adjusted configurations. The predicted results highly coincide with the DFT calculation results near the Fermi level at the conduction band bottom and valence band top, accurately reproducing the band gap characteristics and dispersion relation of α-IGZO. A comparison of the lower-side density of states (DOS) is also included. Figure 6 f-6j shows that for systems containing MM bond defects, the model accurately predicts the peak values ​​of defect states located deep in the band gap and the tailing effect at the bottom of the conduction band. Local Density of States (LDOS) demonstrates that the model can autonomously correlate local structural deformations with defect levels in the band structure; LDOS increases with proximity to the defect site, such as... Figure 6 As shown in k-6o. Taking an α-IGZO system as an example, the differences in intracellular transition matrices between defect-free and defective supercells are further examined. Figure 6 The scatter plot of p was crucial in verifying the model’s “defect-aware” characteristic. The results showed that the predicted changes in the diagonal elements of the Hamiltonian were highly correlated with the distance from the defect center. At the far end of the defect, the parameter differences quickly returned to zero, and the differences were found to be concentrated around the cation sites adjacent to the oxygen vacancy.

[0078] This indicates that through co-training, the model automatically learns to treat "defects" as a local perturbation rather than a global change, thus providing a physical basis for seamlessly splicing Hamiltonian matrices of defect-free and defective supercells.

[0079] Step S3: Construction and transport simulation of the channel Hamiltonian matrix of a defective, aperiodic semiconductor device:

[0080] The predicted TB model based on the verification in step S2 exhibits enhanced sparsity and physically-aware locality, and this application solves the problem of aperiodic interface coupling. Figure 7 As shown, the defect channel Hamiltonian in the device is constructed using the interlayer transition matrix H of the defect-free supercell. NN DF-DF As an approximation, tests were conducted on two sets of 84-atom a-IGZO supercells, one defective and one defect-free, and the obtained |H NN DC-DC -H NN DF-DF The distribution of | is as follows Figure 7 As shown in b and 7c. In this embodiment, only sparse off-diagonal elements are activated, and most of them are below 0.1 eV. Furthermore, the band structure was reconstructed using a set of TB models of the system, and the H... NN DC-DC and H NN DF-DF Interchange tests were conducted, and the results were as follows: Figure 7 As shown in d and 7e. Regardless of whether the supercell contains defects, H NN The exchange of electrons hardly causes any change in their electronic structure. In contrast, the results obtained through MLWF transformation are as follows: Figure 7 As shown in f and 7g, the MLWFs transformation does not possess a unified H between defective and defect-free supercells. NN This result validates the ability to construct a unified H across defective / defect-free systems through a co-training strategy. NN DF-DC The feasibility of this method is demonstrated, and the challenges faced in modeling defective channels in the aforementioned background technologies are cleverly solved, thereby enabling high-precision quantum transport simulation of devices with defective systems.

[0081] Furthermore, an a-IGZO device channel was constructed, which consisted of 20 supercells connected in series, and a V0 supercell was introduced at the center of the channel. O The defective supercell of induced MM bond formation, corresponding to V O The concentration was 4.59 × 10⁻⁶. 19 cm -3 (High concentration side within the experimental observation range). A TB model constructed using a physical sensing GNN was combined with NEGF for quantum transport simulation.

[0082] In the electron-rich α-IGZO system, V O Induced MM bond defects not only introduce defect levels below the Fermi level in the band gap, but also reshape the conduction band bottom, significantly raising its energy and changing the effective mass of charge carriers.

[0083] To comprehensively assess this impact, this embodiment randomly selected nine α-IGZO supercells, each containing different MM bond configurations, for simulation. Figure 8 As shown, simulation results indicate that the device turn-on current decays, and all MM-bond configurations significantly reduce the transport performance of the device channel; some configurations even induce a weak negative differential resistance effect. This nonlinear transport characteristic stems from the fact that, in the presence of defects, the first conduction band becomes increasingly flat and decouples from other bands.

[0084] In order to investigate the physical mechanism of current decay, Figure 9 b shows the on-state current (I) on The relationship between I and the rise in the bottom energy of the MM bond directional band. Observation revealed that I on It exhibits an accelerating downward trend. In contrast, this embodiment artificially and independently introduces an equal-amplitude conduction band bottom energy rise at the center of a defect-free channel (changing only the barrier height, not the band shape), and the results show that the resulting I... on The decrease is much smaller than the actual decrease caused by the MM key.

[0085] This comparative experiment powerfully demonstrates that the influence of conduction band bottom reshaping on transport properties is synergistic and complex, and cannot be attributed solely to the barrier effect caused by the rise in conduction band bottom energy. It should also include multiple factors such as changes in effective mass and scattering.

[0086] Figure 9 c and 9d show the gate voltage (V) g The distributions of LDOS and transport coefficients of a-IGZO devices with defective channels at 0 V and 4 V further verify the rationality of the channel Hamiltonian model constructed in this application:

[0087] The figure clearly shows the bandgap widths in different regions, the deep-level defect states in defective supercells, and the smaller LDOS characteristics of the conduction band compared to the valence band. Simultaneously, the analysis of the transport coefficients also verifies the transport blocking effect induced by conduction band bottom remodeling.

[0088] With V g The application of [something] suppresses the rise of the conduction band bottom, thereby enhancing the peak transmission coefficient near the conduction band bottom. It should be noted that, due to the high defect concentration and restricted carrier paths (which must pass through defect regions) used in the simulation, the suppression effect of defects on transport observed in the above simulation may be more significant than in actual large-scale devices. However, this precisely reflects the GNN model's ability to capture local microscopic defect mechanisms.

[0089] In summary, this embodiment not only reproduces the electronic structure consistent with DFT at the microscopic level, but also confirms through macroscopic device simulation that the Hamiltonian constructed by this method can accurately describe the scattering mechanism introduced by defects, providing a reliable quantum transport simulation tool for the design and yield analysis of actual semiconductor devices.

[0090] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that other embodiments can be implemented without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the claims are intended to be included within this application. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0091] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for constructing a tight-binding model of a defective, aperiodic semiconductor device, characterized in that, The method steps are as follows: Step S1: Data acquisition and graph structure construction. Obtain the atomic structure of the defect-free supercell and the defective supercell of the material to be simulated. Calculate the band structure of each supercell using DFT as training labels. Transform the atomic structure of the supercell into a graph structure, where atoms are nodes and chemical bonds are edges. Extract node features, edge features, and global features. Step S2: Construct a physical perception graph neural network, build a GNN architecture including a feature encoder, a symmetric graph attention layer and a parameter predictor. The network updates the node state through symmetric message passing and outputs the parameters of the TB model. The parameters include orbital in-potential energy and transition parameters, and a physical constraint that decays exponentially with distance is applied to the output transition parameters. Step S3: Defect-aware co-training. Input the data of defect-free supercells and defective supercells into the GNN architecture at the same time, and perform co-training using a shared weight strategy. Calculate the weighted error between the predicted band and the DFT calculated band, so that the network establishes a mapping relationship between the changes in the local atomic environment caused by defects and the changes in TB model parameters. Step S4: Device channel Hamiltonian matrix assembly. Construct a device channel structure composed of defect-free supercell chains and embed defective supercells. Based on the trained model, predict the Hamiltonian matrix inside the supercell and between the nearest neighbor supercells. Use the transition matrix between defect-free supercells as the transition matrix between defect-free and defective supercells. Assemble to form a complete defective, non-periodic semiconductor device channel Hamiltonian matrix for quantum transport simulation.

2. The method according to claim 1, characterized in that, In step S1, the connection determination condition for the atomic graph edges constructed by the graph structure is: interatomic distance d ij < R i + R j + Δr In the formula, R i , R j denoted as the covalent radius of the atom, and Δr as the delocalization offset; For cations in oxide semiconductor materials, the delocalization offset Δr is set to a positive value to cover the delocalization effect of the outer electron orbitals of metal atoms, ensuring that the graph structure contains long-range electronic interactions.

3. The method according to claim 1, characterized in that, In step S2, the symmetric graph attention layer employs a symmetric message passing mechanism, with an attention coefficient α. ij The calculation is based on the mean operation of node features: In the formula, x i , x j These are the feature vectors of nodes i and j, respectively, e ij Let α be the edge feature vector. T Here, represents the weight vector, || denotes the concatenation operation, LeakyReLU is the non-linear activation function, and Softmax is the normalization function; The attention coefficient is forced to satisfy α by taking the average. ij = α ji This ensures that the predicted interatomic interactions conform to physical symmetry.

4. The method according to claim 1, characterized in that, In step S2, the physical constraint refers to the transition parameter t of the original output of the network. ij ' Perform the following forced correction: In the formula, r ij r is the interatomic distance. cut This is the cutoff radius parameter; The physical constraints force the transition intensity to decay exponentially with increasing interatomic distance, ensuring that the generated TB model is sparsity and physically interpretable.

5. The method according to claim 1, characterized in that, In step S3, the defect-aware co-training achieves parameter prediction by sharing network weights and utilizing the local environmental equivariance of the GNN architecture: For regions far from the defect center in a defective supercell, the network predicts Hamiltonian parameters that are consistent with those of the defect-free supercell based on the local atomic coordination environment consistent with that of the defect-free supercell. For the region near the defect center, the network automatically adjusts the Hamiltonian parameters to fit the defect state by sensing the geometric distortions of bond length, bond angle, and coordination number.

6. The method according to claim 1, characterized in that, In step S3, the loss function used for co-training includes a Fermi-Dirac weighting factor: In the formula, E is the band eigenvalue and α is the attenuation coefficient; The weighting factor assigns higher weights to the band eigenvalues ​​near the conduction band bottom and valence band top, and the weights decrease exponentially as the energy moves away from the band edge, thereby improving the simulation accuracy within the critical energy window of the device's turn-on and turn-off characteristics.

7. The method according to claim 1, characterized in that, In step S3, the training process employs a fully differentiable band eigenvalue solver. This solver utilizes vectorized matrix operations to construct the Hamiltonian matrix and incorporates Hermitian coercion operations, supporting end-to-end gradient backpropagation during the band eigenvalue solving process.

8. The method according to claim 1, characterized in that, In step S4, the approximate representation of the interface coupling matrix is ​​based on the following strategy: Based on the physical fact that the atomic environment at the edge of a defective supercell has converged to a defect-free state, the transition matrix H between defect-free supercells predicted by the model is directly selected. NN DF-DF H is the coupling transition matrix at the connection between defect-free and defective supercells. NN DF-DC This addresses the problem of not being able to directly calculate interface coupling parameters under aperiodic boundary conditions.

9. The method according to claim 1, characterized in that, In step S4, the full-channel Hamiltonian assembly is performed by assembling the predicted defect-free supercell intracellular matrix, defective supercell intracellular matrix, and interface coupling matrix into a block tridiagonal matrix according to the geometric design of the device channel, and then inputting the matrix into an unbalanced Green's function solver to simulate the current-voltage characteristics and electron transmission spectrum.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein, When the processor executes the program, it can implement the method as described in any one of claims 1-9.

Images