CN-121997773-A - Large-scale scaling method of large-scale supercell tight binding model based on feature screening

Abstract

The invention relates to the technical field of quantum transport simulation, and discloses a large-scale scaling method of a super-cell tight binding model based on feature screening. Aiming at the problem that the Hamiltonian matrix dimension in a large-scale supercell tight binding model increases sharply along with the size, so that the unbalanced Green function quantum transportation calculation cost is too high and a large number of tracks do not substantially contribute to transportation, three-layer technical schemes are provided, namely a node/edge masking mechanism is introduced in a graph neural network prediction stage, an independent item is selectively shielded according to the atom transportation importance, the full-atom Hamiltonian is scaled to an effective subspace, only tracks which are close to a fermi surface and contribute to an energy band structure obviously are reserved, and efficient scaling of the model is realized in the controllable range of key transportation parameter errors such as band gaps, effective quality and the like. The invention can obviously reduce the calculation cost of quantum transport simulation, breaks through the bottleneck of large-scale defect supercell chain simulation, and provides a high-efficiency and feasible scheme for the quantum transport research of a large-scale complex system.

Inventors

• LV YAWEI
• YI TONG
• JIANG CHANGZHONG

Assignees

• 湖南大学

Dates

Publication Date

20260508

Application Date

20260402

Claims (10)

1. 1. The large-scale scaling method of the super-cellular tight binding model based on the feature screening is characterized by comprising the following steps: Step S1, acquiring first sexual principle calculation basic data of a target semiconductor large-scale supercell system, wherein the basic data comprise energy band structures, state densities, atomic coordinates and inter-atomic bonding information of defective supercells and non-defective supercells; S2, introducing a node and edge masking mechanism, carrying out transport importance assessment on atoms in supercells based on state density and energy band characteristics obtained by calculation based on a first sex principle, and selectively shielding irrelevant atomic nodes and associated chemical bond edges which do not substantially contribute to quantum transport properties; Step S3, scaling the large-scale full-atom Hamiltonian matrix based on the unshielded reserved atom nodes in the step S2, extracting matrix rows and columns corresponding to reserved atom tracks, and constructing an effective subspace Hamiltonian matrix containing partial atoms; And S4, inputting the shielded supercell atomic structure characteristics into a graph neural network, performing fitting training by taking the first principle energy band structure in the step S1 as a target, outputting scaled tight binding model parameters within the range of a preset band gap and an effective mass error threshold, and using the parameters for unbalanced Green function quantum transport calculation.
2. 2. The method according to claim 1, wherein the step S2 of evaluating the transport importance of the atoms in the supercell and selectively shielding the atoms comprises the following steps: Based on a state density analysis result obtained by calculation of a first sex principle, screening out an atomic orbit which has obvious contribution to a band-edge state and a defect state near a forbidden band; And calculating the maximum matrix dimension upper limit which can be processed by a preset unbalanced green function solver, sequencing by combining the contribution degree of the atom tracks, determining a finally reserved high-importance atom set, and shielding the rest atoms outside the set.
3. 3. The method according to claim 1, wherein the constructing the effective subspace hamiltonian matrix containing partial atoms in step S3 comprises the following steps: establishing a global index mapping table aiming at a large-scale full-atom Hamiltonian matrix according to the unshielded atomic orbit reservation sequence in the step S2; And according to the global index mapping table, only extracting matrix blocks formed by intersecting rows and columns of corresponding indexes from the all-atom Hamiltonian matrix, and forming a dimension-reduced scaling matrix as an effective subspace.
4. 4. The method according to claim 1, wherein the graph neural network in step S4 includes a symmetric graph attention mechanism, wherein atoms in the supercell are mapped to nodes, interatomic chemical bonds are mapped to edges, and node features, edge features, and supercell global volume features are fused by a symmetric message passing mechanism.
5. 5. The method of claim 4, wherein the node features include at least electronegativity, number of valence electrons, covalent radius, and ionic charge of atoms, and the edge features include at least a bond vector and bond types divided by two atom categories connected.
6. 6. The method of claim 1, wherein the tight-tie model parameters output by the graph neural network in step S4 include a bit energy parameter and a transition integral parameter, and wherein in a prediction phase of the graph neural network, a physical space truncation constraint is applied to the transition integral parameter so as to satisfy a formula of exponential decay with interatomic distance: Where t ij is the output effective transition integral, t ij ' is the initial transition integral predicted by the neural network, r ij is the spatial distance between atom i and atom j, and r cut is the set truncation radius.
7. 7. The method according to claim 1, wherein in the fitting training process of the graph neural network in step S4, instead of performing a full-scale fitting on the full-band energy band, an energy window is used to extract the band-edge state and defect-state energy bands near the forbidden band as core training targets.
8. 8. The method according to claim 7, wherein the loss function of the graph neural network includes penalty terms for bandgap error and effective mass error, and the energy band structure of the output tight tie model near the fermi surface and the error of the first principle calculation result are controlled within a preset threshold value by iteratively optimizing the network weights.
9. 9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the feature screening-based large scale supercell tight tie model scaling method of any of claims 1 to 8 when the program is executed by the processor.
10. 10. A computer readable storage medium having stored thereon a computer program, which when executed by a processor implements a feature screening based large scale supercell tight tie model scaling method according to any of claims 1 to 8.

Description

Large-scale scaling method of large-scale supercell tight binding model based on feature screening Technical Field The invention relates to the technical fields of computational mechanics, quantum transport simulation and semiconductor device design, in particular to a large-scale supercell tight constraint (Tight-Binding, TB) model scale scaling method based on feature screening, which is particularly suitable for quantum transport simulation of a defect-containing and amorphous semiconductor system (such as amorphous oxide and silicon carbide), and can be combined with a multi-layer graph neural network (Graph Neural Network, GNN), density functional theory (Density Functional Theory, DFT) and a Non-equilibrium green Function (Non-Equilibrium Green's Function, NEGF) method containing a symmetrical graph attention mechanism to realize high-efficiency and accurate simulation of a large-scale, defect-containing and Non-periodic supercell chain. Background As the feature size of semiconductor devices continues to shrink to nanometer or even atomic dimensions, quantum effects (such as carrier tunneling and quantum confinement) and defect effects (such as oxygen vacancies and interstitial atoms) become core factors determining the electrical performance of the devices, the conventional drift-diffusion model cannot accurately describe carrier transport behavior, and a transport simulation tool based on quantum mechanics is needed to support the design of next-generation semiconductor devices. The TB model is taken as a core framework of quantum transport simulation, is combined with NEGF, can describe carrier transport of different electronic structural feature materials under reasonable calculation cost, and is a key bridge for connecting first-principle calculation and device-level modeling. However, the existing TB model construction and scale optimization methods have significant technical bottlenecks: The large-scale supercell calculation cost is high, namely as the dimension of the supercell increases (such as a thousand-atom-level supercell containing defects), the Hamiltonian matrix dimension of the TB model increases in square level, so that the calculated amount of subsequent NEGF quantum transport simulation is increased sharply, even the calculated amount exceeds the existing calculation load bearing range, and the simulation of a large-scale defect supercell chain cannot be realized. The existing model lacks a high-efficiency characteristic screening mechanism, namely a great number of atomic orbits in supercells do not substantially contribute to energy band structures and transport properties in the actual quantum transport process, but the existing TB model construction methods, such as empirical parameter fitting and Tennian transformation (Wannier Transformation), all-atomic orbits are required to be reserved, and a selective screening mechanism based on atomic importance, such as defect and band-edge state contribution, is not established, so that a great amount of redundant information exists in the matrix. The periodic dependence and defect system suitability are poor, the TB model construction method based on the Tennian transformation depends on the crystal periodicity, and is difficult to be suitable for non-periodic systems containing defects, amorphous and the like, while the experience fitting method lacks mobility, and the model scale cannot be flexibly adjusted to balance the precision and the efficiency, so that the application of the method in industrial-grade semiconductor devices (such as thin film transistors and power devices) is further limited. The calculation complexity of the existing method is increased in a super-linear way along with the atomic number (for example, the calculation complexity of the Tennial transformation is close to O (N3)), and the simulation requirement on large-scale defect-containing supercells (the atomic number is more than 1000) in the design of a semiconductor device is increasingly urgent, and the contradiction between the two becomes a core obstacle for restricting the industrialized application of quantum transport simulation. Therefore, the method for efficiently screening the core track, scaling the TB model scale and adapting to the aperiodic defect system is developed on the premise of ensuring the precision of key transport parameters, and has important significance for breaking through the bottleneck of large-scale quantum transport simulation and promoting the design precision and efficiency of the semiconductor device. Disclosure of Invention Aiming at the technical bottlenecks that the prior large-scale defect-containing supercell TB model has extremely high cost of NEGF quantum transportation calculation caused by oversized Hamiltonian matrix dimension, a large number of deep tracks do not substantially contribute to transportation property and the like, the invention provides a large-scale supercell tight binding model scale scaling