High-Performance Tight-Binding Framework

• High-performance tight-binding frameworks are software infrastructures that model electronic structures using parameterized Hamiltonians and scalable, optimized algorithms.
• They employ methods like exact diagonalization and Chebyshev propagation to efficiently simulate quantum systems ranging from thousands to billions of orbitals.
• These frameworks integrate advanced features such as machine learning-based parameterization, multi-level parallelism, and GPU acceleration to support comprehensive property calculations.

A high-performance tight-binding (TB) framework is a software and methodological infrastructure for modeling electronic structure, transport, and related properties in large-scale quantum systems using parameterized TB Hamiltonians, specifically engineered to maximize computational efficiency, scalability, and physical accuracy. Modern high-performance TB frameworks integrate advanced algorithms—such as order-$N$ propagation and sparse-matrix techniques—with highly optimized software stacks (leveraging C++, GPU acceleration, hybrid MPI/OpenMP parallelism, and Python APIs) to efficiently simulate materials systems from thousands to billions of orbitals. These frameworks are central to first-principles modeling, device simulations, machine learning–accelerated electronic structure, and rapid high-throughput screening in condensed matter and materials science.

1. Mathematical Foundations of High-Performance Tight-Binding

The canonical TB Hamiltonian is

$$\hat{H} = \sum_i \epsilon_i c_i^\dagger c_i - \sum_{i\neq j} t_{ij} c_i^\dagger c_j$$

where $\epsilon_i$ are on-site energies and $t_{ij}$ are hopping integrals, as in Eqs. (1–3) of (Li et al., 2022). In operator language with orbital basis $\{| \phi_i \rangle\}$,

$$\hat{H} = \sum_{i,j} H_{ij} |\phi_i\rangle\langle\phi_j|,$$

with $H_{ij}$ determined via orbital integrals: $$\epsilon_i = \int \phi_i^*(r) \hat{h}_0(r) \phi_i(r) dr, \qquad t_{ij} = -\int \phi_i^*(r) \hat{h}_0(r) \phi_j(r) dr.$$ For periodic systems, a Fourier transformation produces

$$H_{ij}(k) = \epsilon_i \delta_{ij} - \sum_{R\neq0} t_{ij}(R) e^{i k \cdot (R + \tau_j - \tau_i)}$$

(Eqs. 7–15, (Li et al., 2022)).

Recent frameworks generalize this formalism to include spin–orbit coupling, three-body/environmental corrections, non-orthogonal basis sets, analytic derivatives for Berry curvature, and explicit machine-learning parameterizations as in GPUTB (Wang et al., 8 Sep 2025) and TB3PY (Park et al., 25 Nov 2025, Garrity et al., 2021).

2. Core Computational Algorithms: Diagonalization and Propagation

Two principal algorithmic strategies dominate scalable TB calculations:

Exact Diagonalization: For relatively small systems, the $N \times N$ Hamiltonian is solved using dense eigensolvers. This approach scales as $O(N^3)$ in CPU and $O(N^2)$ in memory, feasible for $N\lesssim10^5$ (Li et al., 2022, Li et al., 30 Sep 2025, Klymenko et al., 2020). Diagonalization yields eigenpairs $(E_{n,k},|\psi_{n,k}\rangle)$ for band structures and localized observables.

Tight-Binding Propagation Methods (TBPM/KPM): For very large $N$, all major frameworks transition to matrix polynomial propagation (Chebyshev or Lanczos recursions) for time evolution or spectral moment evaluation: $$e^{-i H t} \approx \sum_{m=0}^M (2-\delta_{m0})(-i)^m J_m(t'||H||) T_m(\tilde{H})$$ (Eqs. 21–23, (Li et al., 2022, João et al., 2019)). This allows evaluation of spectral densities, dynamical correlations, and response functions via Fourier or Chebyshev-KPM expansion, achieving $O(N M)$ CPU and $O(N)$ memory scaling, enabling calculations on $N\sim10^{10}$ orbitals (João et al., 2019, Li et al., 30 Sep 2025).

Other critical kernels include partial diagonalization (e.g., FEAST contour solver), recursive Green's function (RGF) algorithms for transport (Klymenko et al., 2020), and efficient Hamiltonian assembly via nearest-neighbor graph traversal and sparse-matrix representations (Klymenko et al., 2020, Wang et al., 8 Sep 2025).

3. Advanced Feature Sets and Supported Physical Quantities

High-performance TB frameworks support comprehensive property calculations:

• Electronic structure: Band structures, density of states (DOS), local DOS, quasi-eigenstates
• Optical/transport: AC/DC conductivity (Kubo + KPM/TBPM), Hall conductivity (Kubo-Bastin), diffusion coefficients, carrier velocity, mobility, mean free path, localization length (Eqs. 25–37, (Li et al., 2022, João et al., 2019, Park et al., 25 Nov 2025))
• Topological invariants: $\mathbb{Z}_{2}$ index (Wilson loop), Chern number, Berry curvature (Kubo formula, Wilson loop integration), spin texture (Li et al., 30 Sep 2025)
• Wave-packet dynamics: Real-time/space propagation, quantum quench, decoherence
• Response functions: Polarization, dielectric and loss function, plasmon dispersion and lifetimes
• Parameter learning and extraction: Fitting TB Hamiltonians from DFT ab-initio (projection (Agapito et al., 2015), least-squares fit (Nakhaee et al., 2019)), ML-based parameterizations (GPUTB (Wang et al., 8 Sep 2025), SlaKoNet (Park et al., 25 Nov 2025)), and active learning against materials databases (Garrity et al., 2021, Park et al., 25 Nov 2025).

Frameworks like TBPLaS 2.0 introduce full hardware abstraction (CPU/GPU), analytical model interfaces, and high-level routines for spin/Berry/topological observables, with modular backends (Eigen, MKL, OpenBLAS, CUDA) and GPU-CUDA pipelines (Li et al., 30 Sep 2025, Wang et al., 8 Sep 2025).

4. Software Architectures, Performance Engineering, and Parallel Scalability

Contemporary high-performance TB frameworks are built with clear front-end/back-end separation:

• Front-ends: Python/Cython APIs for model construction, configuration, workflow scripting, and visualization (e.g., TBPLaS (Li et al., 30 Sep 2025), KITE (João et al., 2019), AFLOWπ (Supka et al., 2017))
• Back-ends: Compiled C++/Fortran/CUDA ‘core’ libraries for solvers, matrix algebra, and parallelization (Li et al., 30 Sep 2025, João et al., 2019, Wang et al., 8 Sep 2025, Klymenko et al., 2020)
• Parallelism: Multi-level parallelization with MPI distributing k-points or random seeds and OpenMP (and/or CUDA) threading inner linear algebra, Chebyshev recursions, and message-passing via ghost layers/domain decomposition (Li et al., 30 Sep 2025, João et al., 2019)
• GPU computation: Direct CUDA/CUBLAS utilization in matrix assembly, KPM propagation, and transport, with fused kernels for sparse operations (Li et al., 30 Sep 2025, Wang et al., 8 Sep 2025)
• Data management: Standardized input/output (JSON, XML, HDF5), modular plug-ins for new methods, and post-processing separation (e.g., for spectral folding in KITE (João et al., 2019))

Major frameworks—TBPLaS 2.0 (Li et al., 30 Sep 2025), KITE (João et al., 2019), GPUTB (Wang et al., 8 Sep 2025), TB3PY (Park et al., 25 Nov 2025)—report (for their largest test cases) efficient handling of $10^{7}\!-\!10^{10}$ orbitals with near-linear scaling in $N$, $M$ (Chebyshev steps), and thread/process count, with wall-time for gigantic systems ranging from minutes to a few hours on large-memory CPUs or modern GPUs.

Framework	Max $N$ Demonstrated	Scaling	CPU/GPU Support	Core Language	Python API	Key Solvers
TBPLaS 2.0	$>10^9$	$O(N\,M)$	Yes	C++ (core), Python/Cy	Yes	Diag, TBPM, FEAST
KITE	$10^{10}$	$O(N\,M)$	No (CPU only)	C++, Python	Yes	Chebyshev-GF, KPM
GPUTB	$10^8$–$10^9$	$O(N)$	Yes (CUDA)	Python (PyTorch-CUDA)	Yes	KPM, ML Hamiltonian
NanoNET	$10^5$–$10^6$	$O(N\log N)$	CPU	Python, C++	Yes	NEGF, RGF, BTD
TB3PY	$10^5$ (benchmarks)	$O(N)$	CPU	Python, C++/Julia	Yes	Diag, SCF

5. High-Throughput, Machine Learning, and Parameterization Strategies

Recent frameworks emphasize (1) automation of TB Hamiltonian generation, (2) active learning or ML-based parametrization, and (3) minimal user intervention:

• Automatic TB extraction from DFT: Projection schemes (Agapito et al., 2015) and PAO-based methods (Supka et al., 2017) build minimal-orbital TB Hamiltonians via projection of Bloch states, with filtering to remove low-projectability bands and shifting to avoid spurious eigenvalues.
• ML/NN parameterizations: Frameworks like GPUTB (Wang et al., 8 Sep 2025) and SlaKoNet (Park et al., 25 Nov 2025) map local atomic environments to TB parameters using descriptor–MLP architectures (Chebyshev/Laguerre polynomials, message-passing networks), supporting transferability across functionals and materials classes.
• Industrial-scale fitting and benchmarking: Large parameter databases (e.g., JARVIS-QETB, ThreeBodyTB.jl (Garrity et al., 2021)), active-learning cycles for error reduction, and CHIPS-TB (Park et al., 25 Nov 2025) for systematic benchmarking, ensure wide applicability and robust performance.
• Integration with first-principles and workflows: TB generation and property computation are now driven by high-level workflow managers and standardized APIs (e.g., AFLOWπ (Supka et al., 2017), TBStudio (Nakhaee et al., 2019)), supporting rapid computation of band gaps, elastic constants, dielectric response, phonon dispersions, and more.

6. Specialized Extensions and Domain-Specific Innovations

High-performance TB frameworks have been adapted for various frontier domains:

• Photonic crystals: The transversality-enforced TB (TETB) construction (Morales-Pérez et al., 2023) employs topological quantum chemistry and auxiliary longitudinal modes to overcome the lack of exponential localization of photonic Wannier functions in 3D, yielding minimal TB models with faithful photonic band reproduction and symmetry/topology preservation.
• Kinetic energy functional reconstruction: The tight-binding expansion of nonlocal kinetic energy density functionals (KEDF) dramatically accelerates iterative orbital-free DFT by approximating the nonlocal part via a frozen, superposed atomic density, resulting in up to $10^2$ speedup for $>10^4$ atoms with negligible loss in accuracy (Chen et al., 2024).
• Large-scale semiconductor benchmarking: Standardized frameworks like CHIPS-TB evaluate and compare DFTB, machine-learned, and three-body TB parameterizations against ab initio and experimental data, providing systematic error and transferability analyses for over 50 materials (Park et al., 25 Nov 2025).

7. Performance Benchmarks and Practical Impact

Performance is consistently validated by head-to-head comparisons:

• TBPLaS 2.0 achieves $>10\times$–$1900\times$ speedups in modeling tools and solvers over previous versions and up to $35\times$ acceleration in DC conductivity via TBPM, extending feasible system sizes to $N>10^9$ (Li et al., 30 Sep 2025).
• KITE achieves linear scaling in $N$ and $M$ to $N\sim10^{10}$ orbitals, with wall times of minutes to a few hours per calculation, and 90–98% parallel efficiency (João et al., 2019).
• GPUTB demonstrates full-pipeline electronic structure and transport (DOS, conductivity) on $10^8$–$10^9$ atom systems in hours on a single GPU (Wang et al., 8 Sep 2025).
• TB3PY/CHIPS-TB, via three-body corrections and self-consistent charge, achieve RMS atomization energy errors of $\approx0.02$ eV/atom and band gap errors of $\lesssim0.5$ eV for diverse structure prototypes (Park et al., 25 Nov 2025, Garrity et al., 2021).

This performance landscape enables first-principles–based simulations, device-scale electronic and quantum transport, ab-initio-level high-throughput screening, and the modeling of complex topological, spintronic, and plasmonic phenomena well beyond the reach of DFT alone.

---

References:

• TBPLaS: (Li et al., 2022, Li et al., 30 Sep 2025)
• KITE: (João et al., 2019)
• AFLOW$\pi$: (Supka et al., 2017)
• GPUTB: (Wang et al., 8 Sep 2025)
• CHIPS-TB/TB3PY: (Park et al., 25 Nov 2025, Garrity et al., 2021)
• TETB (photonic crystals): (Morales-Pérez et al., 2023)
• TB-KEDF for OF-DFT: (Chen et al., 2024)
• NanoNET: (Klymenko et al., 2020)
• Projection approaches: (Agapito et al., 2015)
• TBStudio: (Nakhaee et al., 2019)

Each of these frameworks defines reference standards for scalable, robust, and extensible tight-binding–based simulation in contemporary computational materials science.