Environment-dependent tight-binding models from ab initio pseudo-atomic orbital Hamiltonians

This paper introduces an environment-dependent tight-binding framework that augments Slater-Koster hopping integrals with bond-screening functions, fitted to ab initio pseudo-atomic orbital Hamiltonians across multiple configurations to generate transferable, large-scale electronic structure models with ab initio precision.

The Gist

Imagine you are trying to predict how a massive crowd of people will move through a city.

The Old Way (DFT):
Currently, the most accurate way to do this is to treat every single person as a unique individual with their own thoughts, speed, and mood. You simulate every single interaction. This is incredibly accurate, but it's so computationally expensive that you can only simulate a small neighborhood (a few hundred people) before your computer crashes. This is like Density Functional Theory (DFT) in physics: it's the "gold standard" for accuracy but can't handle huge systems.

The Shortcut (Standard Tight-Binding):
To simulate a whole city, scientists used to use a shortcut called Tight-Binding (TB). Instead of simulating every person's mind, they just said, "If two people are standing next to each other, they move at speed X. If they are far apart, they don't interact."

• The Problem: This shortcut usually relies on "rules of thumb" fitted to a specific, perfect city layout. If you try to use those same rules for a city with a park, a highway, or a construction site (different environments), the predictions fall apart. The rules don't adapt to the surroundings.

The New Solution (This Paper's EDTB):
The authors of this paper, led by M. Buongiorno Nardelli, have invented a "Smart Shortcut" called Environment-Dependent Tight-Binding (EDTB).

Here is how it works, using simple analogies:

1. The "Perfect Blueprint" (PAO Hamiltonians)

First, they take the super-accurate, slow simulation (DFT) and translate it into a compact, easy-to-read "blueprint" called a Pseudo-Atomic Orbital (PAO) Hamiltonian.

• Analogy: Imagine taking a high-definition 3D movie of a city and compressing it into a simple 2D map that still keeps all the essential details. This map is the "truth" they want to match.

2. The "Social Butterfly" Effect (Environment Dependence)

In their new model, the "rules" for how atoms interact aren't fixed. They change based on how crowded the neighborhood is.

• Analogy: Think of a handshake between two people.
  • In an empty room, the handshake is firm and direct.
  • In a crowded party, the handshake is weaker because other people are bumping into them, blocking the connection.
  • The Innovation: The authors added a "Screening Factor" to their math. This factor acts like a "crowd meter." If an atom is surrounded by many neighbors, the model automatically weakens the connection between two specific atoms. If the atom is alone, the connection stays strong.
  • This allows the model to work perfectly whether the atoms are in a solid block, on a rough surface, or inside a complex machine part.

3. The "Master Trainer" (Fitting to Many Shapes)

Usually, scientists train these models on just one perfect crystal shape. The authors trained their model on many different shapes at once (stretched, squished, surfaces, interfaces).

• Analogy: Instead of teaching a driver only how to drive on a straight highway, they taught the driver on highways, dirt roads, icy streets, and parking lots simultaneously. Now, the driver (the model) knows how to handle any situation without needing a new lesson for every new road.

4. The Results: From Neighborhoods to Continents

Because this new method is both accurate (it matches the "perfect blueprint") and flexible (it adapts to crowds), they could simulate systems that were previously impossible:

• Twisted Graphene: They simulated a stack of graphene sheets twisted at a specific angle, containing over 4,000 atoms.
• The Scale: Doing this with the old "perfect" method would have taken a supercomputer years. Their new method did it in minutes on a standard workstation.

Why Does This Matter?

This is like upgrading from a hand-drawn map to a GPS that updates in real-time.

• For Scientists: They can now design new materials (like better batteries, faster computer chips, or solar cells) by simulating huge, complex structures with the accuracy of the most advanced physics, but at a fraction of the cost.
• For the Future: It bridges the gap between "theoretical perfection" and "practical application," allowing us to explore materials that are too big or too complex to study any other way.

In a nutshell: They took a super-accurate but slow physics engine, taught it to understand how "crowded" an atom's neighborhood is, and turned it into a fast, adaptable tool that can simulate massive materials with the same precision as the slow method.

Technical Summary

1. Problem Statement

The electronic structure of solids is fundamental to understanding physical phenomena like charge transport and optical response. While Density Functional Theory (DFT) provides a parameter-free, high-accuracy description, its computational cost scales poorly with system size ($N^3$), limiting routine application to unit cells of a few hundred atoms.

Tight-Binding (TB) models offer a computationally efficient alternative by using localized orbitals, but they are traditionally semi-empirical. Standard Slater-Koster (SK) TB models rely on fixed hopping integrals fitted to specific equilibrium geometries. Consequently, they suffer from poor transferability when applied to systems with varying local environments (e.g., surfaces, interfaces, defects, or strained lattices), as the standard two-center integrals cannot capture the screening effects of the surrounding atomic coordination.

2. Methodology

The authors propose a framework to construct Environment-Dependent Tight-Binding (EDTB) models that retain full ab initio accuracy while remaining computationally efficient. The methodology consists of three main pillars:

A. The PAO Reference Framework

The method utilizes the paoflow code to generate an exact ab initio Hamiltonian in a compact Pseudo-Atomic Orbital (PAO) basis directly from plane-wave DFT calculations.

• Projection: Bloch states are projected onto a set of orthonormal pseudo-atomic orbitals.
• Hamiltonian Construction: An exact PAO Hamiltonian ($\bar{H}$) is constructed from the projected eigenstates and eigenvalues. This Hamiltonian spans the same Hilbert space as the DFT calculation, ensuring no loss of accuracy, and serves as the "ground truth" for fitting.

B. The EDTB Formalism

The authors extend the classic Slater-Koster framework by introducing two key physical mechanisms:

1. Bond-Dependent Screening: Each hopping integral $V_{\lambda}$ is modulated by a screening factor dependent on the local coordination environment:
   $$\tilde{V}_{\lambda}(i, j) = V_{\lambda} \exp(-\gamma_{\lambda} S_{ij})$$
   Here, $S_{ij}$ is a bond screening sum calculated over all atoms $k$ near the bond $ij$, representing how "crowded" the local environment is. The parameter $\gamma_{\lambda}$ controls the screening strength for specific orbital channels (e.g., $pp\sigma$, $pd\pi$).
2. Distance-Dependent Hopping: Instead of discrete shell parameters, the model employs a continuous functional form (Goodwin–Skinner–Pettifor) for hopping integrals $V(r)$, allowing for smooth interpolation across continuously varying bond lengths.

C. Multi-Geometry Fitting Protocol

To determine the parameters (hopping amplitudes, decay exponents, and screening strengths), the authors perform a non-linear least-squares fit using the Levenberg–Marquardt algorithm.

• Objective: Minimize the Root-Mean-Square Error (RMSE) between the EDTB eigenvalues and the PAO eigenvalues across a uniform Brillouin zone mesh.
• Analytic Jacobians: Derivatives of eigenvalues with respect to parameters are computed analytically using the Hellmann–Feynman theorem, ensuring high speed and numerical stability.
• Multi-Geometry Training: Crucially, the fitting is performed simultaneously on multiple atomic configurations (e.g., different volumes, surface slabs, or stacking sequences). This breaks the degeneracy between hopping and screening parameters, forcing the model to learn physically meaningful environment dependence rather than just fitting a single geometry.

3. Key Contributions

• Exact Ab Initio Representation: The method bridges the gap between DFT and TB by using PAO Hamiltonians as the reference, avoiding the gauge ambiguities often associated with Wannier function downfolding.
• Transferable Environment Dependence: By explicitly modeling bond screening and distance dependence, the resulting Hamiltonians are transferable to complex geometries (surfaces, interfaces, alloys) without requiring new DFT calculations for the target system.
• Efficient Optimization: The use of analytic Jacobians allows for rapid convergence (typically $\sim 100$ iterations) even for models with 30–40 parameters, overcoming the inefficiency of derivative-free methods like genetic algorithms.
• Scalability: The implementation utilizes block-sparse tensor formulations and sparse eigensolvers (Lanczos algorithm), enabling the simulation of systems with thousands of atoms on standard workstations.

4. Results and Applications

The authors validated the method on four prototypical systems:

1. Bulk Platinum (FCC):

   • Test: Fitting across isotropic strains ($\pm 4\%$).
   • Result: The standard SK model (fitted only at equilibrium) failed at high strain (RMSE $\approx 815$ meV at $-4\%$). The EDTB-multi model maintained high accuracy (RMSE $\approx 244$ meV combined) across all strains, demonstrating genuine transferability.
   • Application: Successfully calculated the intrinsic Spin Hall Conductivity (SHC) by adding spin-orbit coupling (SOC) to the Hamiltonian, matching first-principles relativistic results.

2. Silicon Surfaces:

   • Test: Training on bulk Si and unrelaxed slabs of (001), (110), and (111) surfaces.
   • Result: The model correctly captured the modification of $d$-orbital hybridization at surfaces due to reduced coordination. It successfully generated thick slabs (40 atoms) showing correct projected bulk gaps and surface states.

3. Si/Ge [001] Superlattices:

   • Test: Modeling binary heterostructures with a multi-parameter framework (bulk vs. interface labels).
   • Result: The model reproduced realistic valence and conduction band offsets and correctly predicted the valley splitting ($\Delta E_{vs}$) in strained Si quantum wells. The splitting oscillated with well width, matching the expected zone-folding physics of the Si $\Delta$-valley.

4. Twisted Bilayer Graphene (TBG):

   • Test: Modeling the continuous variation of interlayer coupling in moiré supercells.
   • Result: The model reproduced Dirac cone dispersion and band splitting across various twist angles.
   • Scalability: The authors successfully computed band structures for supercells containing up to 4,324 atoms (38,916 basis functions). Using sparse solvers, memory usage was reduced from $\sim 11$ GB (dense) to $\sim 250$ MB, making large-scale TBG simulations feasible on a single workstation.

5. Significance

This work presents a robust, scalable, and highly accurate method for generating tight-binding models directly from ab initio data. Its significance lies in:

• Bridging Scales: It enables the study of electronic, optical, and transport properties in large, complex nanostructures (thousands of atoms) with DFT-level accuracy, a regime previously inaccessible to standard DFT.
• Physical Insight: The extracted screening parameters provide physical insight into how local coordination modifies electronic coupling, offering a more interpretable model than black-box machine learning potentials.
• Practical Utility: Integrated into the paoflow software ecosystem, the method is ready for immediate use in calculating complex properties (e.g., topological invariants, transport coefficients) for materials design and discovery.