Generalized deformation potential and machine-learning approaches for electron-phonon coupling and thermoelectric transport in semiconductors

This paper introduces two computationally efficient methods—a generalized acoustic deformation potential model and a machine-learning interpolation technique—that utilize a small number of first-principles electron-phonon matrix elements to accurately predict thermoelectric transport properties in semiconductors, with the machine-learning approach demonstrating superior accuracy and ease of implementation.

The Gist

Imagine you are trying to predict how fast a crowd of people (electrons) can run through a busy, bumpy hallway (a semiconductor crystal). The hallway isn't empty; it's filled with vibrating walls and floor tiles (phonons) that constantly bump into the runners, slowing them down.

In the world of physics, calculating exactly how these "bumps" happen is like trying to map every single step of every runner in a stadium while the stadium is shaking. It's incredibly accurate, but it takes a supercomputer years to do the math. This is the "old way" described in the paper: using complex first-principles calculations (DFPT) to get the exact answer, but at a huge cost of time and energy.

The authors of this paper, Ransell D'Souza and Ivana Savić, asked: "Can we get a nearly perfect answer without doing all that heavy lifting?"

They say yes, and they offer two clever shortcuts to predict how well a material called MoS₂ (a thin, two-dimensional sheet of Molybdenum Disulfide) conducts electricity and heat. They tested these shortcuts on a single layer of MoS₂ and found they work almost as well as the super-computer method.

Here is how their two "shortcuts" work, explained simply:

1. The "Rulebook" Method (The Model Approach)

Think of this like a simplified rulebook for a board game.
Instead of simulating every single collision, the authors created a set of mathematical rules (a "deformation potential model") that describe how the electrons interact with the vibrating walls.

• The Trick: They figured out that you don't need to know the rules for every possible bump. You only need to calculate the rules for about 10 specific, representative bumps (using the expensive computer method once).
• The Result: Once they have those 10 numbers, they plug them into their general rulebook. This rulebook is flexible enough to work for any shape of crystal, not just perfect cubes. It's like having a master key that opens any door, rather than making a new key for every single lock.

2. The "Smart Guess" Method (The Machine-Learning Approach)

Think of this like teaching a student to draw a landscape.
Instead of asking the student to draw every single leaf on every tree (which is the expensive computer method), you show them about 100 sample sketches of the landscape.

• The Trick: The computer (Machine Learning) looks at these 100 sketches and learns the pattern of how the lines curve and where the bumps are. It then uses that pattern to "fill in the blanks" and draw the rest of the landscape perfectly.
• The Result: The computer learns the relationship between the electron and the vibration so well that it can predict the behavior on a super-detailed map without ever having to calculate the physics from scratch for every single point.

The Results: Did the Shortcuts Work?

The authors tested both methods on MoS₂, a material famous for being a great candidate for future electronics.

• Accuracy: Both shortcuts produced results that were almost identical to the expensive, super-accurate computer method.
• The Winner: The Machine-Learning method was the clear champion. It was not only more accurate but also much easier to set up. The "Rulebook" method required the authors to do a lot of complex math to derive the rules for different crystal shapes, whereas the Machine-Learning method just needed the data and let the computer figure out the rest.
• Real-World Check: When they used these methods to predict how well MoS₂ conducts electricity (mobility), the numbers matched the best experimental data available. They found that MoS₂ can conduct electricity very well, matching the top-performing samples found in real labs.

The Bottom Line

This paper is about efficiency. The authors showed that you don't need to run a marathon (the expensive computer calculation) to know the finish time. You can use a smart shortcut (Machine Learning) or a good rulebook (Physics Model) to get the same result in a fraction of the time.

They proved that for materials like MoS₂, these fast methods are reliable enough to be used for designing future electronic devices, saving researchers years of computing time.

Technical Summary

Technical Summary

Problem Statement
While first-principles methods combining density functional perturbation theory (DFPT) and Wannier or Fourier interpolation have enabled predictive calculations of electron-phonon (el-ph) coupling and thermoelectric transport in crystalline materials, these approaches remain computationally expensive. The necessity of using extremely dense reciprocal space grids (on the order of 100 points per direction) to accurately capture charge carrier transport near the Fermi level makes screening materials or studying large unit cells (e.g., heterostructures) prohibitive. Existing simplified models, such as the acoustic deformation potential model, often rely on oversimplifications (e.g., using a single parameter) that fail to capture the complexity of arbitrary crystal symmetries and band extrema locations. Furthermore, while machine learning (ML) has been applied to predict physical properties derived from el-ph coupling, its direct application to interpolating el-ph matrix elements on dense grids for transport calculations has not been extensively explored.

Methodology
The authors present two computationally inexpensive methods to obtain thermoelectric transport properties using a small number of first-principles calculated el-ph matrix elements (approximately 10 to 100 per electronic band and phonon mode), applied to single-layer MoS$_2$.

1. Generalized Model Approach:

   • The authors formulate a generalized acoustic deformation potential (DP) model applicable to arbitrary crystal symmetries and band extrema locations. This extends the Herring and Vogt approach, deriving explicit formulas for hexagonal symmetry with band extrema along the $\Gamma$-K line.
   • The model accounts for intravalley scattering via LA and TA acoustic modes using a DP tensor with independent components derived from $\sim$10 matrix elements.
   • It incorporates non-polar optical and intervalley scattering using deformation potential models, and polar optical scattering using a Fröhlich-like interaction model for 2D materials.
   • A specific parametrization is introduced for homopolar optical (ZO$_1$) modes to account for their small long-range contributions.

2. Machine-Learning (ML) Approach:

   • This method uses regression techniques to interpolate $\sim$100 el-ph matrix elements onto dense reciprocal space grids in the Brillouin zone regions relevant for transport.
   • Training datasets consist of first-principles matrix elements calculated on a coarse grid (DFPT+Wannier). The authors tested "uniform" datasets (q-points along high-symmetry lines) and "random" datasets.
   • For acoustic modes (LA, TA), a Polynomial Features (PFR) algorithm is used, reflecting the polynomial dependence of the deformation potential model.
   • For polar (LO$_2$) and homopolar (ZO$_1$) optical modes, a Gaussian Process Regressor (GPR) with a radial-basis function (RBF) kernel is employed to capture the more complex wave-vector dependence.
   • The resulting matrix elements are used to calculate scattering rates and momentum relaxation times via Fermi's golden rule, assuming a Kane model for band dispersion.

Key Contributions

• Generalized Deformation Potential Theory: The paper derives a rigorous formulation of the acoustic deformation potential model for arbitrary crystal symmetries and band extrema, specifically providing explicit tensor components for hexagonal systems with extrema at the K point.
• ML Interpolation Framework: The authors demonstrate a novel application of ML to directly predict el-ph matrix elements on dense grids using a sparse training set, bypassing the need for full Wannier interpolation of the Hamiltonian and gradients.
• Validation on 2D MoS$_2$: Both methods are applied to single-layer MoS$_2$, a material with significant thermoelectric power factors near 300 K. The study includes a detailed breakdown of scattering mechanisms (acoustic, polar optical, homopolar optical, and intervalley) and their respective contributions to mobility and scattering rates.

Results

• Agreement with State-of-the-Art: Both the generalized model and the ML approach show very good agreement with the computationally intensive DFPT+Wannier approach (implemented via the EPW code).
• Accuracy of ML: The ML method is found to be more accurate than the model approach, particularly for transverse acoustic (TA) modes where the model neglects piezoelectric long-range interactions. The ML approach achieves coefficients of determination above 90% with training sets of $\sim$100 matrix elements.
• Transport Properties:
  • Mobility: The calculated room-temperature electron drift mobility using the ML approach is 159 cm$^2$/(Vs), and 156 cm$^2$/(Vs) using the EPW code. These values align closely with the highest experimental reports (148 cm$^2$/(Vs)) and are consistent with recent first-principles studies that accurately treat long-range interactions.
  • Thermoelectric Coefficients: The calculated electrical conductivity and Seebeck coefficient as a function of carrier concentration agree well with experimental data. The study notes that discrepancies with lower experimental values can be attributed to extrinsic disorder effects (charged impurities, substrate scattering) rather than intrinsic limitations of the calculation.
• Scattering Mechanisms: The analysis confirms that longitudinal acoustic (LA) phonons dominate scattering, followed by polar optical (LO$_2$) modes. The homopolar (ZO$_1$) mode exhibits a cusp near the $\Gamma$ point due to long-range interactions, a feature correctly captured by both methods.

Significance
The paper claims that these methods provide a pathway to efficient, predictive calculations of thermoelectric transport properties without the prohibitive cost of full first-principles workflows on dense grids. The ML approach is highlighted as being particularly straightforward to implement and highly accurate, making it suitable for screening materials with desired transport properties or predicting properties for complex systems like heterostructures. The generalized deformation potential model offers a physically transparent framework that can be reused in subsequent studies (e.g., device modeling) without repeating the full ab-initio workflow. The work bridges the gap between simplified analytical models and expensive numerical simulations, offering a practical tool for semiconductor transport analysis.