Reaching the thermodynamic limit of periodic CCSD cohesive energies and band gaps with denser Brillouin zone sampling

This paper presents a scalable distributed-memory implementation of periodic CCSD theory that enables dense Brillouin zone sampling (up to 216 k-points) to reliably extrapolate cohesive energies and band gaps to the thermodynamic limit, providing definitive benchmark values for eight semiconductors and insulators with errors of approximately 0.1–0.2 eV and 0.4 eV, respectively, compared to experimental data.

The Gist

Imagine you are trying to understand the behavior of a massive, perfect crystal, like a giant diamond or a block of salt. In the real world, these crystals are huge, containing trillions of atoms. However, to study them on a computer, scientists usually have to shrink the problem down to a tiny, manageable "mini-crystal" (a supercell) and pretend it repeats itself infinitely.

The problem with this mini-crystal approach is that it's like trying to understand the weather of an entire continent by only looking at a single backyard. You miss the big picture, leading to "finite-size errors"—mistakes caused by the sample being too small.

The Challenge: The "Pixelated" Crystal
In the world of quantum chemistry, the "size" of your sample is determined by how many points you sample in a mathematical space called the Brillouin zone. Think of these points as pixels in a digital image.

• Low resolution (few pixels): You get a blurry, inaccurate picture of the crystal's properties (like how tightly the atoms stick together or how much energy is needed to jump an electron across the gap).
• High resolution (many pixels): You get a crystal-clear, accurate picture.

The catch is that calculating these properties using the gold-standard method called CCSD (Coupled Cluster with Single and Double excitations) is incredibly expensive. It's like trying to render a 4K movie on a calculator. Because it's so computationally heavy, previous studies could only afford "low-resolution" images (small pixel grids), forcing them to guess what the high-resolution picture would look like. These guesses often led to significant errors.

The Solution: A Super-Powered Computer Team
The authors of this paper built a new, super-efficient software program that acts like a massive team of workers (running on up to 12 computer nodes with 96 cores each). This team can work together to handle a much larger "image" than ever before.

Instead of looking at a tiny grid of pixels, they were able to sample up to 216 points (a $6 \times 6 \times 6$ grid) in the Brillouin zone. This is like upgrading from a blurry thumbnail to a high-definition 4K image. With this new clarity, they could finally see the true, "thermodynamic limit"—the perfect, infinite crystal behavior—without needing to guess.

What They Found: The "Gold Standard" Benchmarks
Using this high-definition approach, the team calculated two key properties for eight common materials (like magnesium oxide, silicon, and diamond):

1. Cohesive Energy: How much energy it takes to pull the crystal apart into individual atoms.
2. Band Gap: The energy "gap" an electron must jump to conduct electricity (essentially, whether the material is an insulator or a semiconductor).

They compared their high-definition results to real-world experiments:

• Cohesive Energy: Their predictions were very close to reality, usually off by only 0.1 to 0.2 eV. They tended to slightly underestimate how tightly the atoms stick together.
• Band Gaps: Their predictions were also quite good, off by about 0.4 eV, but they tended to slightly overestimate the gap (predicting the material is a slightly better insulator than it actually is).

The "Indirect" Puzzle
Some materials have "indirect" band gaps, which are trickier to calculate. It's like trying to measure the distance between two points that aren't directly visible to each other. The authors discovered that standard guessing methods (extrapolation) often failed here, underestimating the gap. They developed a clever "composite" strategy—measuring the direct path first, then adding a correction for the indirect route—to get a much more accurate result.

The Titanium Dioxide Test
To prove their method works on more complex materials, they applied it to Rutile Titanium Dioxide (a common white pigment and photocatalyst). Their calculation predicted a band gap of 4.17 eV. This is slightly higher than the experimental value (which is around 3.9 eV after corrections), but the authors note that this small error is consistent with the method's known limitations and suggests that even more complex physics (like triple excitations) might be needed for perfect accuracy.

The Bottom Line
This paper doesn't just give new numbers; it provides a definitive benchmark. By using a massive computer team to generate "high-definition" data, the authors have established a new, reliable standard for how well the CCSD method works. They showed that while the method is excellent, it still has a small, predictable "blur" (error) of about 0.3–0.4 eV when compared to the real world. This helps other scientists know exactly how much they can trust similar calculations in the future.

Technical Summary

Technical Summary: Reaching the Thermodynamic Limit of Periodic CCSD Cohesive Energies and Band Gaps

Problem Statement
Periodic coupled-cluster theory with single and double excitations (CCSD) is a promising nonperturbative method for calculating the electronic structure of solids, offering systematic improvability over mean-field theories like Density Functional Theory (DFT). However, its application to the thermodynamic limit (TDL)—the physically relevant limit where finite-size errors are eliminated—has been severely constrained by computational cost. The storage and execution time of periodic CCSD scale as $O(n^4 N_k^3)$ and $O(n^6 N_k^4)$ respectively (where $n$ is the number of atoms per unit cell and $N_k$ is the number of $k$-points). Consequently, previous studies have been limited to $k$-point meshes no larger than $4^3$ for three-dimensional solids. This restriction prevents reliable extrapolation to the TDL, particularly for band gaps, which converge more slowly ($N_k^{-1/3}$) than ground-state energies ($N_k^{-1}$). Without access to sufficiently dense $k$-point sampling, predictions for cohesive energies and band gaps suffer from unquantified finite-size errors, hindering the establishment of definitive benchmarks for the CCSD level of theory.

Methodology
To address these limitations, the authors developed and applied a new, optimized distributed-memory software implementation of periodic CCSD and equation-of-motion CCSD (EOM-CCSD) based on the PySCF framework. Key methodological features include:

• Parallelization and Symmetry: The code utilizes both distributed and shared memory parallelism and exploits space-group symmetry to work with the irreducible Brillouin zone, enabling calculations on up to 12 nodes (96 cores each) with 1.5 TB of memory per node.
• Sampling Density: This implementation allows for Brillouin zone sampling with up to $N_k = 6^3 = 216$ $k$-points in double-zeta (DZ) basis sets, $5^3$ in triple-zeta (TZ), and $4^3$ in quadruple-zeta (QZ) basis sets without approximations.
• Extrapolation Schemes: The study employs specific extrapolation models to reach the TDL:
  • For cohesive energies, Hartree-Fock (HF) energies are fitted with a three-parameter model ($N_k^{-1} + N_k^{-2}$), while correlation energies (MP2 and CCSD) use a two-parameter model ($N_k^{-1}$).
  • For band gaps, the authors evaluate three models: Model A (leading order $N_k^{-1/3}$), Model AB (up to second order), and Model ABC (up to third order, including $N_k^{-1}$ terms).
  • A "GW-EOM" scheme is also tested, which assumes a proportionality between the finite-size error coefficients of EOM-CCSD and $G_0W_0@HF$ calculations to predict TDL band gaps using smaller $k$-meshes.
• Composite Approach for Indirect Gaps: Recognizing that indirect band gaps (sum of ionization potential and electron affinity) are difficult to converge due to $k$-mesh shifting requirements for the conduction band minimum (CBM), a composite approach is proposed. This separates the calculation into a direct gap at the valence band maximum (VBM) and a CBM-VBM offset, allowing for more stable extrapolation.
• Basis Sets and Potentials: Calculations utilize correlation-consistent basis sets (cc-pVXZ, $X=D, T, Q$) and GTH pseudopotentials. Basis set incompleteness errors are corrected via extrapolation (e.g., $X^{-3}$ form) or by comparing DZ, TZ, and QZ results.

Key Contributions and Results
The authors report converged CCSD cohesive energies and band gaps for eight simple semiconductors and insulators (C, Si, BN, BP, MgO, LiH, LiF, LiCl) and a new application to rutile TiO$_2$.

• Cohesive Energies: The study provides definitive benchmark numbers for CCSD cohesive energies, converged to approximately 0.1 eV.

  • CCSD tends to underestimate cohesive energies by an average of 0.1–0.2 eV compared to experiment (corrected for zero-point energy), while MP2 tends to overestimate them.
  • The accuracy of CCSD is found to be comparable to or better than standard DFT functionals (PBE, PBEsol, SCAN).
  • Pseudopotential errors are generally small (<0.05 eV) for most solids, though slightly larger for BN and C.

• Band Gaps: The paper establishes that reliable TDL extrapolation for band gaps requires $k$-meshes significantly larger than previously accessible ($N_k \ge 3^3$).

  • Extrapolation Models: Simple single-term extrapolations (Model A) consistently underestimate band gaps by 0.3–0.5 eV when applied to small meshes ($N_k \le 4^3$). The three-parameter Model ABC, fitted over $N_k=3^3$–$6^3$, is identified as the most robust reference.
  • GW-EOM Scheme: The GW-EOM approach (using $G_0W_0$ to correct CCSD finite-size errors) shows mixed reliability. It works well for some materials (MgO, LiCl, LiF) but overestimates gaps by 0.4–0.5 eV for others (BN, BP, Si, C), indicating the proportionality assumption of error coefficients does not universally hold.
  • Indirect Gaps: The composite approach successfully resolves convergence issues for indirect gaps (e.g., in BN), revealing that direct gap extrapolations at the VBM are stable, while the CBM offset converges rapidly.
  • Accuracy: The final TDL EOM-CCSD band gaps exhibit a mean absolute error (MAE) of 0.39 eV (with pseudopotentials) and 0.42 eV (all-electron) relative to zero-point corrected experimental values. The method typically overestimates band gaps.
  • TiO$_2$: For rutile TiO$_2$, the TDL EOM-CCSD prediction is 4.17 eV, overestimating the ZPR-corrected experimental gap (~3.9 eV) by ~0.3 eV.

• Single-Excitation Character: The authors note that materials with larger deviations from experiment (specifically rock-salt structures like MgO and LiH) tend to have lower single-excitation characters ($n_{1}^{IP} \approx 93-94\%$) compared to those with smaller errors ($>95\%$), suggesting that triple excitations may be necessary for higher accuracy in these systems.

Significance
This work represents the largest periodic CCSD calculations based on canonical orbitals to date, successfully resolving finite-size and basis-set errors to provide the first reliable benchmarks for CCSD cohesive energies and band gaps in the thermodynamic limit. The study demonstrates that while CCSD offers high accuracy for cohesive energies, its application to band gaps is sensitive to the extrapolation scheme and $k$-point density. The authors conclude that for practical calculations limited to $N_k \le 4^3$, the AB extrapolation model is the most accurate available approach, though it still carries a residual error of ~0.25 eV relative to the fully converged TDL. The results highlight that the remaining discrepancy between EOM-CCSD and experiment (0.3–0.4 eV) likely stems from insufficient treatment of electron correlation (specifically the need for triple excitations) or electron-phonon interactions, rather than finite-size errors. This sets a new standard for benchmarking more affordable methods and clarifies the current limitations of CCSD for solid-state electronic structure.