betaFeedback

WorksPhysics and Astronomy

Correlation-Consistent Gaussian Basis Sets for Periodic Solids. I. Hydrogen and p-block Elements from Boron to Chlorine

Abstract

We present the cc-pVXZ-s series (X = D, T, and Q) of correlation-consistent Gaussian basis sets for solids, covering hydrogen and the p-block elements from boron to chlorine. The series includes basis sets of consistent quality for the all-electron potential and for various pseudopotentials and effective core potentials commonly used in periodic calculations. Our design combines Dunning's construction of the molecular cc-pVXZ series with a revised material-constrained atomic optimization (MCAO) algorithm for improved numerical performance. MCAO regularizes standard atomic optimization to suppress overly diffuse basis functions and ensure numerical stability in a set of reference solids. Benchmarks against standard cc-pVXZ calculations for molecules and plane-wave calculations for periodic solids establish the accuracy, transferability, and numerical stability of cc-pVXZ-s in both mean-field and correlated wavefunction calculations and confirm reliable extrapolation to the complete-basis-set (CBS) limit. As an application, we use all-electron cc-pVXZ-s to compute basis-set-converged lattice constants, bulk moduli, and cohesive energies for 7 bulk semiconductors. Using a hierarchy of correlated wavefunction methods up to local coupled cluster with singles, doubles, and perturbative triples [CCSD(T)], we find systematic convergence toward experimental bulk moduli and cohesive energies, while a small systematic overestimation of lattice constants suggests potential limitations in the approximate finite-size treatment. This work provides a foundation for the further development of systematically convergent Gaussian basis sets for periodic calculations.

Introduction

Atom-centered Gaussian basis sets are a fundamental ingredient of modern quantum chemistry. Standard libraries of well-benchmarked Gaussian basis sets, such as the correlation-consistent cc-pVXZ series, [1] the def2 series, [2, 3] and the polarization-consistent series, [4, 5] have made molecular quantum chemical simulations routine and systematically improvable at both mean-field and post-mean-field levels of theory. Augmentation with diffuse functions [6] or core-correlating functions [7] and adaptation to relativistic treatments [8, 9] have further expanded the scope of molecular quantum chemistry. These libraries and their extensions provide a reliable foundation for the development and application of new methods and algorithms. [10] [11]

The compactness and systematic improvability of Gaussian basis sets also make them attractive for correlated wavefunction calculations of periodic solids. 12 Most Gaussian basis sets used in molecular calculations are optimized on isolated atoms [1, 2, 3] or small molecules. [4, 5] Such optimization gives standard Gaussian basis sets broad transferability across chemical environments, [10] but can also produce diffuse basis functions that cause linear dependence in periodic solids. 13–16 This problem has been recognized since the early development of periodic Hartree-Fock (HF) and density functional theory (DFT) codes using Gaussian basis functions. 18,19,20,21 Basis-set linear dependence can destabilize self-consistent-field (SCF) calculations 20 and make it difficult to use the large basis sets needed for reliable convergence to the complete-basis-set (CBS) limit. 11,15 Common remedies include removing diffuse primitives 22–25 (e.g., those with exponents below 0.05 or 0.1 a.u.) and discarding nearly redundant linear combinations through canonical orthogonalization. 26,27

An alternative is to reoptimize basis-set parameters to improve numerical stability. Methods developed for this purpose fall into three broad classes. The first class combines molecular or solid-state energy optimization with an explicit linear-dependence penalty. In the molecular optimization (MOLOPT) approach of VandeVondele and Hutter, 13 the objective combines the energy and the overlap-matrix condition number κ\kappa evaluated on a set of reference molecules. Because κ\kappa measures basis-set linear dependence in a given system, it serves as an explicit target for improving numerical stability. The resulting MOLOPT basis sets and their successors 28,29 generally show high numerical stability in periodic systems. A recent large-scale DFT benchmark study demonstrated this stability even for highly compact solids in unconventional phases. 30 Several groups independently adapted the MOLOPT strategy to optimize basis sets on selected reference solids, 31–34 minimizing a combination of the energy and κ\kappa for those solids. The resulting basis sets are often system-specific, 31,32 and their transferability beyond the reference solids remains less well characterized than that of MOLOPT basis sets.

The second class uses direct solid-state energy optimization without an explicit linear-dependence penalty. The pob-TZVP basis sets developed by Peintinger and co-workers 14 are a prominent example. The authors selectively removed diffuse primitives from standard atom-optimized def2-TZVP basis sets [3] and reoptimized the remaining basis parameters on reference solids. Benchmarks using the CRYSTAL code35 showed that the pob-TZVP basis sets and their successors, the pob-rev2 series,36 are accurate and numerically stable in periodic DFT calculations of compact solids. For challenging systems, however, the pob-TZVP basis sets may require augmentation with diffuse functions,14,36 as recently demonstrated for Cu solids.16 Another example in this class is the work of Morales and Malone,37 who optimized Gaussian basis sets for individual solids to accelerate basis-set convergence in auxiliary-field quantum Monte Carlo calculations.

The third class regularizes standard atomic optimization to control linear dependence in periodic solids. One example is the GTH-cc-pVXZ series of Ye and Berkelbach,15 which comprises atom-optimized, correlation-consistent Gaussian basis sets for periodic calculations with Goedecker-Teter-Hutter (GTH) pseudopotentials.38,39 The key design principle is to limit the number of valence primitives, thereby constraining the smallest Gaussian exponent in each valence angular momentum channel to improve the numerical stability of the full basis set. The GTH-cc-pVXZ basis sets have been used in several recent studies employing correlated wavefunction methods for both ground-state properties25,40–42 and electronic excitations,43–47 demonstrating robust transferability across materials and methods, much like standard atom-optimized basis sets in molecules. We recently extended the Ye–Berkelbach approach through the material-constrained atomic optimization (MCAO) framework,16 which combines atomic energy optimization with a system-specific linear-dependence penalty to enable continuous control of linear dependence in periodic solids. We used MCAO to generate correlation-consistent Gaussian basis sets for Cu with various nuclear potentials and relativistic treatments and validated their accuracy, transferability, and numerical stability in both mean-field and correlated wavefunction calculations of Cu-containing solids, including systems beyond the reference solids used in the optimization.

Building on these developments, we present the cc-pVXZ-s series (X=D,T,and QX = \mathrm{D}, \mathrm{T}, \mathrm{and}\ \mathrm{Q}) of correlation-consistent Gaussian basis sets for solids, covering hydrogen and the pp-block elements from boron to chlorine with various nuclear potentials. We extend the original MCAO algorithm to accommodate an arbitrary number of reference solids and improve its numerical performance. This allows us to make cc-pVXZ-s numerically stable across a wide range of solids, including highly compact unconventional phases that cause strong linear dependence in atom-optimized basis sets. Benchmarks against standard cc-pVXZ calculations for molecules and plane-wave (PW) calculations for periodic solids establish the accuracy, transferability, and numerical stability of cc-pVXZ-s in both mean-field and correlated wavefunction calculations, including the reliable extrapolation to the CBS limit. As an application, we use all-electron cc-pVXZ-s to compute basis-set-converged lattice constants, bulk moduli, and cohesive energies for 7 bulk semiconductors. Using a hierarchy of correlated wavefunction methods up to local coupled cluster with singles, doubles, and perturbative triples48 [CCSD(T)], we find systematic convergence toward experimental bulk moduli and cohesive energies, while a small systematic overestimation of lattice constants suggests potential limitations in the approximate finite-size treatment.

The rest of the paper is organized as follows. Section II presents general theoretical considerations for addressing basis-set linear dependence in solids and describes a revised MCAO algorithm with improved numerical performance. Section III presents the design of the cc-pVXZ-s basis sets for various nuclear-potential treatments. Section IV gives the computational details. Section V presents numerical calculations that validate various aspects of the cc-pVXZ-s basis sets and demonstrate their use in correlated wavefunction studies of solids. Section VI concludes with an outlook on future work.

Theory

Periodic Gaussian basis sets

Throughout this work, we consider a periodic solid described by NAON_{\mathrm{AO}} translationally adapted atom-centered Gaussian-type orbitals (GTOs) per unit cell,

χkμ(r)=∑Reik⋅Rμ(r−R),(1)\chi_{k\mu}(r) = \sum_R e^{ik\cdot R}\mu(r-R), \tag*{(1)}

where μ(r−R)\mu(r-R) denotes an isolated GTO translated to the unit cell at lattice vector RR, and k∈Kk \in\mathcal{K} denotes a point in a uniform kk-point mesh K\mathcal{K} of size NkN_k sampling the first Brillouin zone. The isolated GTOs can be either primitive or contracted. A primitive GTO (pGTO) ξilm\xi_{ilm} is a single Gaussian function characterized by its exponent αi\alpha_i and angular-momentum quantum numbers ll and mm,

ξilm(r)=Nilrle−αir2Ylm(r^),(2)\xi_{ilm}(r) = N_{il}r^l e^{-\alpha_i r^2}Y_{lm}(\hat{r}), \tag*{(2)}

where NilN_{il} is the normalization constant and YlmY_{lm} is a real-valued spherical harmonic. A contracted GTO (cGTO) ξμlm\xi_{\mu lm} is formed by combining pGTOs with the same angular momentum using fixed contraction coefficients,

ξμlm(r)=∑idiμξilm(r).(3)\xi_{\mu lm}(r) = \sum_i d_{i\mu}\xi_{ilm}(r). \tag*{(3)}

A Gaussian basis set is defined by its Gaussian exponents and contraction coefficients,

p={α,d}.(4)\boldsymbol{p} = \{\boldsymbol{\alpha},\boldsymbol{d}\}. \tag*{(4)}

For given α\alpha and ll, the 2l+12l+1 pGTOs with different values of mm, i.e., {ξilm}m=12l+1\{\xi_{ilm}\}_{m=1}^{2l+1}, form a primitive shell. Similarly, {ξμlm}m=12l+1\{\xi_{\mu lm}\}_{m=1}^{2l+1} forms a contracted shell. Basis-set optimization is performed at the shell level.

Characterization of linear dependence

Linear dependence among GTOs is characterized by the eigenvalues and eigenvectors of the basis-set overlap matrix,

Skμν≡⟨χkμ∣χkν⟩=∑p=1NAOλkpUkμpUkνp∗.(5)S_{k\mu\nu} \equiv\langle\chi_{k\mu} \mid\chi_{k\nu} \rangle= \sum_{p=1}^{N_{\mathrm{AO}}}\lambda_{kp}U_{k\mu p}U_{k\nu p}^{*}. \tag*{(5)}

In particular, eigenvectors associated with small eigenvalues, empirically λ<10−6∼10−8\lambda< 10^{-6}\sim10^{-8}, represent nearly redundant linear combinations of basis functions and thus indicate linear dependence.20,26 In this work, we quantify the linear dependence of a basis set using the condition number κ\kappa and the linear-dependence weight ww.

The condition number κ\kappa, defined as the ratio of the largest to the smallest eigenvalue, is the most commonly used metric for quantifying linear dependence.13,15,31–33 Most previous studies use a kk-point definition of κ\kappa,

κk-point=max⁡k∈Kmax⁡pλkpmin⁡k∈Kmin⁡pλkp(6)\kappa_{k\text{-point}}=\frac{\max_{k\in\mathcal{K}}\max_p\lambda_{kp}}{\min_{k\in\mathcal{K}}\min_p\lambda_{kp}} \tag*{(6)}

whereas we use the supercell definition,

κsupercell=max⁡k∈Kmax⁡pλkpmin⁡k∈Kmin⁡pλkp.(7)\kappa_{\mathrm{supercell}}=\frac{\max_{k\in\mathcal{K}}\max_p\lambda_{kp}}{\min_{k\in\mathcal{K}}\min_p\lambda_{kp}}. \tag*{(7)}

Because κsupercell≥κk-point\kappa_{\mathrm{supercell}}\ge\kappa_{k\text{-point}}, the supercell definition provides a more stringent measure of linear dependence. In addition, κsupercell\kappa_{\mathrm{supercell}} remains a meaningful metric in the special case of a single basis function per unit cell, whereas κk-point=1\kappa_{k\text{-point}}=1 regardless of how diffuse the basis function is. We henceforth omit the subscript “supercell”.

Despite its simplicity and widespread use, κ\kappa reflects only the conditioning of the most singular mode, whereas basis-set linear dependence involves all singular modes. To account for these modes, we introduce the linear-dependence weight (LDW) as an alternative metric,

w(λ0)=1Nk∑k∈K∑p=1NAOf(λkp;λ0),(8)w(\lambda_0)=\frac{1}{N_k}\sum_{k\in\mathcal{K}}\sum_{p=1}^{N_{\mathrm{AO}}}f(\lambda_{kp};\lambda_0), \tag*{(8)}

where f(λ;λ0)f(\lambda;\lambda_0) is a filter function that takes appreciable values when λ≪λ0\lambda\ll\lambda_0 and vanishes when λ≫λ0\lambda\gg\lambda_0, with λ0\lambda_0 a user-selected threshold for identifying linear dependence. In practice, λ0\lambda_0 can be related to a chosen condition-number threshold κ0\kappa_0 by

λ0=max⁡k∈Kmax⁡pλkpκ0.(9)\lambda_0=\frac{\max_{k\in\mathcal{K}}\max_p\lambda_{kp}}{\kappa_0}. \tag*{(9)}

In what follows, we use the thresholds λ0\lambda_0 and κ0\kappa_0 interchangeably. The filter function ff is not unique. Here, we choose ff to be a squared-softplus function,

fsoftplus2(λ;λ0,β)={β−1ln⁡[1+(λ0λ)β]}2,(10)f_{\mathrm{softplus}^2}(\lambda;\lambda_0,\beta)=\left\{\beta^{-1}\ln\left[1+\left(\frac{\lambda_0}{\lambda}\right)^\beta\right]\right\}^2, \tag*{(10)}

with β=2\beta=2 providing a reasonably sharp transition near λ=λ0\lambda=\lambda_0. For λ≪λ0\lambda\ll\lambda_0, fsoftplus2(λ;λ0)f_{\mathrm{softplus}^2}(\lambda;\lambda_0) grows asymptotically as [ln⁡(λ0/λ)]2[\ln(\lambda_0/\lambda)]^2, thereby assigning larger weights to eigenvectors associated with more severe linear dependence.

In addition to the total LDW in eq. (8), we also define the shell-resolved LDW,

wμl(λ0)=1Nk∑k∈K∑p=1NAOf(λkp;λ0)∑m=12l+1∣Uk(μlm)p∣2,(11)w_{\mu l}(\lambda_0)=\frac{1}{N_k}\sum_{k\in\mathcal{K}}\sum_{p=1}^{N_{\mathrm{AO}}}f(\lambda_{kp};\lambda_0)\sum_{m=1}^{2l+1}\left|U_{k(\mu lm)p}\right|^2, \tag*{(11)}

which identifies the primitive or contracted shells that contribute most to the overall linear dependence of a basis set. Normalization of the overlap eigenvectors implies that the total LDW is the sum of all shell-resolved weights,

w(λ0)=∑μlwμl(λ0).(12)w(\lambda_0)=\sum_{\mu l}w_{\mu l}(\lambda_0). \tag*{(12)}

Both κ\kappa and ww exhibit finite-size effects, meaning that their values depend on the size of the kk-point mesh K\mathcal{K} or, equivalently, the supercell. Unless otherwise specified, all κ\kappa and ww values reported in this work were evaluated using a sufficiently large mesh K\mathcal{K}, with a corresponding supercell containing at least 300 atoms.

Factors affecting a basis set’s linear dependence

Before presenting our approach to controlling linear dependence, we use a few examples to illustrate the factors that affect it. Consider an elemental solid described by a minimal basis set consisting of a single primitive shell per atom with Gaussian exponent α\alpha and angular momentum ll. The linear dependence of this basis set—characterized by the metrics defined in section II B—is then fully determined by α\alpha and ll for a given solid. Figure 1 shows the condition number κ\kappa as a function of α\alpha for different values of ll in diamond, graphite, and silicon at their respective equilibrium geometries. As expected, κ\kappa increases monotonically as α\alpha decreases in all cases, illustrating that increasingly diffuse basis functions eventually lead to strong linear dependence (e.g., κ≳109\kappa\gtrsim10^9). The smallest α\alpha that can be used before encountering strong linear dependence depends strongly on angular momentum, crystal structure, and elemental species.

Figure 1. Condition number of a basis set consisting of a single primitive Gaussian shell per atom with exponent α\alpha and angular momentum ll evaluated for three elemental solids at their respective equilibrium geometries, namely diamond (top), graphite (middle), and silicon (bottom). The red dashed horizontal line indicates κ=109\kappa=10^9, an empirical threshold for strong basis-set linear dependence.15,16

  1. In general, κ\kappa is more sensitive to α\alpha for primitives with higher ll. For example, in diamond, an ss primitive with α≈0.07\alpha\approx0.07 gives a value of κ\kappa similar to that of an ff primitive with α≈0.2\alpha\approx0.2.

  1. Structural anisotropy can make certain ll channels more sensitive than others, as illustrated by graphite. Compared with diamond, the layered structure of graphite leads to a markedly different ordering of the κ\kappa–α\alpha profiles. The pp and dd profiles nearly overlap, while the ff channel exhibits substantially faster growth in κ\kappa.

  1. Finally, different elements can exhibit very different κ\kappa–α\alpha profiles even when they share the same crystal structure. Silicon, for example, has the same crystal structure as diamond but can tolerate more diffuse primitives than diamond in all ll channels. This difference can be attributed to the larger equilibrium lattice constant (5.431 Å versus 3.567 Å) and hence the greater interatomic separation in silicon, which reduces the overlap between neighboring basis functions.

An important lesson from this model study is that the notion of a “diffuse primitive” is not well defined by the Gaussian exponent alone. Angular momentum, crystal structure, and elemental species all affect whether a particular primitive should be considered diffuse in the context of linear dependence. For this reason, discarding all primitives whose exponents fall below a fixed threshold (e.g., α<0.1\alpha< 0.1 a.u.) is an imperfect remedy for basis-set linear dependence. This criterion may remove important primitives with small exponents that cause little linear dependence while retaining problematic primitives with larger exponents. Consistent with this observation, recent periodic wavefunction-based calculations of solids have used system-specific truncation thresholds to address strong linear dependence in standard cc-pVXZ basis sets.24,25

Solidcc-pVQZdef2-QZVPcc-pVQZ-s
Diamond3.4×10123.4 \times10^{12}9.9×10159.9 \times10^{15}3.5×1083.5 \times10^8
Graphite7.9×10127.9 \times10^{12}3.7×10163.7 \times10^{16}8.4×1078.4 \times10^7
Silicon1.1×10101.1 \times10^{10}1.0×10111.0 \times10^{11}1.4×1071.4 \times10^7

Table 1. Condition number κ\kappa of different Gaussian basis sets for three solids at their respective equilibrium geometries.

The model study also helps understand the linear dependence of realistic multi-shell basis sets. Table I shows the condition numbers of two standard atom-optimized basis sets, cc-pVQZ¹ and def2-QZVP,² evaluated for the three solids studied above. All κ\kappa values are significantly higher than 10910^9, indicating strong linear dependence in both basis sets for these solids. To identify the primitive or contracted shells responsible for this linear dependence, Figure 2 shows the shell-resolved LDWs wμlw_{\mu l} [eq. (11)] alongside the basis-set structures. For diamond, the most diffuse primitives in both the ss and pp channels have wμl>10w_{\mu l}>10, indicating strong contributions to near-singular overlap eigenvectors. The most diffuse dd primitive also makes a noticeable contribution despite its relatively large exponent of ca. 0.2. For graphite, structural anisotropy causes the most diffuse pp primitive to acquire a noticeably larger weight than its ss counterpart. For silicon, the ss channel contributes little to linear dependence despite having a smaller minimum exponent than that in the corresponding carbon basis, whereas the most diffuse pp primitive has a moderate wμl≈1w_{\mu l}\approx1 for cc-pVQZ and a more problematic wμl≈7w_{\mu l}\approx7 for def2-QZVP. These trends are fully consistent with those observed in Figure 1.

Figure 2. Shell-resolved linear-dependence weight [eq. (11)] for the molecular cc-pVQZ and def2-QZVP basis sets and the solid-state cc-pVQZ-s basis sets developed in this work. Results are shown for diamond (top), graphite (middle), and silicon (bottom). Circles and boxes denote primitive and contracted shells, respectively, with each contracted shell placed at the exponent of its most diffuse primitive. The magnitude of the linear-dependence weight is indicated by the color scale shown on the right. The threshold λ0\lambda_0 is determined using κ0=109\kappa_0 = 10^9.

To preview the basis sets developed in this work, table I and Figure 2 also show κ\kappa and wμlw_{\mu l} for our cc-pVQZ-s basis sets, which are derived from the standard cc-pVQZ sets through controlled reoptimization with MCAO (section III). The cc-pVQZ-s basis sets are substantially more numerically stable than cc-pVQZ, with κ<109\kappa<10^9 for all three solids and wμl<0.01w_{\mu l}<0.01 for all shells. This improved stability arises primarily from increasing the Gaussian exponents of the most diffuse ss and pp shells relative to those in cc-pVQZ, consistent with the LDW analysis. Importantly, the higher-ll polarization functions remain largely unchanged from cc-pVQZ, preserving the correlation-consistent character of the parent basis sets, as we will demonstrate numerically in section V A.

Material-constrained atomic optimization with improved numerical performance

The numerical experiments above motivate us to ask whether we can largely preserve the structure of an atom-optimized basis set, and thus its accuracy, transferability, and systematic improvability, while regularizing its linear dependence in solids. If so, across what range of solids can these regularized atom-optimized basis sets maintain both numerical stability and accuracy? We developed material-constrained atomic optimization16 (MCAO) to address the first question. We address the second through systematic numerical validation and benchmarking in section V.

MCAO determines the Gaussian exponents α\boldsymbol{\alpha} of a basis set by minimizing the objective function

L(α;{M},ε0,κ0)=Eatom(α)+ε0P(α;{M},κ0),(13)L(\boldsymbol{\alpha}; \{\mathcal{M}\}, \varepsilon_0, \kappa_0) = E_{\mathrm{atom}}(\boldsymbol{\alpha}) + \varepsilon_0 P(\boldsymbol{\alpha}; \{\mathcal{M}\}, \kappa_0), \tag*{(13)}

where EatomE_{\mathrm{atom}} is the atomic energy and PP is a dimensionless linear-dependence penalty evaluated over a set of reference solids {M}\{\mathcal{M}\}. The strength and onset of the penalty are controlled by two parameters, ε0\varepsilon_0 and κ0\kappa_0. ε0\varepsilon_0 has units of energy and sets the energy scale of the penalty, whereas κ0\kappa_0 acts approximately as a target condition number and serves as the primary parameter for tuning the conditioning of the optimized basis sets. The key hypothesis behind MCAO is that optimizing EatomE_{\mathrm{atom}}, as in conventional atomic optimization, helps retain the overall structure of atom-optimized basis sets while reducing sensitivity to the choice of reference solids. MCAO thus provides a middle ground between conventional atomic optimization,[1, 2, 3] which minimizes only EatomE_{\mathrm{atom}}, and solid-specific optimizations,31–34 which minimize both EE and κ\kappa for solids.

In our previous application of MCAO to Cu basis-set optimization,16 we considered the simple case of a single reference solid with a κ\kappa-based penalty,

Pκ(α;M,κ0)=κ(α;M)κ0.(14)P_{\kappa}(\boldsymbol{\alpha};M,\kappa_0) = \frac{\kappa(\boldsymbol{\alpha};M)}{\kappa_0}. \tag*{(14)}

In this work, we introduce two extensions to the original MCAO algorithm to improve its numerical performance.

First, we define a revised penalty based on the LDW eq. (8) that naturally accommodates multiple reference solids,

Pw(α;{M},κ0)=∑Mw(α;M,κ0),(15)P_w(\boldsymbol{\alpha};\{M\},\kappa_0) = \sum_M w(\boldsymbol{\alpha};M,\kappa_0), \tag*{(15)}

where ww is evaluated using the squared-softplus filter function defined in eq. (10). We find that PwP_w substantially improves the stability and convergence rate of the optimization by smoothly incorporating all ill-conditioned overlap eigenvalues. By contrast, PκP_{\kappa} is determined by the extremal eigenvalues, whose identities may change between optimization cycles, introducing nonsmoothness into the objective function.

Second, if the initial basis set is extremely ill-conditioned for the chosen reference solids, evaluating the penalty in standard double-precision floating-point arithmetic can introduce substantial numerical errors and destabilize the optimization. In our previous work,16 we mitigated this issue by manually restricting the valence space of the initial basis set so that its condition number was reasonably close to the target value κ0\kappa_0. Here, we introduce a lattice-scaling procedure that allows MCAO to start from basis sets with essentially arbitrary conditioning. Specifically, we determine an initial scaling factor γ≥1\gamma\ge1 that uniformly expands the lattice constants of all reference solids so that the largest κ\kappa of the initial basis set across the reference solids is slightly below min⁡(κ0,1010)\min(\kappa_0,10^{10}). Starting from these expanded geometries, we decrease γ\gamma toward 1 in steps of 0.03 until the original reference solids are recovered. At each value of γ\gamma, we perform a complete MCAO optimization using the corresponding γ\gamma-scaled reference solids, with the optimized basis set from the preceding value of γ\gamma serving as the initial guess. This procedure keeps the basis set within a numerically well-conditioned regime throughout the optimization and avoids instabilities associated with an excessively ill-conditioned starting point.

The cc-pVXZ-s basis sets for H and p-block elements from B to Cl

Overall design principle

Our design of the cc-pVXZ-s basis sets for X=D,T,X = \mathrm{D}, \mathrm{T}, and Q\mathrm{Q} closely follows Dunning’s construction of the cc-pVXZ series.[1] We make slight modifications to the structure of the lower-zeta basis sets, as described below, to improve their accuracy while preserving their correlation consistency. The selected elements—H and pp-block elements from B to Cl—provide an excellent test set for investigating the trade-off between basis-set accuracy and numerical stability, while avoiding complications from (semi)core electrons,[7] multireference character,49 and relativistic effects[8] associated with other elements. For each element, we first generate atom-optimized basis sets through the QZ level (§ 3.1.3 B), evaluate their linear dependence in the reference solids (§ 3.1.3 C), and apply MCAO to restore numerical stability when strong linear dependence is found (§ 3.1.3 D). The same protocol is applied to generate cc-pVXZ-s basis sets suitable for calculations using different nuclear-potential treatments (§ 3.1.3 E). Compact auxiliary basis sets optimized for density fitting50,51 are also developed (§ 3.1.3 F).

Atomic optimization

For each element, we first construct atom-optimized cc-pVXZ-s basis sets using the following steps.

  1. A valence set consisting of low-ll primitives (ss for H and ss and pp for pp-block elements) is optimized by minimizing the atomic restricted open-shell HF (ROHF) energy,

αval∗←min⁡αvalEHF(αval).(16)\boldsymbol{\alpha}_{\mathrm{val}}^{*} \leftarrow\min_{\boldsymbol{\alpha}_{\mathrm{val}}} E_{\mathrm{HF}}(\boldsymbol{\alpha}_{\mathrm{val}}). \tag*{(16)}
  1. With the optimized valence primitives held fixed, a polarization set consisting of high-ll primitives (l≥1l \ge1 for H and l≥2l \ge2 for pp-block elements) is optimized by minimizing the frozen-core atomic coupled-cluster singles and doubles52 (CCSD) energy,

αpol∗←min⁡αpolECCSD(αval∗,αpol).(17)\boldsymbol{\alpha}_{\mathrm{pol}}^{*} \leftarrow\min_{\boldsymbol{\alpha}_{\mathrm{pol}}} E_{\mathrm{CCSD}}(\boldsymbol{\alpha}_{\mathrm{val}}^{*},\boldsymbol{\alpha}_{\mathrm{pol}}). \tag*{(17)}

For H, the atomic CCSD correlation energy is zero, so we instead minimize the CCSD correlation energy of H2_2 at a bond length of 0.7414 Å.15

  1. The optimized valence primitives from step 1 are contracted using ROHF atomic natural orbital53 (ANO) coefficients and then combined with the optimized polarization functions from step 2. Optionally, the polarization primitives are contracted using frozen-core CCSD ANO coefficients (vide infra).

  1. Additional valence-space polarization primitives are optimized in the presence of the contracted valence and polarization shells to further increase basis-set flexibility.

This general protocol requires several choices. The first is the size of the valence set, i.e., the number of primitives in each valence ll channel. In conventional cc-pVXZ, the valence set grows with zeta level, while its size remains fixed across each period at a given zeta level. This leads to two issues in periodic solids. First, valence sets for elements early in a period (e.g., groups 13 and 14) contain very diffuse primitives at TZ and QZ levels, resulting in strong linear dependence in solids containing these elements. Second, the HF energy error increases with atomic number in each period, making basis sets for later elements (e.g., groups 15–17) systematically less accurate than those for earlier elements. To address these issues, we use a sufficiently large valence set at all zeta levels, subject to the linear-dependence constraints described in § 3.1.3 D. Specifically, we use valence sets of QZ quality for group 15–17 elements, TZ quality for group 14 elements and H, and DZ quality for group 13 elements. The precise choice is less critical because MCAO automatically adjusts the most

diffuse valence exponents in each ll channel during optimization.

The second choice is the size of the polarization set. Conventional cc-pVXZ uses ldld, 2l d1f2l\,d1f, and 3d2f1g3d2f1g primitives for pp-block elements and 1p1p, 2p1d2p1d, and 3p2d1f3p2d1f for H at the DZ, TZ, and QZ levels, respectively, without contraction. This hierarchy defines the correlation-consistent structure[1] and provides the foundation for the systematic improvability of conventional cc-pVXZ.11 However, large basis-set errors at lower zeta levels have been reported for molecules containing third-period elements, which require additional compact dd primitives, as in the cc-pV(X+dX+d)Z basis sets.54 We find a similar issue in our periodic benchmarks. With the conventional polarization structures, solids containing third-period elements can exhibit larger errors than their second-period counterparts at DZ and sometimes TZ. We therefore adopt a (3d)→[1d](3d) \rightarrow[1d] structure at DZ and (3d1f)→[2d1f](3d1f) \rightarrow[2d1f] at TZ for all pp-block elements, while retaining the QZ structure. Here, [ml][ml] denotes contraction of nn ll-type primitive shells into mm ll-type contracted shells. For H, the conventional structures, 1p1p for DZ and 2p1d2p1d for TZ, are retained because there are no CCSD ANOs for a one-electron system. The QZ polarization structure is retained throughout, and the DZ and TZ contraction patterns preserve the overall basis-set size relative to cc-pVXZ. As we demonstrate numerically in section V A, the revised polarization structure improves the accuracy of cc-pVXZ-s at DZ and TZ while preserving the reliable CBS extrapolation of conventional cc-pVXZ.

Finally, for the additional valence-space polarization functions, we follow Dunning’s original recipe and add one and two primitives per valence ll channel at DZ and TZ levels, respectively. To avoid additional linear dependence, we decontract an existing valence primitive instead of adding a new function when its exponent is close to that of an optimized valence-space polarization function. We define closeness by an exponent ratio in [γ−1,γ][\gamma^{-1},\gamma] with γ=1.3\gamma= 1.3. In addition, if a valence-space polarization function is more diffuse than the most diffuse primitive in a given ll channel, the latter is decontracted to avoid introducing a more diffuse primitive. At QZ, we directly decontract the three most diffuse valence primitives from the contracted valence shells.

Reference solids

The choice of reference solids for evaluating linear dependence and performing MCAO directly controls the range of systems to which the resulting basis sets can be applied. In this work, we adopt the set of solids constructed by Bosoni and co-workers30 for a recent large-scale DFT benchmark study (henceforth referred to as the Bosoni set). For each element, the Bosoni set contains 10 solids spanning a range of coordination environments and oxidation states, as shown in Figure 3. These comprise 4 elemental solids in simple cubic (SC), body-centered cubic (BCC), face-centered cubic (FCC), and diamond phases, and 6 oxides with different stoichiometries, X2_2O, XO, X2_2O3_3, XO2_2, X2_2O5_5, and XO3_3. For many elements, the Bosoni set covers thermodynamically stable phases at ambient conditions (e.g., diamond for C and Si, FCC for Al), and also includes highly compact unconventional phases that cause strong linear dependence in atom-optimized basis sets. The latter are particularly useful for developing broadly applicable basis sets. Other thermodynamically stable solids lie outside the Bosoni set (e.g., binary solids formed by group 13 and 15 elements, such as BN and AlN). For the elements considered here, we will show that the cc-pVXZ-s basis sets optimized using the Bosoni set are also well-conditioned for these more conventional phases. Nonetheless, the Bosoni set can also be augmented with additional solids when broader coverage is needed.

Crystal structures of the Bosoni reference solids

Figure 3. Crystal structures of the Bosoni reference solids30 used for MCAO, comprising 4 elemental solids with simple cubic (SC), body-centered cubic (BCC), face-centered cubic (FCC), and diamond structures, and 6 oxides with increasing oxygen-to-element ratios, X2_2O, XO, X2_2O3_3, XO2_2, X2_2O5_5, and XO3_3. Conventional cubic cells are shown for BCC, FCC, diamond, X2_2O, XO, and XO2_2, while the calculations use the corresponding primitive cells. Yellow and red denote X and O, respectively.

To assess the linear dependence of atom-optimized basis sets, Figure 4 shows the largest condition number κmax⁡\kappa_{\max} over the 10 reference solids for each element using standard cc-pVXZ basis sets. Strong linear dependence is observed at TZ and QZ for H and group 13 and 14 elements, and at QZ for group 15 elements. These basis sets are natural targets for MCAO. Basis sets for elements in group 16 and higher remain largely well-conditioned even at QZ, consistent with their less diffuse primitives arising from larger effective nuclear charges. In particular, atom-optimized O basis sets remain numerically stable through QZ. This simplifies MCAO with the Bosoni set because the penalty for an oxide can be evaluated while keep-

Maximum condition number over all 10 reference solids

Figure 4. Maximum condition number (κmax⁡\kappa_{\max}) over all 10 reference solids of each element evaluated with the conventional cc-pVXZ basis sets and the cc-pVXZ-s sets developed in this work. The cc-pVXZ-s sets have κmax⁡<109\kappa_{\max} < 10^9 for all elements at all zeta levels, as indicated by the red horizontal line.

ing the atom-optimized O basis set fixed.

Figs. S1 and S2 show κmax⁡\kappa_{\max} over the Bosoni set for two other atom-optimized basis sets, the def2-SVP/TZVP/QZVP series[2, 3] and the ccECP-cc-pVXZ series optimized for the ccECP55. Both series show patterns similar to those in Figure 4, demonstrating that linear dependence is a general issue for atom-optimized basis sets across basis structures and nuclear potentials. Fig. S3 also shows strong linear dependence for the GTH-cc-pVXZ sets15—which were designed to be numerically stable for a manually selected set of thermodynamically stable solids—for H and group 13–15 elements at TZ and QZ. The comparison therefore highlights the broader range of environments covered by the Bosoni reference solids.

MCAO

We perform MCAO to improve the numerical stability of all atom-optimized basis sets from section III B with κmax⁡>107\kappa_{\max} > 10^7 on the Bosoni reference solids. In particular, MCAO is used to regularize the valence and polarization primitives, i.e., steps 1 and 2 in section III B. The subsequent contraction of primitives (step 3) and determination of valence-space polarization primitives (step 4) use only atomic calculations, and neither introduces additional linear dependence. In conventional atomic optimization, valence and polarization primitives are determined separately [eqs. (16) and (17)], whereas MCAO couples these optimizations through the penalty term. We therefore alternate between valence and polarization optimization until convergence,

αval(n+1)←min⁡αvalEHF(αval)+ε0valPw(αval,αpol(n))(18)\boldsymbol{\alpha}_{\mathrm{val}}^{(n+1)} \leftarrow\min_{\boldsymbol{\alpha}_{\mathrm{val}}} E_{\mathrm{HF}}(\boldsymbol{\alpha}_{\mathrm{val}}) + \varepsilon_{0}^{\mathrm{val}} P_w(\boldsymbol{\alpha}_{\mathrm{val}},\boldsymbol{\alpha}_{\mathrm{pol}}^{(n)}) \tag*{(18)}
αpol(n+1)←min⁡αpolECCSD(αval(n+1),αpol)+ε0polPw(αval(n+1),αpol)(19)\boldsymbol{\alpha}_{\mathrm{pol}}^{(n+1)} \leftarrow\min_{\boldsymbol{\alpha}_{\mathrm{pol}}} E_{\mathrm{CCSD}}(\boldsymbol{\alpha}_{\mathrm{val}}^{(n+1)},\boldsymbol{\alpha}_{\mathrm{pol}}) + \varepsilon_{0}^{\mathrm{pol}} P_w(\boldsymbol{\alpha}_{\mathrm{val}}^{(n+1)},\boldsymbol{\alpha}_{\mathrm{pol}}) \tag*{(19)}

where nn labels the MCAO macrocycle, and the dependence of PwP_w on κ0\kappa_0 and {M}\{M\} is omitted for brevity. The basis parameters typically converge within 3 macrocycles,16 and we use a conservative choice of 5 cycles for all basis sets reported here.

The energy scaling parameter ε0\varepsilon_0 is set to 10 mEhmE_{\mathrm{h}} and 1 mEhmE_{\mathrm{h}} for the valence and polarization primitives, respectively. The weaker penalty on the polarization set helps retain the basis set’s correlation-consistent structure.[1] [16] The linear-dependence penalty PwP_w [eq. (15)] is evaluated only for the subset of Bosoni reference solids satisfying κ≥0.01κmax⁡\kappa\ge0.01\kappa_{\max}, with condition numbers computed using the initial atom-optimized basis sets. This subset remains fixed throughout the optimization. For most elements, this criterion selects one (typically the simple cubic phase) or a few elemental solids, reducing the cost of penalty evaluation. For selected cases, we verified that the optimized basis sets are identical to those obtained from the more expensive MCAO with the full Bosoni set.

We use the optimization of all-electron carbon cc-pVXZ-s basis sets to illustrate practical aspects of MCAO and the choice of κ0\kappa_0. This parameter controls the strength and onset of the linear-dependence penalty and balances energy accuracy against numerical conditioning. We vary κ0\kappa_0 from 101010^{10} for a weak penalty to 10710^7 for a strong penalty. We monitor linear dependence through κmax⁡\kappa_{\max} over the Bosoni reference solids and energy accuracy through atomic HF energy and CCSD correlation-energy errors relative to cc-pV5Z reference calculations. In all cases, MCAO starts from atom-optimized cc-pVXZ-s basis sets generated using the protocol in § 3.1.3 B. The results are summarized in Figure 5.

Variation of the highest condition number and energy errors over the Bosoni solids for carbon cc-pVXZ-s basis sets

Figure 5. Variation of the highest condition number κmax⁡\kappa_{\max} over the Bosoni solids (top), the atomic ROHF energy error (middle), and the atomic CCSD correlation energy error (bottom) for carbon cc-pVXZ-s basis sets optimized using MCAO with different κ0\kappa_0. The energy errors are relative to reference values obtained with the standard cc-pV5Z basis set. Results for the initial atom-optimized basis sets are denoted by κ0=∞\kappa_0 = \infty. In the top panel, the red shaded area indicates strong linear dependence with κmax⁡≥109\kappa_{\max} \ge10^9. The final cc-pVDZ/TZ/QZ-s basis sets selected by κ0=3×107\kappa_0 = 3 \times10^7, 10810^8, and 3×1083 \times10^8, respectively, are highlighted by stars and vertical shaded bands in all panels.

The top panel shows that MCAO effectively tunes numerical conditioning. As κ0\kappa_0 is tightened from 101010^{10} to 10710^7, the optimized κmax⁡\kappa_{\max} decreases smoothly and closely follows κ0\kappa_0. All three zeta levels exhibit nearly identical κmax⁡\kappa_{\max}–κ0\kappa_0 profiles despite their very different initial conditioning, indicating that MCAO simultaneously adjusts both the valence and polarization exponents.

Improved numerical conditioning comes at the cost of larger HF energy and CCSD correlation-energy errors, as shown in the lower two panels of Figure 5. At a given κ0\kappa_0, the increase in both errors relative to unconstrained atomic optimization (denoted by κ0=∞\kappa_0 = \infty in the figure) grows with zeta level, because the larger polarization sets at higher zeta levels require more substantial exponent adjustments to achieve comparable numerical conditioning. To maintain consistent increases in energy error across zeta levels, we use κ0=3×107\kappa_0 = 3 \times10^7, 10810^8, and 3×1083 \times10^8 to select the final cc-pVXZ-s basis sets at DZ, TZ, and QZ, respectively, as highlighted by stars in Figure 5. With this choice, the increase in error relative to atom-optimized basis sets is consistent across zeta levels, ca. 1.5 mEhmE_{\mathrm{h}} for HF energies and 1 mEhmE_{\mathrm{h}} for CCSD correlation energies. The latter preserves the hierarchy of correlation-energy accuracy found in atom-optimized cc-pVXZ-s.

The Gaussian exponents of the final all-electron cc-pVQZ-s basis set of carbon are shown in Figure 6. Relative to conventional cc-pVQZ, MCAO increases the valence primitive exponents and effectively reduces the valence set to between DZ and TZ quality. The smallest exponent in the dd polarization channel increases slightly, while the remaining polarization exponents are largely unchanged. These modifications are fully consistent with the ill-conditioned primitives identified by the LDW analysis in section II B, validating the effectiveness of MCAO to identify and regularize these problematic primitives.

Figure 6. Gaussian exponents of carbon cc-pVQZ-s basis sets optimized for different nuclear potentials, including the all-electron potential, the GTH-HF and GTH-PBE pseudopotentials, and the ccECP. Dunning’s atom-optimized all-electron cc-pVQZ basis set is also included for comparison. Circles denote primitive shells. Squares mark the smallest exponent in each contracted shell. Orange and blue denote valence and polarization channels, respectively.

The same choices of κ0\kappa_0 are used to select the final cc-pVXZ-s basis sets for all other elements to which MCAO is applied. With these choices, the final all-electron cc-pVXZ-s basis sets have κmax⁡<109\kappa_{\max} < 10^9 for all elements considered here at all zeta levels, as shown in Figure 4. This improved numerical stability across a range of solids enables reliable calculations with cc-pVXZ-s up to QZ and extrapolation to the CBS limit, as demonstrated numerically in section V.

Pseudopotentials and effective core potentials

In addition to all-electron basis sets, we use the same protocol to generate cc-pVXZ-s basis sets for the GTH pseudopotentials optimized for HF calculations (GTH-HF56) and DFT calculations with the Perdew–Burke–Ernzerhof57 (PBE) exchange-correlation functional (GTH-PBE58) and the correlation-consistent effective core potential55 (ccECP) optimized for correlated wavefunction calculations. MCAO is largely insensitive to the choice of nuclear potential because differences between potentials primarily affect compact primitives in the core region. The cc-pVXZ-s basis sets optimized for all four nuclear potentials therefore have nearly identical valence exponents, as illustrated for carbon cc-pVQZ-s in Figure 6. They also exhibit similar numerical stability across the Bosoni reference solids, as shown in figs. S2 and S3. Consequently, calculations using cc-pVXZ-s with different nuclear potentials are expected to exhibit very similar basis-set convergence, as demonstrated numerically in section V A 4.

This unique feature allows different aspects of cc-pVXZ-s to be benchmarked against suitable references. For example, the accuracy of cc-pVXZ-s in mean-field calculations can be assessed using the basis sets optimized for the GTH pseudopotentials, for which a sufficiently large PW basis provides a CBS reference. The accuracy of cc-pVXZ-s in correlated wavefunction calculations can be assessed using the all-electron basis sets and comparing the results with conventional cc-pVXZ calculations. We use these comparisons to validate the cc-pVXZ-s basis sets in section V A.

Auxiliary and MINAO basis sets

To enable efficient periodic calculations using density fitting50,51 (DF), we also developed optimized auxiliary basis sets for cc-pVXZ-s. The details of the optimization are provided in the Supporting Information. These auxiliary basis sets are termed cc-pVXZ-s-FIT and are designed for both mean-field and correlated wavefunction calculations. The average size ratio of cc-pVXZ-s-FIT to cc-pVXZ-s is approximately 8, 5, and 3 for the all-electron DZ, TZ, and QZ sets and 4, 3, and 2 for the GTH- and ccECP-based DZ, TZ, and QZ sets, respectively. For all-electron calculations, cc-pVXZ-s-FIT is therefore slightly larger than the commonly used cc-pVXZ-JKFIT59 sets optimized for cc-pVXZ (6, 3, and 2 for DZ, TZ, and QZ, respectively), especially at larger zeta levels. This increase mainly reflects the enlarged valence and dd polarization primitive sets in cc-pVXZ-s. The cc-pVXZ-s-FIT sets are typically much smaller than automatically generated auxiliary basis sets, such as the even-tempered basis (ETB) used in PySCF60–62 and the AutoAux basis63 used by ORCA.64

Finally, we also generate the minimal atomic orbital (MINAO) basis sets corresponding to cc-pVXZ-s by fully contracting the atom-optimized valence primitives before MCAO. The MINAO basis sets can serve as projection bases for population analysis and the generation of intrinsic atomic orbitals65 (IAOs). These IAOs serve as the localized-orbital basis for the periodic local coupled-cluster calculations in section V B.

Computational details

Basis-set optimization

Atomic optimization and MCAO are implemented in the open-source PyGTO code,66 which uses PySCF60–62 for electronic structure calculations. The Gaussian exponents are optimized in logarithmic space using the Broyden–Fletcher–Goldfarb–Shanno (BFGS) algorithm67 with numerical gradients. An optimization is considered converged when changes in the objective function and logarithmic Gaussian exponents fall below 10−9Eh10^{-9} E_h and 10−310^{-3}, respectively. For some elements treated with pseudopotentials or ECPs, HF energy minimization can cause two exponents to become nearly equal, leading to numerical instability. We avoid this variational collapse68 by penalizing the ratio of each pair of adjacent exponents α1<α2\alpha_1 < \alpha_2.

Pexp−rat(α1,α2)=εexp−ratfsoftplus2(α2/α1;rmin⁡,β),(20)P_{\mathrm{exp-rat}}(\alpha_1,\alpha_2)=\varepsilon_{\mathrm{exp-rat}}f_{\mathrm{softplus}}^2(\alpha_2/\alpha_1;r_{\min},\beta), \tag*{(20)}

where εexp−rat=10Eh\varepsilon_{\mathrm{exp-rat}} = 10 E_h, fsoftplus2f_{\mathrm{softplus}}^2 is the squared-softplus function defined in eq. (10), rmin⁡=1.5r_{\min}=1.5 is the minimum allowed ratio, and β=100\beta=100 tightly constrain the exponent ratio to r≥rmin⁡r \ge r_{\min}. The contribution of each reference point is evaluated from the overlap-matrix eigenvalues for the smallest supercell containing more than 300 atoms. For cubic Bosoni reference solids, the supercells are constructed by repeating the unit cell equally along all three lattice directions.

Test sets

We benchmark different aspects of the cc-pVXZ-s basis sets using the following tests.

  1. Atomization energies of 19 molecular oxides, AlO (2), BO (2), ClO (2), ClO2_2 (2), CO (1), CO2_2 (1), FO (2), H2_2O (1), H2_2O2_2 (1), NO (2), NO2_2 (2), O2_2 (3), O3_3 (1), PO (2), PO2_2 (2), SiO (1), SiO2_2 (1), SO2_2 (1), and SO3_3 (1), with spin multiplicities 2S+12S+1 shown in parentheses.

  1. Equations of state (EOS) for 110 Bosoni solids, comprising 4 elemental solids (simple cubic, FCC, BCC, and diamond) and 6 oxides (X2_2O, XO, X2_2O3_3, XO2_2, X2_2O5_5, and XO3_3) for H, B, C, N, O, F, Al, Si, P, S, and Cl.

  1. Band structures and EOS for 7 semiconductors, BN (B3), BP (B3), AlN (B3), AlP (B3), C (A4), Si (A4), and SiC (B3), with Strukturbericht designations shown in parentheses. The band-structure tests also include FCC Al as a representative bulk metal.

Structures of all test systems are provided in the Supporting Information.

General computational details

We use the PySCF code for all Gaussian-basis calculations. Electron-repulsion integrals (ERIs) are evaluated using DF with auxiliary basis sets adapted to each calculation. For periodic PBE calculations, we use fast Fourier transform density fitting69 (FFTDF) with a sufficiently large PW fitting basis to make DF errors negligible. This enables direct comparison of our Gaussian-basis PBE results with PW references calculated in Quantum ESPRESSO70,71 using the same GTH-PBE pseudopotentials.

For HF and correlated wavefunction calculations, we use the range-separated density fitting72,73 (RSDF) with optimized Gaussian auxiliary basis sets generated by the algorithm developed in this work. The only exception is molecular atomization-energy calculations with the all-electron cc-pVXZ basis sets, for which we use the standard cc-pVXZ-JKFIT auxiliary basis sets.59 The standard auxiliary basis sets cause numerical problems in periodic calculations because of strong linear dependence in the DF metric.24,74 Our reoptimized auxiliary basis sets resolve these problems. Canonical orthogonalization is not used in any periodic calculations.

Results obtained with finite Gaussian basis sets are extrapolated to the CBS limit using standard protocols from the literature. We extrapolate HF energies using the exponential formula,75

EHF(X)=EHF(CBS)+Aexp⁡(−βX),(21)E_{\mathrm{HF}}(X) = E_{\mathrm{HF}}(\mathrm{CBS}) + A \exp(-\beta X), \tag*{(21)}

where β=1.54\beta= 1.54 for (D,T) and (T,Q) extrapolations and 1.95 for (Q,5) extrapolations. Correlation energies are extrapolated using the X−3X^{-3} formula,76,77

Ecorr(X)=Ecorr(CBS)+AX−3.(22)E_{\mathrm{corr}}(X) = E_{\mathrm{corr}}(\mathrm{CBS}) + A X^{-3}. \tag*{(22)}

In correlated calculations, we freeze the He and Ne cores for elements in the second and third periods, respectively.

All EOS calculations use 7 cell volumes uniformly spanning [0.94, 1.06] times the reference volume V0V_0. For the Bosoni set, the reference volumes are taken from the PBE-optimized structures in ref 30, while experimental volumes compiled by Schmika and co-workers78 are used for the 7 semiconductors. The resulting energy curves for wavefunction calculations and free-energy curves for PBE calculations (vide infra) are fitted to the Birch–Murnaghan EOS to extract equilibrium lattice constants a0a_0 and bulk moduli B0B_0. The cohesive energy EcohE_{\mathrm{coh}} is calculated at the reference volume V0V_0, with atomic reference energies corrected for basis-set superposition error79 (BSSE) using two shells of ghost atoms.

We use box plots to visualize error distributions. Each box spans the middle 50% of the data (i.e., the interquartile range), and the horizontal line inside the box marks the median. The whiskers extend to the most extreme points within 1.5 times the interquartile range of the corresponding quartile. Data points beyond the whiskers are shown individually as outliers.

PBE calculations

For the Bosoni set, EOS calculations use spin-restricted PBE with the GTH-PBE pseudopotentials. Following ref 30, we use Fermi smearing with a width of 0.045 Ry = 0.0225 EhE_{\mathrm{h}} and fit the EOS using the free energy F=E−TSF = E - TS, where SS is the electronic entropy. The Brillouin zone is sampled using uniform kk-point meshes of size 4×4×44 \times4 \times4 for BCC, FCC, and simple cubic lattices, 3×3×33 \times3 \times3 for diamond and XO lattices, and 2×2×22 \times2 \times2 for X2_2O, XO2_2, XO3_3, X2_2O3_3, and X2_2O5_5 lattices, respectively. We reduce finite-size errors by shifting the kk-point meshes by the Baldereschi mean-value point,80,81 (0.1477, 0.3112, 0.4588) for FCC, diamond, X2_2O, XO, and XO2_2 lattices, (0.25, 0.25, 0.5833) for the BCC lattice, and (0.25, 0.25, 0.25) for SC, X2_2O5_5, and XO3_3 lattices (all expressed in units of the supercell reciprocal lattice vectors).

The band structures of the 7 semiconductors and aluminum are also calculated using spin-restricted PBE with the GTH-PBE pseudopotentials. SCF calculations use an unshifted 4×4×44 \times4 \times4 kk-point mesh, followed by non-SCF calculations along the selected band path. Calculations for the 7 semiconductors use zero temperature without smearing, while aluminum calculations use the same smearing setup as the Bosoni set.

Wavefunction-based calculations

We calculate the EOS of the 7 semiconductors at zero temperature using HF, MP2, SCS-MP2,24,82 CCSD, and CCSD(T). The integrable divergence in the HF exchange energy is treated using the probe-charge Ewald method.83,84 This shifts the occupied HF orbital energies by the supercell Madelung constant vMv_{\mathrm{M}} and the HF total energy by NvM/2Nv_{\mathrm{M}}/2, where NN is the number of electrons per unit cell. The (SCS-)MP2 correlation energies and the (T) contribution to CCSD(T) use Madelung-corrected HF orbital energies, while CCSD calculations use uncorrected HF orbital energies. With these choices, all energy components have a leading-order finite-size scaling of O(Nk−1)O(N_k^{-1}).12,85 For the basis-set convergence tests in section V A 4, we use an unshifted 3×3×33 \times3 \times3 kk-point mesh for all 7 semiconductors.

For the comparison with experiment in section V B, we first extrapolate the calculated energy–volume curves to the combined basis-set and thermodynamic limits (TDL), then fit them to the EOS to extract a0a_0 and B0B_0. For HF and (SCS)-MP2, we use the composite correction

E(CBS,TDL)=E(CBS,3×3×3)+E(TZ,TDL)−E(TZ,3×3×3),(23)E(\mathrm{CBS},\mathrm{TDL}) = E(\mathrm{CBS},3 \times3 \times3) + E(\mathrm{TZ},\mathrm{TDL}) - E(\mathrm{TZ},3 \times3 \times3), \tag*{(23)}

where CBS extrapolation on a 3×3×33 \times3 \times3 kk-point mesh uses TZ and QZ basis sets. The TZ results on n×n×nn \times n \times n kk-point meshes with n=3,4n = 3,4 and 5 are fitted to the asymptotic form24

E(Nk)=E(TDL)+A1Nk−1+A2Nk−2.(24)E(N_k) = E(\mathrm{TDL}) + A_1N_k^{-1} + A_2N_k^{-2}. \tag*{(24)}

For EcohE_{\mathrm{coh}}, we refine the TDL correction in eq. (23) using n=4,5n = 4,5 and 6.

For CCSD and CCSD(T), we use the local natural orbital (LNO) approximation40,86,87 with systematically tightened truncation thresholds to obtain accurate estimates of canonical results with TZ basis sets and a 3×3×33 \times3 \times3 kk-point mesh. The atomic CCSD and CCSD(T) reference energies used to evaluate EcohE_{\mathrm{coh}} are calculated within the frozen natural orbital (FNO) approximation88,89 using one shell of ghost atoms. We then apply an MP2-based composite correction for the change from one to two shells of ghost atoms. The FNO calculations use a tight percentage occupation threshold90 of 99.995% to select the natural orbitals. The estimated errors arising from the LNO approximation are below 0.003 Å for a0a_0, 0.5 GPa for B0B_0, and 0.02 eV for EcohE_{\mathrm{coh}}, sufficient for the present benchmarks. Detailed convergence studies are provided in the Supporting Information. We treat the remaining basis-set and finite-size errors in the CC correlation energy approximately using a scaled MP2 correction,

ECC(CBS,TDL)≈ECC(TZ,3×3×3)+α[EMP2(CBS,TDL)−EMP2(TZ,3×3×3)],(25)E_{\mathrm{CC}}(\mathrm{CBS},\mathrm{TDL}) \approx E_{\mathrm{CC}}(\mathrm{TZ},3 \times3 \times3) + \alpha\left[E_{\mathrm{MP2}}(\mathrm{CBS},\mathrm{TDL}) - E_{\mathrm{MP2}}(\mathrm{TZ},3 \times3 \times3)\right], \tag*{(25)}

where EMP2(CBS,TDL)E_{\mathrm{MP2}}(\mathrm{CBS},\mathrm{TDL}) is obtained using eq. (23). The scaling factor α\alpha accounts for differences in the basis-set and finite-size convergence of CC and MP2 correlation energies. We estimate α\alpha as 0.7 using canonical MP2 and CCSD correlation energies reported in ref 25 for four bulk semiconductors, BN, BP, Si, and C, at their equilibrium geometries. We apply this fitted α\alpha to all 7 semiconductors and all cell volumes considered here. We denote these CC results as L-CC+Δ\DeltaMP2 to emphasize the approximate treatment of basis-set and finite-size errors.

Results and discussion

Validation of cc-pVXZ-s basis sets
Molecular atomization energies

We start by benchmarking cc-pVXZ-s against standard cc-pVXZ in molecular calculations. Our aim is to assess whether cc-pVXZ-s retains the accuracy and reliable CBS extrapolation of standard cc-pVXZ in both mean-field and correlated wavefunction calculations. We select atomization energies because they directly probe the total energy recovered by a basis set, with less benefit from error cancellation.

Figure 7 shows the basis-set errors in HF and MP2 atomization energies for 19 molecular oxides calculated at DZ, TZ, and QZ, as well as CBS extrapolations using (D,T) and (T,Q) for both cc-pVXZ-s and standard cc-pVXZ. The reference values are obtained by CBS(Q,5) extrapolation with the standard basis sets. Overall, cc-pVXZ-s converges rapidly with zeta level and extrapolates to the CBS reference values at both the HF and MP2 levels. At the HF level, cc-pVXZ-s substantially outperforms standard cc-pVXZ at DZ and, to a lesser extent, at TZ. The improvement is greatest for third-period elements and can be attributed to the expanded dd-polarization primitive set in cc-pVXZ-s, which serves a role similar to the additional dd primitives in cc-pV(X+dX +d)Z.54 At the MP2 level, cc-pVXZ-s closely follows the rapid convergence of standard cc-pVXZ, yielding reasonable accuracy with CBS(D,T) extrapolation and chemical accuracy with CBS(T,Q). Taken together, these results suggest that cc-pVXZ-s largely retains the systematic convergence of standard cc-pVXZ despite being regularized to improve numerical conditioning.

Figure 7. Box plots of basis-set errors in atomization energies of 19 molecular oxides calculated at the HF and MP2 levels using cc-pVXZ and cc-pVXZ-s for X=DX = \mathrm{D}, T, and Q and CBS extrapolations using (D,T) and (T,Q). The panels show HF, MP2 correlation, and MP2 total atomization-energy errors from top to bottom. The errors are relative to CBS(Q,5) reference values obtained using cc-pVQZ and cc-pV5Z and are normalized by the number of atoms in each molecule.

PBE equation of state

Turning to periodic solids, we first use the Bosoni set to validate the mean-field accuracy of cc-pVXZ-s. The diverse lattice structures and chemical environments covered by the 110 solids in the Bosoni set make this set an excellent testbed for probing the accuracy, transferability, and numerical stability of our basis sets. We perform PBE EOS calculations using cc-pVXZ-s with the GTH-PBE pseudopotentials. We compare the resulting lattice constants a0a_0, bulk moduli B0B_0, and free energies per atom F0F_0 at the equilibrium volume with reference results obtained using a converged PW basis. The basis-set errors are summarized in Figure 8. A breakdown of these errors by element is shown in figs. S4–S14.

Figure 8. Box plots of basis-set errors in lattice constants, bulk moduli, and equilibrium free energies per atom calculated at the PBE level for the Bosoni test set using cc-pVXZ-s for X=DX = D, T, and Q. MOLOPT-DZVP and MOLOPT-TZV2P results are included for comparison. Each box plot includes results for 110 solids corresponding to the 11 elements considered in this work. All calculations employed the GTH-PBE pseudopotentials. The reference data were generated with a large plane-wave basis and the same pseudopotentials.

For all three properties, cc-pVXZ-s converges rapidly toward the PW reference as the zeta level increases, with improvements in both the median error and its spread. DZ already gives reasonable median errors of approximately 0.01 Å for a0a_0, −2 GPa for B0B_0, and 4 mEh for F0F_0. The error distributions are relatively wide, however, with several sizable outliers. Increasing the basis size to TZ and QZ further improves the medians and systematically narrows the error distributions. At QZ, the median errors for all three properties are close to zero. The absolute errors across all 110 solids are below 0.01 Å for a0a_0, 4 GPa for B0B_0, and 2 mEh for F0F_0, demonstrating consistent accuracy across this diverse test set.

For comparison, fig. 8 includes results for the MOLOPT-DZVP and MOLOPT-TZV2P basis sets optimized using the UZH protocol for PBE calculations.13 Both basis sets are molecularly optimized and have the same sizes as cc-pVDZ-s and cc-pVTZ-s, respectively, for the elements considered here. The cc-pVDZ-s and cc-pVTZ-s results closely match their MOLOPT counterparts for a0a_0 and B0B_0, while cc-pVTZ-s gives smaller errors in F0F_0. Both basis families show similar patterns of large outliers at DZ, which suggests that the errors primarily reflect limitations of the DZ basis structure, particularly the absence of ff functions, rather than differences in parameterization. The SO3_3 solid, for example, has the largest absolute errors in bulk modulus and free energy with both DZ basis sets. This can be attributed to the high formal oxidation state of S, whose accurate description requires ff functions and hence a TZ basis set. Overall, the results in fig. 8 suggest that the correlation-consistent cc-pVXZ-s basis sets, optimized using HF and correlated wavefunction methods, can be reliably applied to periodic DFT calculations without re-optimization and match the accuracy of state-of-the-art DFT basis sets.

PBE band structures

Figure 9 further illustrates the accuracy of cc-pVXZ-s through PBE band structures for the 7 semiconductors and aluminum metal. Relative to the PW reference, cc-pVDZ-s already achieves quantitative accuracy for all occupied bands and the lowest few virtual bands. Higher-energy virtual bands are more challenging for cc-pVDZ-s, but their accuracy improves systematically with cc-pVTZ-s and cc-pVQZ-s. The latter closely reproduces all PW reference bands within 30 eV above the Fermi level, including strongly entangled virtual bands in small-gap semiconductors (e.g., Si and AlP) and aluminum metal. These results suggest that accurate band structures emerge naturally from our regularized atom-optimized Gaussian basis sets. This differs from previous approaches that explicitly optimize basis sets against PW band structures of reference solids.32,34

Figure 9. PBE band structures for the 7 semiconductors and FCC Al calculated with the cc-pVXZ-s Gaussian basis sets and a plane-wave basis, using the GTH-PBE pseudopotentials.

MP2 equation of state

The molecular MP2 and periodic PBE benchmarks above establish the accuracy of cc-pVXZ-s in both mean-field and correlated calculations. As a final validation, we examine cc-pVXZ-s in periodic wavefunction calculations through MP2 EOS calculations for the 7 semiconductors. The left column of Figure 10 shows the basis-set errors in a0a_0, B0B_0, and EcohE_{\mathrm{coh}} for cc-pVXZ-s optimized for the all-electron potential, the GTH-HF pseudopotential, and the ccECP. Here, errors are calculated against the CBS(T,Q) extrapolations using the respective cc-pVXZ-s basis sets. For all three properties, the cc-pVXZ-s basis sets developed for different nuclear potentials exhibit nearly identical rapid convergence with zeta level. The cc-pVDZ-s basis sets achieve reasonable accuracy for the two structural properties, with a median error of ca. 0.03 Å for a0a_0 and −7 GPa for B0B_0, reduced to below 0.01 Å and −2 GPa with cc-pVQZ-s. As with the molecular atomization energies in section V A 1, EcohE_{\mathrm{coh}} shows stronger basis-set dependence. The smallest basis, cc-pVDZ-s, gives a sizable median error of ca. −0.6 eV/atom, which is reduced to ca. −0.2 and −0.1 eV/atom using cc-pVTZ-s and cc-pVQZ-s, respectively.

Figure 10. Box plots of basis-set errors in a0a_0, B0B_0, and EcohE_{\mathrm{coh}} calculated at the MP2 level for the 7 semiconductors using different Gaussian basis sets and nuclear potentials. The left panels show results obtained using the cc-pVXXZ-s (X=DX = \mathrm{D}, T\mathrm{T}, and Q\mathrm{Q}) basis sets developed in this work for the all-electron potential, the GTH-HF pseudopotential, and the ccECP. Errors are relative to the respective CBS(T,Q) extrapolations. The two right columns show all-electron and GTH-HF results obtained with basis sets from the literature. The all-electron results use cc-pVXXZ1^1 (X=DX = \mathrm{D} and T\mathrm{T}), pob-TZVP,14^{14} and pob-XXZVP-rev236^{36} (X=DX = \mathrm{D} and T\mathrm{T}), while the GTH-HF results use GTH-cc-pVXXZ15^{15} (X=DX = \mathrm{D}, T\mathrm{T}, and Q\mathrm{Q}). Errors are relative to CBS(T,Q) extrapolations obtained with our cc-pVXXZ-s basis sets for the corresponding nuclear potential.

For all three properties, CBS(D,T) extrapolation achieves accuracy comparable to cc-pVQZ-s and offers a computationally efficient alternative. Similar systematic convergence of cc-pVXXZ-s in HF EOS calculations is presented in fig. S15. In these calculations, both the median error and its spread converge even faster with zeta level, with cc-pVTZ-s essentially reaching CBS accuracy.

The two right columns of Figure 10 show all-electron and GTH-HF results obtained with basis sets from the literature. In all-electron calculations, standard cc-pVXXZ is less accurate than cc-pVXXZ-s at DZ for all three properties, while the two families achieve similar accuracy at TZ and with CBS(D,T) extrapolation. For GTH-HF, the GTH-cc-pVXXZ basis sets15 show a similar trend, with larger errors than cc-pVXXZ-s at DZ and comparable accuracy at TZ and QZ. The CBS(T,Q) extrapolations from the two families also agree closely, with median differences of ca. 0.005 Å for a0a_0, 3 GPa for B0B_0, and 0.05 eV/atom for EcohE_{\mathrm{coh}}. Together with the molecular benchmarks in section V A 1, these results support the accuracy of cc-pVXXZ-s in periodic wavefunction calculations. Regularized atomic optimization also enables numerically stable calculations with cc-pVQZ-s in solids where conventional cc-pVQZ suffers from

Comparison with the solid-optimized all-electron pob-family basis sets in Figure 10 provides additional insight. For all three properties, pob-TZVP gives substantially larger errors than cc-pVTZ-s in both the median and spread. This can be attributed to the overly compact valence primitives in pobTZVP, which cause sizable BSSE in molecular or periodic calculations.36 The pob-rev2 series was designed to mitigate this limitation.36 As shown in Figure 10, pob-DZVP-rev2 and pob-TZVP-rev2 improve upon pob-TZVP, especially for a0a_0. Both pob-rev2 basis sets, however, still have much wider error distributions than cc-pVXZ-s, and the improvement from DZVP to TZVP is less clear, making them less appealing for periodic wavefunction calculations. As suggested previously,14 augmenting pob-family basis sets with suitable diffuse valence primitives may substantially improve their accuracy. In principle, MCAO provides a systematic way to determine these additional valence primitives as demonstrated in our previous work,16 although we do not explore this possibility here.

Bulk properties of semiconductors from wavefunction methods

To illustrate the practical use of cc-pVXZ-s, we calculate a0a_0, B0B_0, and EcohE_{\mathrm{coh}} for the 7 semiconductors using a hierarchy of wavefunction methods, including HF, MP2, SCS-MP2, L-CCSD+Δ\DeltaMP2, and L-CCSD(T)+Δ\DeltaMP2. All calculations are performed using the all-electron potential and extrapolated to the combined basis-set and thermodynamic limits using the protocol described in section IV E. The final results are visualized in Figure 11 and summarized in section A.

Figure 11. Box plots of errors in a0a_0, B0B_0, and EcohE_{\mathrm{coh}} calculated using HF, MP2, SCS-MP2, L-CCSD+Δ\DeltaMP2, and L-CCSD(T)+Δ\DeltaMP2 for the 7 semiconductors, relative to experimental references corrected for zero-point anharmonic effects.78 All calculations have been extrapolated to the combined basis-set and thermodynamic limits using the protocol described in section IV E.

For all three properties, the theoretical predictions converge systematically as the treatment of electron correlation is refined. For B0B_0 and EcohE_{\mathrm{coh}}, our highest level of theory, L-CCSD(T)+Δ\DeltaMP2, achieves excellent overall agreement with the experimental references. For a0a_0, however, the median error is smallest at MP2 and increases at higher levels of theory, with L-CCSD(T)+Δ\DeltaMP2 systematically overestimating the experimental values by ca. 0.02 Å. Given the reliable CBS extrapolations enabled by cc-pVXZ-s, the remaining errors in our CCSD(T) calculations likely arise from the approximate MP2-based finite-size treatment in eq. (25). These results therefore demonstrate the predictive power of correlated wavefunction methods for solids while motivating the development of more scalable periodic local correlation methods.40

Despite the good overall agreement with experiment, an outlier is apparent in the bulk-modulus comparison in Figure 11. Closer inspection identifies this outlier as BN. Table II compares its bulk modulus from several theoretical methods with experimental values from multiple sources. Except for L-CCSD+Δ\DeltaMP2, which predicts a relatively high value of 405 GPa, all correlated wavefunction predictions cluster around 395 GPa. The experimental reference used in Figure 11 is 410 GPa, obtained by Schimka and co-workers78 by adding a zero-point correction of ca. 10 GPa to the room-temperature experimental value of 400 GPa reported by Grim­sdich and co-workers.91 Goncharov and co-workers94 assigned this measurement a large error bar of 20 GPa, making it a less precise benchmark for comparison with theory. Table II also lists B0B_0 from three other independent measurements using various experimental methods with smaller uncertainties.92–94 After applying the zero-point correction of 10 GPa and an estimated correction of 2 GPa (based on the thermal expansion coefficient reported by Goncharov and co-workers93), these measurements give reference values of 384–399 GPa, with a mean of ca. 394 GPa, in excellent agreement with the theoretical consensus. This case study illustrates the difficulty of rigorous comparisons between theory and experiment, even for “simple” solids like BN. The cc-pVXZ-s basis sets developed here help remove basis-set incompleteness as a prominent source of error in correlated wavefunction calculations, facilitating such comparisons in future studies.

MethodB0B_0 (GPa)
HF429
MP2397
SCS-MP2395
L-CCSD+Δ\DeltaMP2405
L-CCSD(T)+Δ\DeltaMP2392
Brillouin412 ±\pm 20 (400 ±\pm 20)
Ultrasonic384 ±\pm 4 (372 ±\pm 4)
X-ray399 ±\pm 4 (387 ±\pm 4)
Brillouin398 ±\pm 4 (386 ±\pm 4)

Table II. Bulk modulus of cubic boron nitride from theoretical calculations and experimental measurements. All calculations are extrapolated to the combined basis-set and thermodynamic limits using the all-electron cc-pVXZ-s basis sets developed in this work. The experimental values include corrections for zero-point motion (+10 GPa) and thermal expansion (+2 GPa). The uncorrected room-temperature values are listed in parentheses.

Conclusion

In summary, we have developed and numerically validated the cc-pVXZ-s series (XX = D, T, and Q) of correlation-consistent Gaussian basis sets for solids, covering hydrogen and the pp-block elements from boron to chlorine. We provide basis sets for the all-electron potential, the GTH pseudopotentials optimized for HF and PBE calculations, and the ccECPs. Using an improved MCAO algorithm, we obtain cc-pVXZ-s basis sets that are numerically stable across a diverse set of reference materials drawn from a recent large-scale DFT benchmark study. Benchmarks against standard cc-pVXZ calculations for molecules and plane-wave calculations for periodic solids establish the accuracy, transferability, and numerical stability of cc-pVXZ-s in both mean-field and correlated wavefunction calculations and confirm reliable extrapolation to the CBS limit. Periodic wavefunction calculations with the all-electron cc-pVXZ-s basis sets demonstrate systematic convergence toward experimental bulk moduli and cohesive energies for the 7 semiconductors, while a small systematic overestimation of lattice constants suggests potential limitations in the approximate finite-size treatment.

The present work offers two main contributions to future solid-state basis-set design. Technically, it establishes MCAO as an effective tool for fine-tuning basis-set numerical stability across an essentially arbitrary number of reference solids. Chemically, for the selected H and pp-block elements, it demonstrates that regularization against judiciously chosen reference solids yields basis sets that balance numerical stability, energy accuracy, and transferability across diverse solids. Extending MCAO to other elements will require element-specific adaptations, such as adding semicore-correlating functions for ss-block metals,95 treating multireference character in dd-block metals,49 and accounting for relativistic effects in heavy elements.[8, 9] The advances reported here provide a foundation for these extensions, which we plan to present in subsequent papers in this series.

Predicted lattice constants, bulk moduli, and cohesive energies for the 7 semiconductors

Tables III to V report our predictions of a0a_0, B0B_0, and EcohE_{\mathrm{coh}} for the 7 semiconductors at the HF, MP2, SCS-MP2, L-CCSD+Δ+\DeltaMP2, and L-CCSD(T)+Δ+\DeltaMP2 levels of theory. The tables also report the mean absolute error (MAE) and mean absolute relative error (MARE) of each method relative to experiment.

SolidHFMP2SCS-MP2L-CCSD +ΔMP2L-CCSD(T) +ΔMP2Expt.
BN3.5993.5963.6033.6013.6103.592
BP4.5854.5004.5214.5364.5414.525
AlN4.3674.3904.3914.3794.3914.368
AlP5.5425.4485.4675.4825.4855.451
C3.5513.5423.5523.5543.5633.553
Si5.5075.4005.4245.4465.4495.421
SiC4.3704.3544.3614.3624.3724.346
MAE (Å)0.0390.0140.0110.0150.022⋯
MARE (%)0.80.30.20.30.5⋯

Table III. Predicted lattice constants at the indicated levels of theory, including the mean absolute error (MAE) and mean absolute relative error (MARE) relative to experiment. Experimental values, which have been corrected for zero-point motion, are from ref 78.

SolidHFMP2SCS-MP2L-CCSD +ΔMP2L-CCSD(T) +ΔMP2Expt.
BN429397395405392394
BP176182178178173168
AlN229205206214207206
AlP959695969487
C497464458466449455
Si104103102103100101
SiC244231230233225229
MAE (GPa)196383⋯
MARE (%)7.53.62.44.12.1⋯

Table IV. Predicted bulk moduli at the indicated levels of theory, including the mean absolute error (MAE) and mean absolute relative error (MARE) relative to experiment. Experimental values, corrected for zero-point motion, are from ref 78, except for BN, for which we use the revised experimental average discussed in section V B.

SolidHFMP2SCS-MP2L-CCSD +ΔMP2L-CCSD(T) +ΔMP2Expt.
BN4.767.237.056.626.776.76
BP3.435.725.475.055.235.14
AlN3.906.166.035.635.785.85
AlP2.754.614.454.144.304.32
C5.328.087.787.357.537.55
Si3.055.144.904.574.764.68
SiC4.556.966.696.296.466.48
MAE (eV/atom)1.860.450.230.160.04⋯
MARE (%)32.47.83.92.80.8⋯

Table V. Predicted cohesive energies at the indicated levels of theory, including the mean absolute error (MAE) and mean absolute relative error (MARE) relative to experiment. Experimental values, which have been corrected for zero-point motion, are from ref 78.

Supporting Information

See the Supporting Information for (i) geometry and basis-set data files; (ii) raw data for the molecular atomization energy and periodic EOS calculations; (iii) condition numbers of literature basis sets for solids in the Bosoni set; (iv) a breakdown of Bosoni-set PBE EOS errors by element; (v) HF EOS errors for the 7 semiconductors; (vi) computational details of periodic local coupled cluster calculations; and (vii) the algorithm for optimizing auxiliary basis sets for density fitting.

Conflict of interest

The authors declare no competing financial interest.

Data availability

The data that support the findings of this study are available in the GitHub repository: https://github.com/hongzhouye/supporting_data/tree/main/2026/ccpvxz_solids

Acknowledgments

This work was supported by the National Science Foundation under Grant No. CHE-2543461. We acknowledge computing resources provided by the Division of Information Technology at the University of Maryland, College Park.

Supporting Information for Correlation-Consistent Gaussian Basis Sets for Periodic Solids. I. Hydrogen and p-block Elements from Boron to Chlorine

Jincheng Yu; Hong-Zhou Ye

si.pdf

Supporting Information for Correlation-Consistent Gaussian Basis Sets for Periodic Solids. I. Hydrogen and p-block Elements from Boron to Chlorine

Geometry and basis-set data

Geometry files (in XYZ or VASP POSCAR format) and Gaussian basis-set data files (in NWChem format) are available in the following GitHub repository:

https://github.com/hongzhouye/supporting_data/tree/main/2026/ccpvxz_solids

The geometry files include

  • the Bosoni set, reproduced from ref 1 using the reference cell volumes V0V_0;

  • the 19 molecular oxides, optimized at the ω\omegaB97X-D3^{3} [2, 3, 4]/def2-TZVPP5^{5} level using ORCA;[6]

  • the 7 semiconductors and FCC aluminum, reproduced from ref 7 using the experimental lattice constants a0a_0 listed in Table S1.

MaterialLatticea0a_0 (Å)
BNB33.592
BPB34.525
CA43.553
AlNB34.368
AlPB35.451
SiA45.421
SiCB34.346
AlA14.018

Table S1. Lattice constants of the 7 semiconductors and aluminum.

The basis-set data files include

  • the cc-pVXZ-s basis sets and the associated cc-pVXZ-s-FIT auxiliary basis sets for all four nuclear potentials;

  • the FIT auxiliary basis sets optimized for the cc-pVXZ, pob-TZVP, pob-D/TZVP, and MOLOPT-DZVP/TZV2P basis sets.

Raw data

The GitHub repository linked above also provides the raw data for the benchmarks presented in the main text, including

  • the all-electron HF and MP2 molecular atomization energies calculated using the standard cc-pVXZ (X=DX = D, T, Q, and 5) and cc-pVXZ-s (X=DX = D, T, and Q) basis sets;

  • the PBE free energies for the Bosoni EOS calculations with the GTH-PBE pseudopotentials, including results obtained with cc-pVXZ-s (X=DX = D, T, and Q), MOLOPT-DZVP, MOLOPT-TZV2P, and a converged PW basis;

  • the HF, MP2, LNO-CCSD, and LNO-CCSD(T) energies used in EOS calculations for the 7 semiconductors, including HF and MP2 results with various pseudopotentials and all-electron LNO-CC results.

Supplementary figures

Figure S1. Maximum condition number (κmax⁡\kappa_{\max}) over the 10 reference solids for each element, evaluated with the conventional def2-family basis sets and the all-electron cc-pVXZ-s sets developed in this work. The horizontal red line indicates κmax⁡=109\kappa_{\max}=10^{9}.

Figure S2. Maximum condition number (κmax⁡\kappa_{\max}) over the 10 reference solids for each element, evaluated with the ccECP-cc-pVXZ basis sets provided with ccECP[8] and the cc-pVXZ-s sets developed for ccECP in this work. The horizontal red line indicates κmax⁡=109\kappa_{\max}=10^{9}.

Figure S3. Maximum condition number (κmax⁡\kappa_{\max}) over the 10 reference solids for each element, evaluated with the GTH-cc-pVXZ basis sets from [9] and the cc-pVXZ-s sets developed for the GTH-HF pseudopotential in this work. The horizontal red line indicates κmax⁡=109\kappa_{\max} = 10^{9}.

Figure S4. Same as Figure 8 in the main text, but for the 10 solids in the H subset of the Bosoni test set.

Figure S5. Same as Figure 8 in the main text, but for the 10 solids in the B subset of the Bosoni test set.

Figure S6. Same as Figure 8 in the main text, but for the 10 solids in the C subset of the Bosoni test set.

Figure S7. Same as Fig. 8 in the main text, but for the 10 solids in the N subset of the Bosoni test set.

Figure S8. Same as Figure 8 in the main text, but for the 10 solids in the O subset of the Bosoni test set.

Figure S9. Same as Fig. 8 in the main text, but for the 10 solids in the F subset of the Bosoni test set.

Figure S10. Same as Figure 8 in the main text, but for the 10 solids in the Al subset of the Bosoni test set.

Figure S11. Same as Figure 8 in the main text, but for the 10 solids in the Si subset of the Bosoni test set.

Figure S12. Same as Figure 8 in the main text, but for the 10 solids in the P subset of the Bosoni test set.

Figure S13. Same as Figure 8 in the main text, but for the 10 solids in the S subset of the Bosoni test set.

Figure S14. Same as Figure 8 in the main text, but for the 10 solids in the Cl subset of the Bosoni test set.

Figure S15. Same as Figure 10 in the main text, but for errors at the HF level.

Local natural orbital–coupled cluster

In this section, we describe the computational details of the local natural orbital–coupled cluster (LNO-CC) method implemented in PySCF.[10] [11, 12, 13] We use this method to accurately estimate canonical CCSD and CCSD(T) equations of state (EOS) for the 7 semiconductors. LNO-CC rewrites the canonical CC correlation energy as a sum over fragment contributions,

ELNO−CC(η)=∑infragECC(i)(η)+ΔEMP2ext(η).(26)E_{\mathrm{LNO-CC}}(\eta)=\sum_{i}^{n_{\mathrm{frag}}} E_{\mathrm{CC}}^{(i)}(\eta)+\Delta E_{\mathrm{MP2}}^{\mathrm{ext}}(\eta). \tag*{(26)}

Each fragment correlation energy ECC(i)(η)E_{\mathrm{CC}}^{(i)}(\eta) is calculated within a local active space consisting of (i) a small set of localized orbitals and (ii) their optimized bath orbitals. In this work, we use the atom-based fragments defined in ref 13. Each fragment uses all intrinsic atomic orbitals14 (IAOs) on the corresponding atom as its localized orbitals. The bath orbitals are chosen to be the MP2 local natural orbitals with occupations beyond user-selected thresholds, ηocc\eta_{\mathrm{occ}} and ηvir\eta_{\mathrm{vir}}, for the occupied and virtual spaces, respectively. The ratio ηocc/ηvir\eta_{\mathrm{occ}}/\eta_{\mathrm{vir}} is fixed at 10 for optimal convergence in gapped systems, and only η≡ηvir\eta\equiv\eta_{\mathrm{vir}} is reported below. Orbitals outside the local active space contribute to the LNO-CC correlation energy through the external correction ΔEMP2ext\Delta E_{\mathrm{MP2}}^{\mathrm{ext}}, evaluated using MP2. As η\eta is tightened, the LNO-CCSD and LNO-CCSD(T) correlation energies systematically converge to their canonical counterparts in the same basis set and on the same kk-point mesh. Canonical CC results can be estimated from LNO-CC calculations at finite thresholds by linear extrapolation against ΔEMP2ext\Delta E_{\mathrm{MP2}}^{\mathrm{ext}},

ELNO−CC(η)=ECC+AΔEMP2ext(η)(27)E_{\mathrm{LNO-CC}}(\eta)=E_{\mathrm{CC}}+A\Delta E_{\mathrm{MP2}}^{\mathrm{ext}}(\eta) \tag*{(27)}

In this work, we use two-point extrapolations following ref 15, denoted by (η1,η2)(\eta_1,\eta_2), where η1,2\eta_{1,2} are the thresholds for the two calculations.

As discussed in the main text, the final L-CCSD+Δ\DeltaMP2 and L-CCSD(T)+Δ\DeltaMP2 total energies are obtained by adding the HF energy and the scaled MP2-based basis-set and finite-size correction to the LNO-CCSD and LNO-CCSD(T) correlation energies evaluated using the TZ basis sets and a 3×3×33 \times3 \times3 kk-point mesh. Because the HF energy and scaled MP2 correction are independent of η\eta, the final L-CCSD+Δ\DeltaMP2 and L-CCSD(T)+Δ\DeltaMP2 energies can be extrapolated to the canonical CC limit using eq. (2).

For diamond, we performed L-CC+Δ\DeltaMP2 calculations using η=10−4\eta= 10^{-4}, $3 \times 10^{-5}$, and 10−510^{-5}, allowing two extrapolations: a loose extrapolation, (10−4,3×10−5)(10^{-4},3 \times10^{-5}), and a tight extrapolation, $(3 \times 10−510^{-5}, 10−5)10^{-5}). As shown in Table S2, the loose and tight extrapolations agree to within ca. 0.001 Å for a0a_0 and 0.5 GPa for B0B_0, suggesting that the loose extrapolation is sufficient for both structural properties. Figure S16 further supports this choice for all 7 semiconductors: results at η=3×10−5\eta= 3 \times10^{-5} differ from the loose extrapolation by ca. 0.005 Å for a0a_0 and 1 GPa for B0B_0. We estimate the errors of the loose extrapolation as half of these differences, giving ca. 0.003 Å for a0a_0 and 0.5 GPa for B0B_0.

η\etaLNO-CCSDLNO-CCSD(T)
a0a_0 (Å)B0B_0 (GPa)EcohE_{\mathrm{coh}} (eV/atom)a0a_0 (Å)B0B_0 (GPa)EcohE_{\mathrm{coh}} (eV/atom)
10−410^{-4}3.551464.67.593.558452.67.66
3×10−53 \times10^{-5}3.552465.47.463.561450.47.59
10−510^{-5}3.553466.37.393.562449.97.55
(10−4,3×10−5)(10^{-4}, 3 \times10^{-5})3.554465.77.383.563448.87.54
(3×10−5,10−5)(3 \times10^{-5}, 10^{-5})3.554466.17.353.563449.37.53

Table S2. L-CCSD+Δ\DeltaMP2 and L-CCSD(T)+Δ\DeltaMP2 values of a0a_0, B0B_0, and EcohE_{\mathrm{coh}} for diamond, evaluated at several LNO truncation thresholds η\eta and extrapolated using eq. (2).

Figure S16. Convergence of a0a_0, B0B_0, and EcohE_{\mathrm{coh}} with the LNO truncation threshold η\eta for L-CCSD+Δ\DeltaMP2 (left) and L-CCSD(T)+Δ\DeltaMP2 (right). The references are the (10−4,3×10−5)(10^{-4}, 3 \times10^{-5}) extrapolation for a0a_0 and B0B_0 and the (3×10−5,10−5)(3 \times10^{-5}, 10^{-5}) extrapolation for EcohE_{\mathrm{coh}}.

For EcohE_{\mathrm{coh}}, Table S2 shows that the loose and tight extrapolations differ by 0.03 eV/atom for CCSD and 0.01 eV/atom for CCSD(T), supporting the use of the tight extrapolation for EcohE_{\mathrm{coh}}. Figure S16 shows that the loose–tight differences remain well within 0.03 eV/atom for CCSD and 0.01 eV/atom for CCSD(T) across all 7 semiconductors. We similarly estimate the error of the tight extrapolation for EcohE_{\mathrm{coh}} as half of the loose–tight difference, giving ca. 0.02 eV/atom.

Optimization of auxiliary basis sets

The cc-pVXZ-s-FIT auxiliary basis sets are uncontracted and parameterized using the Legendre polynomial-based approach described in ref 16, with up to 6 terms through degree 5 (i.e., kmax⁡=5k_{\max}=5). We found that this parameterization yields more compact optimized auxiliary basis sets than the ETB-based parameterization used in our previous work.17 The basis parameters are optimized by minimizing an objective that combines atomic DF errors from HF and MP2 calculations, inspired by previous work on auxiliary basis-set optimization,18–24

Laux=max⁡{γJKΔJ, γJKΔK, ΔEJ, ΔEK, ΔEMP2, γMP2ΔIMP2SS, γMP2ΔIMP2OS}+Pmin−exp(28)L_{\mathrm{aux}}=\max\left\{\gamma_{JK}\Delta J,\ \gamma_{JK}\Delta K,\ \Delta E_J,\ \Delta E_K,\ \Delta E_{\mathrm{MP2}},\ \gamma_{\mathrm{MP2}}\Delta I_{\mathrm{MP2}}^{\mathrm{SS}},\ \gamma_{\mathrm{MP2}}\Delta I_{\mathrm{MP2}}^{\mathrm{OS}}\right\}+P_{\mathrm{min-exp}} \tag*{(28)}

The terms in LauxL_{\mathrm{aux}} are defined below. (For H, all MP2-related terms are calculated using H2_2 at a bond length of 0.7414 Å.) Minimizing LauxL_{\mathrm{aux}} yields the optimized auxiliary basis set for a fixed size. Starting from a sufficiently large initial auxiliary basis set, we remove primitives from each ll channel, proceeding from high to low ll, to obtain the smallest auxiliary basis set satisfying

Laux≤10−5 a.u.(29)L_{\mathrm{aux}}\leq10^{-5}\ \mathrm{a.u.} \tag*{(29)}

This criterion yields atomic HF and MP2 energy errors below 10 μEh\mu E_{\mathrm{h}}, comparable to those of other optimized auxiliary basis sets.18,19

The first four terms in eq. (3) measure DF errors in HF calculations. ΔJ\Delta J and ΔK\Delta K are the maximum absolute errors in the HF Coulomb and exchange matrix elements,

ΔJ=max⁡μν∣J~μν−Jμν∣,(30)\Delta J=\max_{\mu\nu}\left|\widetilde{J}_{\mu\nu}-J_{\mu\nu}\right|, \tag*{(30)}
ΔK=max⁡μν∣K~μν−Kμν∣,(31)\Delta K=\max_{\mu\nu}\left|\widetilde{K}_{\mu\nu}-K_{\mu\nu}\right|, \tag*{(31)}

and ΔEJ=∣E~J−EJ∣\Delta E_J=\left|\widetilde{E}_J-E_J\right| and ΔEK=∣E~K−EK∣\Delta E_K=\left|\widetilde{E}_K-E_K\right| are errors in the HF and exchange energies. In all cases, X~\widetilde{X} and XX denote quantities evaluated using the density-fitted electron-repulsion integrals (ERIs) V~pqrs\widetilde{V}_{pqrs} and the exact ERIs VpqrsV_{pqrs}, respectively. All quantities are evaluated using a fixed ROHF density matrix calculated without DF, so the errors arise solely from the DF approximation.

The last three terms inside the maximum in eq. (3) measure DF errors in frozen-core MP2 calculations. ΔEMP2=∣E~MP2−EMP2∣\Delta E_{\mathrm{MP2}}=\left|\widetilde{E}_{\mathrm{MP2}}-E_{\mathrm{MP2}}\right| is the error in the MP2 correlation energy, and ΔIMP2SS\Delta I_{\mathrm{MP2}}^{\mathrm{SS}} and ΔIMP2OS\Delta I_{\mathrm{MP2}}^{\mathrm{OS}} are errors in same-spin and opposite-spin MP2 wavefunction amplitudes adapted from literature,18,19,23,24

ΔEMP2SS=1∣EMP2SS∣∑iajb∣ΔViajb∗(ΔViajb−ΔVibja)∣Δiajb,(32)\Delta E_{\mathrm{MP2}}^{\mathrm{SS}} = \frac{1}{\lvert E_{\mathrm{MP2}}^{\mathrm{SS}}\rvert} \sum_{iajb} \frac{\left\lvert\Delta V_{iajb}^{*}\left(\Delta V_{iajb}-\Delta V_{ibja}\right)\right\rvert}{\Delta_{iajb}}, \tag*{(32)}
ΔEMP2OS=1∣EMP2OS∣∑iajb∣ΔViajb∣2Δiajb(33)\Delta E_{\mathrm{MP2}}^{\mathrm{OS}} = \frac{1}{\lvert E_{\mathrm{MP2}}^{\mathrm{OS}}\rvert} \sum_{iajb} \frac{\left\lvert\Delta V_{iajb}\right\rvert^{2}}{\Delta_{iajb}} \tag*{(33)}

where ΔViajb=V~iajb−Viajb\Delta V_{iajb}=\widetilde{V}_{iajb}-V_{iajb}, Δiajb=−εi+εa−εj+εb\Delta_{iajb}=-\varepsilon_i+\varepsilon_a-\varepsilon_j+\varepsilon_b is the orbital energy difference, and EMP2SSE_{\mathrm{MP2}}^{\mathrm{SS}} and EMP2OSE_{\mathrm{MP2}}^{\mathrm{OS}} are the same-spin and opposite-spin MP2 correlation energies, respectively.

The parameters γJK=0.1\gamma_{JK}=0.1 and γMP2=10\gamma_{\mathrm{MP2}}=10 impose looser tolerances on errors in the JJ and KK matrices and stricter tolerances on errors in the MP2 amplitudes. The former is motivated by the observation that ΔJ\Delta J and ΔK\Delta K are typically larger than ΔEJ\Delta E_J and ΔEK\Delta E_K when using standard cc-pVXZ-JKFIT basis sets, while the latter follows literature,18,19 which requires small errors in both the MP2 amplitudes and the MP2 energy.

The last term in eq. (3), Pmin−expP_{\mathrm{min-exp}}, penalizes overly diffuse primitives that can cause strong linear dependence in the DF metric matrix and destabilize DF-based SCF calculations.25,26 Let α0(l)\alpha_0^{(l)} and β0(l)\beta_0^{(l)} be the smallest exponents in angular-momentum channel ll of the AO and auxiliary basis sets, respectively. Define

βmin(l)=min⁡(l1,l2)α0(l1)+α0(l2),(34)\beta_{\mathrm{min}}^{(l)}=\min_{(l_1,l_2)}\alpha_0^{(l_1)}+\alpha_0^{(l_2)}, \tag*{(34)}

where the minimization is carried out over all (l1,l2)(l_1,l_2) pairs such that ∣l1−l2∣≤l≤l1+l2\lvert l_1-l_2\rvert\le l\le l_1+l_2 and l1+l2+ll_1+l_2+l is even. We require the auxiliary exponents to satisfy

β0(l)βmin(l)≳rmin−exp(35)\frac{\beta_0^{(l)}}{\beta_{\mathrm{min}}^{(l)}}\gtrsim r_{\mathrm{min-exp}} \tag*{(35)}

for all ll channels in the auxiliary basis set. The scaling factor is set to rmin−exp=0.45r_{\mathrm{min-exp}}=0.45, obtained by analyzing the exponents of cc-pVXZ and cc-pVXZ-JKFIT. The constraint in eq. (8) is enforced using the exponent-ratio penalty in eq. (19) of the main text. To obtain Pmin−expP_{\mathrm{min-exp}}, we apply this penalty to β0(l)/βmin(l)\beta_0^{(l)}/\beta_{\mathrm{min}}^{(l)} with threshold rmin−expr_{\mathrm{min-exp}} and sum over all ll channels. The minimum-exponent penalty helps regularize the initial stages of optimization but has no effect on the optimized cc-pVXZ-s-FIT auxiliary basis sets, whose exponents readily satisfy eq. (8), with a negligible Pmin−expP_{\mathrm{min-exp}} (below 10−3010^{-30}).

References

References

  1. [1]E. Bosoni, L. Beal, M. Bercx, P. Blaha, S. Blügel, J. Bröder, M. Callsen, S. Cottenier, A. De-gomme, V. Dikan, K. Eimre, E. Flage-Larsen, M. Fornari, A. Garcia, L. Genovese, M. Giantomassi, S. P. Huber, H. Janssen, G. Kastlunger, M. Krack, G. Kresse, T. D. Kühne, K. Lejaeghere, G. K. H. Madsen, M. Marsman, N. Marzari, G. Michalichek, H. Mirhosseini, T. M. A. Müller, G. Petretto, C. J. Pickard, S. Poncé, G.-M. Rignanese, O. Rubel, T. Ruh, M. Sluÿts, D. E. P. Vanpoucke, S. Vijay, M. Wolloch, D. Wortmann, A. V. Yakutovich, J. Yu, A. Zadoks, B. Zhu, and G. Pizzi, “How to verify the precision of density-functional-theory implementations via reproducible and universal workflows,” Nat. Rev. Phys. 6, 45–58 (2024).
  2. [2]S. Grimme, J. Antony, S. Ehrlich, and H. Krieg, “A consistent and accurate ab initio parametrization of density functional dispersion correction (DFT-D) for the 94 elements H–Pu,” The Journal of Chemical Physics 132, 154104 (2010).DOI
  3. [3]J.-D. Chai and M. Head-Gordon, “Systematic optimization of long-range corrected hybrid density functionals,” The Journal of Chemical Physics 128, 084106 (2008).DOI
  4. [4]Y.-S. Lin, G.-D. Li, S.-P. Mao, and J.-D. Chai, “Long-range corrected hybrid density functionals with improved dispersion corrections,” Journal of Chemical Theory and Computation 9, 263–272 (2013).DOI
  5. [5]F. Weigend and R. Ahlrichs, “Balanced basis sets of split valence, triple zeta valence and quadruple zeta valence quality for h to rn: Design and assessment of accuracy,” Phys. Chem. Chem. Phys. 7, 3297–3305 (2005).pubmed.ncbi.nlm.nih.gov/16240044
  6. [6]F. Neese, “the orca program system,” Wiley Interdiscip. Rev. Comput. Mol. Sci 2, 73–78 (2012).
  7. [7]L. Schimka, J. Harl, and G. Kresse, “Improved hybrid functional for solids: The hsesol functional,” J. Chem. Phys. 134, 024116 (2011).DOI
  8. [8]M. C. Bennett, C. A. Melton, A. Annaberdiyev, G. Wang, L. Shulenburger, and L. Mitas, “A new generation of effective core potentials for correlated calculations,” J. Chem. Phys. 147, 224106 (2017).DOI
  9. [9]H.-Z. Ye and T. C. Berkelbach, “Correlation-consistent gaussian basis sets for solids made simple,” J. Chem. Theory Comput. 18, 1595–1606 (2022).DOI
  10. [10]Q. Sun, T. C. Berkelbach, N. S. Blunt, G. H. Booth, S. Guo, Z. Li, J. Liu, J. D. McClain, E. R. Sayfuty ar ova, S. Sharma, S. Wouters, and G. K.-L. Chan, “Pyscf: the python-based simulations of chemistry framework,” Wiley Interdiscip. Rev. Comput. Mol. Sci 8, e1340 (2018).arxiv.org/abs/1701.08223

Paper details

vlnc:W_cwq3g6qq
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents