betaFeedback

WorksPhysics and Astronomy

Projected gradient descent algorithm for ab initio\textit{ab initio} crystal structure relaxation under a fixed unit cell volume

Abstract

This paper is concerned with ab initio\textit{ab initio} crystal structure relaxation under a fixed unit cell volume, which is a step in calculating the static equations of state and forms the basis of thermodynamic property calculations for materials. The task can be formulated as an energy minimization with a determinant constraint. Widely used line minimization-based methods (e.g., conjugate gradient method) lack both efficiency and convergence guarantees due to the nonconvex nature of the feasible region as well as the significant differences in the curvatures of the potential energy surface with respect to atomic and lattice components. To this end, we propose a projected gradient descent algorithm named PANBB. It is equipped with (i) search direction projections onto the tangent spaces of the nonconvex feasible region for lattice vectors, (ii) distinct curvature-aware initial trial step sizes for atomic and lattice updates, and (iii) a nonrestrictive line minimization criterion as the stopping rule for the inner loop. It can be proved that PANBB favors theoretical convergence to equilibrium states. Across a benchmark set containing 223 structures from various categories, PANBB achieves average speedup factors of approximately 1.41 and 1.45 over the conjugate gradient method and direct inversion in the iterative subspace implemented in off-the-shelf simulation software, respectively. Moreover, it normally converges on all the systems, manifesting its unparalleled robustness. As an application, we calculate the static equations of state for the high-entropy alloy AlCoCrFeNi, which remains elusive owing to 160 atoms representing both chemical and magnetic disorder and the strong local lattice distortion. The results are consistent with the previous calculations and are further validated by experimental thermodynamic data.

Introduction

Crystal structure relaxation seeks equilibrium states on the potential energy surface (PES). It underpins various applications such as crystal structure prediction 1 and high-throughput calculations in material design 2. Relaxing crystal structures under a fixed unit cell volume is a step in calculating the static equations of state (EOS) for materials 3 and also presents the volume-dependent configurations for thermal models such as the quasiharmonic approximation 4, classical mean-field potential 5, and electronic excitation based on Mermin statistics 6, etc.

Mathematically, relaxing the crystal structure under a fixed unit cell volume can be formulated as minimizing the potential energy with a determinant constraint, where a nonconvex feasible region arises. Developing convergent methods for general nonconvex constrained optimization problems remains challenging, since classical tools (e.g., projection) and theories (e.g., duality) become inapplicable. As far as we know, little attention has been paid to the determinant-constrained optimization.

In practice, line minimization (LM)-based methods are usual choices for this purpose. A flowchart of the widely used LM-based methods is depicted in Figure 1. Each iteration of these methods primarily consists of updating the search directions followed by an LM process to determine the step sizes.

Flowchart of the widely used LM-based methods

Figure 1. A flowchart of the widely used LM-based methods. The bold text “KS” refers to solving Kohn-Sham (KS) equations to obtain new energy, atomic forces, and lattice stress whenever the configuration gets updated. The LM process is marked out by the shaded box. The subscript “inter” indicates the intermediate lattice matrix.

Among the popular representatives in this class are the conjugate gradient method (CG) [21] and direct inversion in the iterative subspace (DIIS) [22]. Specifically, CG searches along the conjugate gradient directions made up by previous directions and current forces, while DIIS employs historical information to construct the quasi-Newton directions. In most cases, CG exhibits steady convergence but can be unsatisfactory in efficiency. DIIS converges rapidly in the small neighborhoods of local minimizers but can easily diverge when starting far from equilibrium states.

As shown in Figure 1, the performance of the LM-based methods crucially depends on three key factors: search directions, initial trial step sizes, and LM criterion. However, the fixed-volume relaxation presents substantial difficulties on their designs. Firstly, due to the nonconvex nature of the feasible region, employing search directions for lattice vectors without tailored modifications can result in considerable variations in the configurations, giving rise to drastic fluctuations in the potential energies; for an illustration, please refer to Figure 2 later. Secondly, on account of the significant differences in the curvatures of PES with respect to the atomic and lattice components, adopting shared initial trial step sizes, as in the widely used LM-based methods (see S2 in Figure 1), can be detrimental to algorithm efficiency. This point is recognized through extensive numerical tests, where the relative displacements in the lattice vectors during relaxation are typically much smaller than those in the atomic positions. We refer readers to Figure 3 later for an illustration. Thirdly, the monotone criterion utilized in the widely used LM-based methods mandates a strict reduction in the potential energy after each iteration, often making the entire procedure trapped in the LM process.

Figure 2. Left: a comparison between iterations with and without the gradient projection (3) in two-dimensional context. The xx- and yy-axes refer to the first and second singular values of lattice matrix AA, respectively. The black solid line stands for the feasible region det⁡(A)=1\det(A)=1, the black square denotes the last iterate A(k)A^{(k)}, and the black dashdotted line refers to the tangent space T(k)\mathcal{T}^{(k)} of feasible region at A(k)A^{(k)}. The black and red arrows represent respectively the lattice force Flatt(k)F_{\mathrm{latt}}^{(k)} and its projection onto T(k)\mathcal{T}^{(k)}, namely, F~latt(k)\widetilde{F}_{\mathrm{latt}}^{(k)}. The black and red circles show the next iterates without and with the gradient projection (3), respectively, after performing the scaling operation indicated by the black dotted lines. Right: the relaxation history of PANBB on an 8×1×18 \times1 \times1 silicon supercell (64 atoms), where “#KS” denotes the number of solving the KS equations, “w/ grad. proj.” and “w/o grad. proj.” stand for the results using F~latt(k)\widetilde{F}_{\mathrm{latt}}^{(k)} and Flatt(k)F_{\mathrm{latt}}^{(k)}, respectively.

Figure 3. The relaxation on silicon supercells (8 ∼ 64 atoms). Left: the relaxation history of PANBB on an 8 × 1 × 1 silicon supercell. “Distinct” and “Shared” stand for the results of PANBB using distinct and shared step sizes for atomic and lattice components, respectively. Right: the average #KS by CG and PANBB for relaxing silicon supercells of different sizes. Each system size is associated with 10 initial configurations, which are generated by random perturbations from equilibrium states; the deviations in the atomic positions are less than 0.1 Å and the deviations in the lattice vectors are less than 0.02 Å.

According to incomplete statistics, the CG implemented in off-the-shelf simulation software spends approximately 60% of its overhead on the trials rejected by the LM criterion, as supported by the Supplemental Material. It is worth noting that the authors of [23] devise a nonmonotone LM criterion but only consider atomic degrees of freedom.

In this work, we propose a projected gradient descent algorithm, called PANBB, that addresses all the aforementioned difficulties. Specifically, PANBB (i) employs search direction projections onto the tangent spaces of the nonconvex feasible region for lattice vectors, (ii) adopts distinct curvature-aware step sizes for the atomic and lattice components, and (iii) incorporates a nonmonotone LM criterion that encompasses both atomic and lattice degress of freedom. The convergence of PANBB to equilibrium states has been theoretically established.

PANBB has been tested across a benchmark set containing 223 systems from various categories, with approximately 68.6% of them are metallic systems, for which the EOS calculations are more demanding. PANBB exhibits average speedup factors of approximately 1.41 and 1.45 over the CG and DIIS implemented in off-the-shelf simulation software, respectively, in terms of running time. In contrast to the failure rates of approximately 4.9% for CG and 25.1% for DIIS, PANBB consistently converges to the equilibrium states across the broad, manifesting its unparalleled robustness. As an application of PANBB, we calculate the static EOS of the high-entropy alloy (HEA) AlCoCrFeNi, whose fixed-volume relaxation remains daunting owing to the strong local lattice distortion (LLD) [24]. The numerical results are consistent with the previous calculations [25] obtained by relaxing only atomic positions. We further apply the modified mean field potential (MMFP) approach [18] to the thermal EOS, which is validated by the X-ray diffraction and diamond anvil cell experiments [26].

The notations are gathered as follows. We use N∈NN \in\mathbb{N} for the number of atoms, and use R=[r1,…,rN]∈R3NR = [r_1, \ldots, r_N] \in\mathbb{R}^{3N} and A:=[a1,a2,a3]∈R3×3A := [a_1, a_2, a_3] \in\mathbb{R}^{3 \times3} to denote the Cartesian atomic positions and lattice matrix, respectively. The reciprocal lattice matrix A−TA^{-\mathsf{T}} is abbreviated as B∈R3×3B \in\mathbb{R}^{3 \times3}. We denote by E(R,A)∈RE(R, A) \in\mathbb{R}, Fatom(R,A)∈R3NF_{\mathrm{atom}}(R, A) \in\mathbb{R}^{3N}, Flatt(R,A)∈R3×3F_{\mathrm{latt}}(R, A) \in\mathbb{R}^{3 \times3}, and Σdev(R,A)∈R3×3\Sigma_{\mathrm{dev}}(R, A) \in\mathbb{R}^{3 \times3}, respectively, the potential energy, atomic forces, lattice force, full stress tensor, and deviatoric stress tensor evaluated at the configuration (R,A)(R, A); if no confusion arises, we omit specifying (R,A)(R, A). From [27, 28], one obtains the relation

Flatt=∇ΣB−FatomRTB.(1)F_{\mathrm{latt}} = \nabla\Sigma B - F_{\mathrm{atom}} R^{\mathsf{T}} B. \tag*{(1)}

When describing algorithms, we use the first and second superscripts in brackets to indicate respectively the outer and inner iteration numbers (e.g., E(k)E^{(k)} and α(k−1,ℓ)\alpha^{(k-1,\ell)}). We denote by “I” the identity mapping from R3×N\mathbb{R}^{3 \times N} to R3×N\mathbb{R}^{3 \times N}. The notation “⟨⋅,⋅⟩\langle\cdot,\cdot\rangle” refers to the standard inner product between two matrices, defined as ⟨M,N⟩:=Tr⁡(MTN)\langle M,N\rangle:= \operatorname{Tr}(M^{\mathsf{T}}N), while “∥⋅∥F\| \cdot\|_{\mathrm{F}}” yields the matrix Frobenius norm as ∥M∥F:=⟨M,M⟩\|M\|_{\mathrm{F}} := \sqrt{\langle M,M\rangle}.

Algorithmic Developments

In the sequel, we expound on our algorithmic developments from three aspects: search directions, initial trial step sizes, and LM algorithm and criterion.

Search Directions

For the atomic part, we simply take the steepest descent directions, i.e.,

Datom(k):=Fatom(k) for all k≥0.D_{\mathrm{atom}}^{(k)} := F_{\mathrm{atom}}^{(k)} \text{ for all } k \ge0.

As for the lattice part, we first note that the steepest descent direction Flatt(k)F_{\mathrm{latt}}^{(k)} (see Eq. (1)) cannot be used directly due to the volume constraint; otherwise, there can be unexpected large variations in the lattice vectors, resulting in drastic fluctuations in the potential energies. For an illustration in two-dimensional context, see the left panel of Figure 2 where one can observe a large distance between the last iterate A(k)A^{(k)} (denoted by the black square) and the next iterate using Flatt(k)F_{\mathrm{latt}}^{(k)} (denoted by the black circle).

Instead, we search along the tangent lines of the feasible region at the current configuration. In mathematics, this amounts to projecting the lattice force Flatt(k)F_{\mathrm{latt}}^{(k)} onto the tangent space T(k)⊆R3×3\mathcal{T}^{(k)} \subseteq\mathbb{R}^{3 \times3} [29] of F:={A∈R3×3:det⁡(A)=V}\mathcal{F}:=\{A \in\mathbb{R}^{3 \times3}:\det(A)=V\} at A(k)A^{(k)}, where

T(k):={D∈R3×3:⟨B(k),D⟩=0}.(2)\mathcal{T}^{(k)} := \left\{D \in\mathbb{R}^{3 \times3} : \left\langle B^{(k)},D\right\rangle= 0\right\}. \tag*{(2)}

The projection favors a closed-form expression:

F~latt(k):=Flatt(k)−⟨B(k),Flatt(k)⟩∥B(k)∥F2B(k).(3)\widetilde{F}_{\mathrm{latt}}^{(k)} := F_{\mathrm{latt}}^{(k)}-\frac{\left\langle B^{(k)},F_{\mathrm{latt}}^{(k)}\right\rangle}{\left\|B^{(k)}\right\|_{\mathrm{F}}^{2}}B^{(k)}. \tag*{(3)}

We then take Dlatt(k):=F~latt(k)D_{\mathrm{latt}}^{(k)}:=\widetilde{F}_{\mathrm{latt}}^{(k)} for k≥0k \ge0. A comparison between the iterations with and without the gradient projection (3) can be found in the left panel of Figure 2. After projecting Flatt(k)F_{\mathrm{latt}}^{(k)} (denoted by the black arrow) onto T(k)\mathcal{T}^{(k)}, we obtain F~latt(k)\widetilde{F}_{\mathrm{latt}}^{(k)} (denoted by the red arrow) and then an iterate closer to A(k)A^{(k)} (denoted by the red circle).

It can be shown theoretically that the gradient projection (3) eliminates the potential first-order energy increment induced by the scaling operation (S4 in Figure 1). We shall point out that the strategy does not necessarily apply to general nonconvex constrained optimization problems. The desired theoretical results are achieved by fully exploiting the algebraic characterization of the tangent space (2). For technical details, please refer to the proof of Lemma 2 in Appendix § 1.6. We also demonstrate the merit of the gradient projection (3) through numerical comparisons; see the right panel of Figure 2.

Initial Trial Step Sizes

Efficient initial trial step sizes are essential for improving the overall performance. Moreover, our extensive numerical tests reveal that the relative displacements in the lattice vectors are typically much smaller than those in the atomic positions during relaxation, particularly for large systems. This observation underscores the substantial differences in the local curvatures of PES with respect to the atomic positions and lattice vectors, or in other words, ill conditions underlying the relaxation tasks. In light of this, we shall treat the atomic and lattice components separately and leverage their respective curvature information.

Motivated by the above discussions, we introduce curvature-aware step sizes for the atomic and lattice components, respectively. For the atomic part, we adopt alternating Barzilai-Borwein (ABB) step sizes [23, 30, 31], defined as

αatom(k−1,0)=αatom,ABB(k):={αatom,BB1(k),if mod⁡(k,2)=0;αatom,BB2(k),if mod⁡(k,2)=1.(4)\alpha_{\mathrm{atom}}^{(k-1,0)}=\alpha_{\mathrm{atom,ABB}}^{(k)}:=\begin{cases}\alpha_{\mathrm{atom,BB1}}^{(k)},&\text{if }\operatorname{mod}(k,2)=0;\\\alpha_{\mathrm{atom,BB2}}^{(k)},&\text{if }\operatorname{mod}(k,2)=1.\end{cases} \tag*{(4)}

for k≥1k \ge1, where

αatom,BB1(k):=∥Satom(k−1)∥F2⟨Satom(k−1),Yatom(k−1)⟩,αatom,BB2(k):=⟨Satom(k−1),Yatom(k−1)⟩∥Yatom(k−1)∥F2,\alpha_{\mathrm{atom,BB1}}^{(k)}:=\frac{\left\|S_{\mathrm{atom}}^{(k-1)}\right\|_{\mathrm{F}}^{2}}{\left\langle S_{\mathrm{atom}}^{(k-1)},Y_{\mathrm{atom}}^{(k-1)}\right\rangle},\quad\alpha_{\mathrm{atom,BB2}}^{(k)}:=\frac{\left\langle S_{\mathrm{atom}}^{(k-1)},Y_{\mathrm{atom}}^{(k-1)}\right\rangle}{\left\|Y_{\mathrm{atom}}^{(k-1)}\right\|_{\mathrm{F}}^{2}},

Satom(k−1):=R(k)−R(k−1)S_{\mathrm{atom}}^{(k-1)}:=R^{(k)}-R^{(k-1)}, and Yatom(k−1):=Fatom(k−1)−Fatom(k)Y_{\mathrm{atom}}^{(k-1)}:=F_{\mathrm{atom}}^{(k-1)}-F_{\mathrm{atom}}^{(k)}. It is not difficult to verify that

αatom,BB1(k)=arg⁡min⁡α∥αSatom(k−1)−Yatom(k−1)∥F,\alpha_{\mathrm{atom},\mathrm{BB1}}^{(k)}=\arg\min_{\alpha}\left\|\alpha S_{\mathrm{atom}}^{(k-1)}-Y_{\mathrm{atom}}^{(k-1)}\right\|_{\mathrm{F}},
αatom,BB2(k)=arg⁡min⁡α∥αYatom(k−1)−Satom(k−1)∥F.\alpha_{\mathrm{atom},\mathrm{BB2}}^{(k)}=\arg\min_{\alpha}\left\|\alpha Y_{\mathrm{atom}}^{(k-1)}-S_{\mathrm{atom}}^{(k-1)}\right\|_{\mathrm{F}}.

In other words, αatom,BB1(k)\alpha_{\mathrm{atom},\mathrm{BB1}}^{(k)} and αatom,BB2(k)\alpha_{\mathrm{atom},\mathrm{BB2}}^{(k)} approximate the inverse atomic Hessian of potential energy around R(k)R^{(k)}. The gradient descent with the above BB step sizes thus resembles the Newton iteration for solving the nonlinear equations Fatom=0F_{\mathrm{atom}}=0.

Now let us derive the ABB step sizes for the lattice part. Due to the volume constraint, the equilibrium condition on the lattice becomes Σdev=0\Sigma_{\mathrm{dev}}=0, namely, the shear stress vanishes and the normal stress cancels out the external isotropic pressure. This is equivalent to F~latt=0\widetilde{F}_{\mathrm{latt}}=0 in view of Eqs. (1) and (3), provided that the atomic forces vanish. Recalling the Newton iteration for solving the nonlinear equations F~latt=0\widetilde{F}_{\mathrm{latt}}=0, we obtain the following ABB step sizes for the lattice vectors

αlatt(k−1,0)=αlatt,ABB(k):={αlatt,BB1(k),if mod(k,2)=0;αlatt,BB2(k),if mod(k,2)=1(5)\alpha_{\mathrm{latt}}^{(k-1,0)}=\alpha_{\mathrm{latt},\mathrm{ABB}}^{(k)}:= \begin{cases} \alpha_{\mathrm{latt},\mathrm{BB1}}^{(k)}, & \text{if mod}(k,2)=0;\\ \alpha_{\mathrm{latt},\mathrm{BB2}}^{(k)}, & \text{if mod}(k,2)=1 \end{cases} \tag*{(5)}

for k≥1k\ge1, where

αlatt,BB1(k):=∥Slatt(k−1)∥F2⟨Slatt(k−1),Ylatt(k−1)⟩F,αlatt,BB2(k):=⟨Slatt(k−1),Ylatt(k−1)⟩F∥Ylatt(k−1)∥F2,\alpha_{\mathrm{latt},\mathrm{BB1}}^{(k)}:=\frac{\left\|S_{\mathrm{latt}}^{(k-1)}\right\|_{\mathrm{F}}^{2}}{\left\langle S_{\mathrm{latt}}^{(k-1)},Y_{\mathrm{latt}}^{(k-1)}\right\rangle_{\mathrm{F}}},\quad \alpha_{\mathrm{latt},\mathrm{BB2}}^{(k)}:=\frac{\left\langle S_{\mathrm{latt}}^{(k-1)},Y_{\mathrm{latt}}^{(k-1)}\right\rangle_{\mathrm{F}}}{\left\|Y_{\mathrm{latt}}^{(k-1)}\right\|_{\mathrm{F}}^{2}},

Slatt(k−1):=A(k)−A(k−1)S_{\mathrm{latt}}^{(k-1)}:=A^{(k)}-A^{(k-1)}, and Ylatt(k−1):=F~latt(k−1)−F~latt(k)Y_{\mathrm{latt}}^{(k-1)}:=\widetilde{F}_{\mathrm{latt}}^{(k-1)}-\widetilde{F}_{\mathrm{latt}}^{(k)}. With these distinct step sizes, we update the trials as

Rtrial(k,ℓ):=R(k)+αatom(k−1,ℓ)Datom(k,ℓ),R_{\mathrm{trial}}^{(k,\ell)}:=R^{(k)}+\alpha_{\mathrm{atom}}^{(k-1,\ell)}D_{\mathrm{atom}}^{(k,\ell)},
Ainter(k,ℓ):=A(k)+αlatt(k−1,ℓ)Dlatt(k,ℓ),A_{\mathrm{inter}}^{(k,\ell)}:=A^{(k)}+\alpha_{\mathrm{latt}}^{(k-1,\ell)}D_{\mathrm{latt}}^{(k,\ell)},
Atrial(k,ℓ):=Vdet⁡(Ainter(k,ℓ))3Ainter(k,ℓ),ℓ=0,1,….(6)A_{\mathrm{trial}}^{(k,\ell)}:=\sqrt[3]{\frac{V}{\det\left(A_{\mathrm{inter}}^{(k,\ell)}\right)}}A_{\mathrm{inter}}^{(k,\ell)},\qquad\ell=0,1,\ldots. \tag*{(6)}

As shown in Figure 3, we demonstrate the benefits brought by the distinct ABB step sizes (4) and (5) through the relaxation of silicon supercells. From the left panel, one can observe the considerable acceleration of the distinct version over the shared one. In the latter case, the ABB step sizes are shared and computed by the quantities from both the atomic and lattice components. In the right panel, we test our new algorithm equipped with the distinct step sizes on the silicon supercells of different sizes (1×1×1∼8×1×11\times1\times1\sim8\times1\times1). Compared with that of the CG built in off-the-shelf simulation software, the number of solving the KS equation of the new algorithm increases slowly as the system size grows, which indicates a preconditioning-like effect of our strategy.

LM Algorithm and Criterion

With the search directions in place, the LM algorithm starts from the initial step sizes and attempts to locate an appropriate trial configuration that meets certain LM criterion. Capturing the local PES information, however, the ABB step sizes can occasionally incur small increments on energies [23]. Extensive numerical simulations suggest that keeping the ABB step sizes intact is more beneficial [23, 32]. The monotone LM criterion (S5 in Figure 1) adopted in the widely used LM-based methods will lead to unnecessary computational expenditure and undermine the merits of the ABB step sizes. Notably, the authors of [23] propose a nonmonotone criterion which encompasses only atomic degrees of freedom.

In the following, a nonrestrictive nonmonotone LM criterion is presented for both atomic and lattice degrees of freedom and the simple backtracking is taken as the LM algorithm. Specifically, the nonmonotone criterion accepts the trial configuration if

Etrial(k,ℓ)≤Eˉ(k)−η(αatom(k−1,ℓ)∥Fatom(k)∥F2+αlatt(k−1,ℓ)∥F~latt(k)∥F2),(7)E_{\mathrm{trial}}^{(k,\ell)} \le\bar{E}^{(k)}-\eta\left(\alpha_{\mathrm{atom}}^{(k-1,\ell)}\left\|F_{\mathrm{atom}}^{(k)}\right\|_{F}^{2}+\alpha_{\mathrm{latt}}^{(k-1,\ell)}\left\|\widetilde{F}_{\mathrm{latt}}^{(k)}\right\|_{F}^{2}\right), \tag*{(7)}

where Etrial(k,ℓ):=E(Rtrial(k,ℓ),Atrial(k,ℓ))E_{\mathrm{trial}}^{(k,\ell)}:=E\left(R_{\mathrm{trial}}^{(k,\ell)},A_{\mathrm{trial}}^{(k,\ell)}\right), η∈(0,1)\eta\in(0,1) is a constant, and {Eˉ(k)}\{\bar{E}^{(k)}\} is a surrogate sequence, defined recursively as

Eˉ(k+1):=Eˉ(k)+μ(k)q(k)E(k+1)1+μ(k)q(k),(8)\bar{E}^{(k+1)}:=\frac{\bar{E}^{(k)}+\mu^{(k)}q^{(k)}E^{(k+1)}}{1+\mu^{(k)}q^{(k)}}, \tag*{(8)}
q(k+1):=μ(k)q(k)+1,q^{(k+1)}:=\mu^{(k)}q^{(k)}+1,

with Eˉ(0):=E(0)\bar{E}^{(0)}:=E^{(0)}, q(0):=1q^{(0)}:=1, and μ(k)∈[0,1]\mu^{(k)}\in[0,1]. The existence of αatom(k−1,ℓ)\alpha_{\mathrm{atom}}^{(k-1,\ell)} and αlatt(k−1,ℓ)\alpha_{\mathrm{latt}}^{(k-1,\ell)} satisfying Eq. (7) is proved in Lemma 2 of Appendix A. It is not hard to verify that Eˉ(k)\bar{E}^{(k)} is a weighted average of the past energies {E(j)}j=0k\{E^{(j)}\}_{j=0}^{k} and is always larger than E(k)E^{(k)}. Hence, small energy increments are tolerable. That being said, the criterion (7) is sufficient to ensure the convergence of the entire procedure; see Theorem 1 in Appendix A.

By virtue of the nonmonotone nature of the criterion (7), we take the simple backtracking as the LM algorithm. Whenever the last trial configuration is rejected, we set

αatom(k−1,ℓ):=αatom(k−1,ℓ−1)δatom,αlatt(k−1,ℓ):=αlatt(k−1,ℓ−1)δlatt(9)\alpha_{\mathrm{atom}}^{(k-1,\ell)}:=\alpha_{\mathrm{atom}}^{(k-1,\ell-1)}\delta_{\mathrm{atom}},\quad\alpha_{\mathrm{latt}}^{(k-1,\ell)}:=\alpha_{\mathrm{latt}}^{(k-1,\ell-1)}\delta_{\mathrm{latt}} \tag*{(9)}

where δatom,δlatt∈(0,1)\delta_{\mathrm{atom}},\delta_{\mathrm{latt}}\in(0,1) are constants. Without carefully selected δatom\delta_{\mathrm{atom}} and δlatt\delta_{\mathrm{latt}}, we observe in most simulations that the ABB step sizes can directly meet the nonrestrictive criterion (7) and the rejected trials account for approximately 1.8% of computational costs, in stark contrast to approximately 58.8% in the widely used CG.

Implementation Details

With the aforementioned developments, we have devised a projected gradient descent algorithm called PANBB for crystal structure relaxation under a fixed unit cell volume. More specifically, for the initial trial step sizes, we set αatom(−1,0)=4.8×10−2 A˚2/eV\alpha_{\mathrm{atom}}^{(-1,0)}=4.8\times10^{-2}\,\text{\AA}^{2}/\text{eV}, αlatt(−1,0)=10−6 A˚2/eV\alpha_{\mathrm{latt}}^{(-1,0)}=10^{-6}\,\text{\AA}^{2}/\text{eV} during the first iteration [33] and compute in the subsequent iterations the truncated absolute values of the ABB step sizes:

αatom(k−1,0)=max⁡{min⁡{∣αatom,ABB(k)∣,τatom(k)},10−5},(10)\alpha_{\mathrm{atom}}^{(k-1,0)}=\max\left\{\min\left\{\left|\alpha_{\mathrm{atom,ABB}}^{(k)}\right|,\tau_{\mathrm{atom}}^{(k)}\right\},10^{-5}\right\}, \tag*{(10)}
αlatt(k−1,0)=max⁡{min⁡{∣αlatt,ABB(k)∣,τlatt(k)},10−7},(11)\alpha_{\mathrm{latt}}^{(k-1,0)}=\max\left\{\min\left\{\left|\alpha_{\mathrm{latt,ABB}}^{(k)}\right|,\tau_{\mathrm{latt}}^{(k)}\right\},10^{-7}\right\}, \tag*{(11)}

where

τatom(k):=γatom(k)max⁡{−log⁡10(∥Fatom(k)∥F/N),1},\tau_{\mathrm{atom}}^{(k)}:=\gamma_{\mathrm{atom}}^{(k)}\max\left\{-\log_{10}\left(\left\|F_{\mathrm{atom}}^{(k)}\right\|_{F}/N\right),1\right\},
τlatt(k):=γlatt(k)max⁡{−log⁡10(∥F~latt(k)∥F/N),1}.\tau_{\mathrm{latt}}^{(k)}:=\gamma_{\mathrm{latt}}^{(k)}\max\left\{-\log_{10}\left(\left\|\widetilde{F}_{\mathrm{latt}}^{(k)}\right\|_{F}/N\right),1\right\}.

The factors γatom(k),γlatt(k)\gamma_{\mathrm{atom}}^{(k)},\gamma_{\mathrm{latt}}^{(k)} take respectively the values of 1,10−31,10^{-3} for k=0k=0 and later are modified according to the previous iterations: if they appear to be stringent from the recent history for αatom(k−1,0)\alpha_{\mathrm{atom}}^{(k-1,0)} and αlatt(k−1,0)\alpha_{\mathrm{latt}}^{(k-1,0)} to pass the criterion (7), we increase them; otherwise, we decrease them. For instance,

γatom(k):={2γatom(k−1),Case I;0.5γatom(k−1),Case II;γatom(k−1),otherwise,\gamma_{\mathrm{atom}}^{(k)}:= \begin{cases} 2\gamma_{\mathrm{atom}}^{(k-1)}, & \text{Case I;}\\ 0.5\gamma_{\mathrm{atom}}^{(k-1)}, & \text{Case II;}\\ \gamma_{\mathrm{atom}}^{(k-1)}, & \text{otherwise,} \end{cases}

where

  • Case I: the truncation has been active but αatom(k−1,0)\alpha_{\mathrm{atom}}^{(k-1,0)} has directly passed the criterion (7) for twice in the past min⁡{k−k^,20}\min\{k-\hat{k},20\} iterations;

  • Case II: αatom(k−1,0)\alpha_{\mathrm{atom}}^{(k-1,0)} has failed to pass the criterion (7) for twice in the past min⁡{k−k^,20}\min\{k-\hat{k},20\} iterations.

Initially, we set k^:=0\hat{k}:=0; whenever the factor γatom(k)\gamma_{\mathrm{atom}}^{(k)} gets changed, we set k^:=k\hat{k}:=k. The update on γlatt(k)\gamma_{\mathrm{latt}}^{(k)} follows similar rules. Note that the box boundaries in Eq. (10), such as 10−510^{-5} and 1010 for αatom(k−1,0)\alpha_{\mathrm{atom}}^{(k-1,0)}, are present for the ease of theoretical analysis. In practice, these boundaries are rarely touched. The criterion (7) is equipped with η=10−4\eta=10^{-4} and μ(k)=0.05\mu^{(k)}=0.05 following [23]. We take δatom=0.1\delta_{\mathrm{atom}}=0.1 and δlatt=0.5\delta_{\mathrm{latt}}=0.5 for the backtracking (9) [34]. Under mild conditions, we establish the convergence of PANBB to equilibrium states; see Theorem 1 in Appendix A.

Numerical Simulations

Experimental Settings

We have implemented PANBB in the in-house plane-wave code \textsc{CESSP} [35, 36, 37, 38]. The exchange-correlation energy is described by the generalized gradient approximation [39]. Electron-ion interactions are treated with the projector augmented-wave (PAW) potentials based on the open-source ABINIT Jollet-Torrent-Holzwarth data set library in the PAW-XML format [40]. The total energies are calculated using the Monkhorst-Pack mesh [41] with a kk-mesh spacing of 0.12∼0.200.12\sim0.20 \AA−1^{-1}. The typical plane-wave energy cutoffs are around 500∼600500\sim600 eV. The KS equations are solved by the preconditioned self-consistent field (SCF) iteration [37].

In addition to PANBB, we test the performance of the CG and DIIS built in \textsc{CESSP}. Notably, compared with the one documented in textbook, the CG in test uses deviatoric stress tensors in search direction updates for the fixed-volume relaxation. We terminate all the three methods whenever the number of solving the Kohn-Sham equations (abbreviated as #KS) reaches 1000 or the maximum atomic force and the maximum stress components divided by NN fall below 0.01 eV/\AA. The convergence criterion for the SCF iteration is 10−510^{-5} eV. To evaluate the performance of the three methods, we record the #KS and running time in seconds (abbreviated as T) as required for fulfilling the force tolerance.

Benchmark Test

We have performed numerical comparisons among CG, DIIS, and PANBB on a benchmark set consisting of 223 systems from various categories, such as metallic systems, organic molecules, semiconductors, ABX3ABX_{3} perovskites, heterostructures, and 2D materials. About 68.6% of them are metallic systems, for which the EOS calculations are more demanding. The number of atoms in one system ranges from 2 to 215. Some of them are available from the Materials Project [42] and the Organic Materials Database [43]. In the beginning, these systems are at their ideal configurations for defects, heterostructures, and substitutional alloys, or simply place a molecule on top of a surface. More information about the benchmark set can be found in the Supplemental Material.

To provide a comprehensive comparison among CG, DIIS, and PANBB over the benchmark set, we adopt the performance profile [44] as shown in Figure 4, where #KSmin⁡_{\min} and Tmin⁡_{\min} represent the minimum #KS and T required by the three methods on a given system, respectively. For example, in the right panel of Figure 4, the intercepts of the three curves indicate that CG, DIIS, and PANBB are the fastest on approximately 11.3%, 28.9%, and 59.8% of all the systems, respectively. The yy-coordinate touched by the curve of PANBB when the xx-coordinate equals two shows that the running time of PANBB is not twice larger than those of the other two methods on approximately 96.5% of all the systems. Basically, a larger area below the curve implies better overall performance of the associated method. We can therefore conclude from Figure 4 that PANBB favors the best overall performance. In addition, the Supplemental Material reveals that PANBB is respectively faster than CG and DIIS on approximately 85.2% and 68.8% of all the systems. The average speedup factors of PANBB over CG and DIIS, in terms of running time, are around 1.41 and 1.45, respectively. We also calculate the average speedup factors of PANBB over CG by the system category; see Figure 5 for an illustration. Among others, the acceleration brought by PANBB is the most evident when relaxing organic molecules, ABX3ABX_{3} perovskites, and metallic systems, with speedup factors of 1.67, 1.50, and 1.40, respectively, in terms of running time. These statistics provide strong evidence for the universal efficiency superiority of PANBB. The robustness of PANBB also deserves a whistle. Thanks to its theoretical convergence guarantees (see Theorem 1 in Appendix § 1.6), PANBB converges consistently across the benchmark set, whereas CG fails on 11 systems due to the breakdown of the LM algorithm, and DIIS diverges on 56 systems.

Figure 4. The performance profiles of CG, DIIS, and PANBB, where #KSmin⁡_{\min} and Tmin⁡_{\min} denote the minimum #KS and T required by the three methods on a given system, respectively. To reach a fair comparison and illustrate the algorithm robustness at the same time, we only retain (i) the systems where the difference between the converged energies obtained by each pair of methods does not exceed 3 meV per atom; (ii) the systems on which at least one method diverges or fails to converge before #KS exceeds 1000, and set the associated #KS and T to ∞\infty. The final number of the systems used for this figure is 204. Left: #KS. Right: T.

Figure 5. The average speedup factors of PANBB over CG by the system category. To achieve a fair comparison, we only include the systems where both PANBB and CG converge normally and the converged energy differences per atom are not larger than 3 meV. Categories whose numbers of the included systems are less than 5 are merged together into “other”. Left: #KS. Right: T.

Calculations of EOS on the AlCoCrFeNi HEA

As an application of crystal structure relaxation under a fixed unit cell volume, we employ PANBB to calculate the static EOS up to approximately 30 GPa for the body-centered cubic AlCoCrFeNi HEA. The local lattice distortion (LLD), induced mainly by the large atomic size mismatch of the alloy components, is one of the core effects responsible for the unprecedented mechanical behaviors of HEAs. The LLD in some equimolar HEAs under ambient pressure is studied by ab initio approach [24]. The volume-dependent LLD is investigated and is helpful for understanding the structural properties of HEAs under pressure. The thermodynamic properties have also been investigated in comparison with experimental data for validation.

To incorporate the temperature-induced magnetic structural transition, we need to describe both the ferromagnetic (FM) and paramagnetic (PM) states of AlCoCrFeNi. To characterize the inherent chemical and magnetic disorder, we combine the SAE method [45] and an integer programming approach [46] to construct representative 4×4×54 \times4 \times5 supercells with 160 atoms. Specifically, the SAE method is first invoked for chemical disorder modeling, where we seek to minimize

∑d(C2)<r2cw2f(C2,σ)+∑d(C3)<r3cw3f(C3,σ)\sum_{d(C_2)<r_2^c} w_2 f(C_2,\sigma) + \sum_{d(C_3)<r_3^c} w_3 f(C_3,\sigma)

with respect to the atomic occupation σ\sigma. Here wnw_n (n=2,3n = 2, 3) are appropriate weights and rncr_n^c (n=2,3n = 2, 3) are cutoff radii. The notations f(C2,σ)f(C_2,\sigma) and f(C3,σ)f(C_3,\sigma) represent the similarity functions associated with the diatomic clusters C2C_2 and triatomic clusters C3C_3, respectively. The corresponding diameter of the cluster is denoted by d(C2)d(C_2) or d(C3)d(C_3). We perform 90,000 Monte carlo samplings and minimize f(C2,σ)f(C_2,\sigma) to approximately 0.06 and f(C3,σ)f(C_3,\sigma) to approximately 0.18.

For magnetic disorder modeling, we refrain from using the SAE method again because it necessitates distinguishing spin-up and spin-down elements, leading to a higher-dimensional search space. Instead, we resort to an integer programming approach to distribute spin-up and spin-down sites uniformly in the configuration given by the SAE method, as a way to achieve the disordered local moment approximation [47]. We refer readers to Appendix B for more technical details. The generated atomic and magnetic configurations are illustrated in Figure 6. The rightmost panel displays the radial distribution function of the magnetic elements, demonstrating a proper description of the magnetic disorder by the integer programming approach.

Figure 6. The generated atomic and magnetic configurations. From left to right: the atomic configuration generated by the SAE method, the 64 spin-up and 64 spin-down sites chosen by the integer programming approach for magnetic configuration, and the radial distribution function of the magnetic elements.

To calculate the static EOS for both the FM and PM states, we distribute 18 unit cell volumes in the range between 10.29 A˚3\mathrm{\mathring{A}}^{3}/atom and 13.12 A˚3\mathrm{\mathring{A}}^{3}/atom, corresponding to the static pressures from approximately 29.4 GPa to −12.9-12.9 GPa. For the FM state, we employ PANBB for the fixed-volume relaxation with colinear calculations. Following [25], the PM state is then described based on the relaxed FM structures, with the initial magnetic configuration determined by the integer programming approach. The energy-volume relations on both the FM and PM AlCoCrFeNi are depicted in Figure 7, along with

Figure 7. The energy-volume relations on FM and PM AlCoCrFeNi. “DFT” indicates the ab initio results, while “Fit” is for the curve fitted to the static BM3 EOS. The red circles and solid line represent the results for the FM state, while the black squares and dotted line represent the results for the PM state.

the fitted third-order Birch-Murnaghan (BM3) EOS [48, 49]. We record the fitted parameters in Table I, where V0V_0 represents the equilibrium volume and E0E_0, B0B_0, and B0′B'_0 denotes the energy, bulk modulus, and its derivative with respect to pressure associated with V0V_0, respectively.

ReferencesV0V_0 (A˚3\mathrm{\mathring{A}}^3/atom)E0E_0 (eV/atom)B0B_0 (GPa)B0′B'_0
FM ([25])11.77-6.85157.328.9
FM (PANBB)11.74-6.86158.495.3
FM (CG)11.73-6.86158.045.8
PM ([25])11.70-6.82161.645.3
PM (PANBB)11.65-6.84167.594.7
PM (CG)11.65-6.84161.644.0

Table I. The fitted parameters of the static BM3 EOS.

Compared with the previous calculations [25], our fitted V0V_0 and B0B_0 for both the FM and PM states, E0E_0 for the FM state, and B0′B'_0 for the PM state are comparable; see Table I. The gaps in B0′B'_0 for the FM state and E0E_0 for the PM state can result from two aspects: (i) we enhance magnetic disorder by the integer programming approach; (ii) [25] relaxes only the atomic positions, while PANBB optimizes both the atomic and lattice degrees of freedom. By the way, our value of B0′B'_0 appears to be more reasonable since it is closer to those obtained for similar compositions [24].

Within the mean-field approximation [50], we can estimate the Curie temperature of AlCoCrFeNi as

23(1−c)kB(E0PM−E0FM)≈210.8 K,\frac{2}{3(1-c)k_B}\left(E_0^{\mathrm{PM}}-E_0^{\mathrm{FM}}\right) \approx210.8\ \mathrm{K},

where cc is the concentration of nonmagnetic atoms, kBk_B is the Boltzmann’s constant, E0FME_0^{\mathrm{FM}} and E0PME_0^{\mathrm{PM}} are the energies of the FM and PM states associated with the equilibrium volumes from the BM3 EOS fitting based on the PANBB results, respectively; see Figure 7. Our estimated Curie temperature lies in the interval of the critical temperature of Alx_xCoCrFeNi [51], which is between 200 K (x=0.75x = 0.75) and 400 K (x=1.25x = 1.25), to some extent demonstrating rational descriptions of the FM and PM states.

To construct the EOS in the temperature (TT) and pressure (PP) space, the modified mean field potential (MMFP) approach [18] is used to describe the ion vibration energy as

FvibMMFP=−kBT[32ln⁡MkBT2πℏ2+ln⁡vf(V,T)],vf(V,T)=4π∫exp⁡[p(r,V)kBT]r2 dr,p(r,V)=12[Es(D+r)+Es(D−r)−2Es(D)]+λr2D[Es(D+r)−Es(D−r)],\begin{aligned} F_{\mathrm{vib}}^{\mathrm{MMFP}} & = -k_B T \left[\frac{3}{2}\ln\frac{M k_B T}{2\pi\hbar^2}+\ln v_f(V,T)\right], \\ v_f(V,T) & = 4\pi\int\exp\left[\frac{p(r,V)}{k_B T}\right]r^2\,\mathrm{d}r, \\ p(r,V) & = \frac{1}{2}\left[E_s(D+r)+E_s(D-r)-2E_s(D)\right] \\ & \quad+\frac{\lambda r}{2D}\left[E_s(D+r)-E_s(D-r)\right], \end{aligned}

where ℏ\hbar denotes the Planck constant, MM is the relative atomic mass, DD is the nearest atomic distance, rr corresponds to the distance of the ion vibration away from its equilibrium position, EsE_s is the static EOS, and pp is the constructed potential. Here, λ\lambda can take three values, −1-1, 00, 11, which originate from different forms of the Grüneisen coefficient. As a powerful tool to calculate the Helmholtz free energy, the MMFP effectively describes the ion vibration with the anharmonic effect, and has been widely used in thermodynamic physical properties simulations of HEAs [25, 45].

Since the effectiveness of the MMFP approach highly depends on the potential field constructed from the static EOS, we can validate the previous calculations by investigating the thermodynamic properties of AlCoCrFeNi. We first calculate the equilibrium volume and bulk modulus at ambient conditions with the MMFP approach using the PM PANBB results. Our values (approximately 11.83 A˚3\mathrm{\mathring{A}}^3/atom and 148.8 GPa) turn out to be close to those from the X-ray diffraction and diamond anvil cell experiments (11.79 A˚3\mathrm{\mathring{A}}^3/atom and 150±2.5150 \pm2.5 GPa) [26]. We then calculate the volume-pressure relation at 300 K with the PM PANBB results; see the solid line with label “300 K (MMFP)” in Figure 8. At equivalent pressures, the maximum relative deviation of the volumes from the diamond anvil cell data [26] does not exceed 0.7%.

Figure 8. The volume-pressure relation on AlCoCrFeNi at 0 K and 300 K. “Cheng exp.” represents the X-ray diffraction and diamond anvil cell experimental results [26], “0 K (FM static EOS)” refers to the relation by fitting the static EOS based on the FM PANBB results, “0 K (PM static EOS)” refers to the relation by fitting the static EOS based on the PM PANBB results, and “300 K (MMFP)” refers to the relation given by the MMFP approach based on the PM PANBB results.

We would also like to compare PANBB with CG on the relaxation of the FM AlCoCrFeNi. PANBB converges normally at all volumes, while CG is found to fail at 4 volumes due to the breakdown of the LM process. Moreover, a sharp jump of c/ac/a is found during the relaxation of CG when the pressure exceeds 21 GPa; see Figure 9. The results of CG at the other 16 volumes are fitted by the BM3 EOS. The fitted parameters of the static EOS of AlCoCrFeNi are summarized in Table I and consistent with those derived from the PANBB results. The average speedup factor of PANBB over CG in terms of running time is 1.36. Note that this speedup factor is calculated at the volumes where both methods converge normally and the energy differences per atom are not larger than 3 meV.

Figure 9. The axis length ratio-volume relation on the FM AlCoCrFeNi. For the volumes where CG fails to converge normally, we use the axis length ratios of the final structures. The axis length ratios “c/ac/a” and “b/ab/a” are indicated by solid and dotted lines, respectively. The results of CG and PANBB are indicated by square and circle markers, respectively.

By the way, we illustrate the LLD achieved by fixed-volume relaxation at pressures of approximately 0 GPa, 21 GPa, and 29 GPa using the smearing radial distribution function. This stands in sharp contrast to the peaks at coordination shells in the ideal lattice, as depicted in Figure 10. Notably, when the pressure does not exceed 21 GPa, the radial distribution functions of the structures relaxed by PANBB and CG display a high degree of consistency. However, as the pressure rises to 29 GPa, the radial distribution function obtained by PANBB exhibits smeared peaks consistent with those of the ideal lattice, while that obtained by CG shows significant displacements. This can be attributed to the previously mentioned issue of strong lattice distortion during the relaxation process of CG. Whether such lattice distortion conforms to physics deserves future investigations.

Figure 10. Radial distribution functions of the FM AlCoCrFeNi at pressures of approximately 0 GPa (left), 21 GPa (middle), and 29 GPa (right). The black solid lines refer to the functions before relaxation, while the blue dashed and red dotted lines represent the functions after relaxation using PANBB and CG, respectively.

Conclusions

In this work, we introduce PANBB, a novel force-based algorithm for relaxing crystal structures under a fixed unit cell volume. By incorporating the gradient projections, distinct curvature-aware initial trial step sizes, and a nonmonotone LM criterion, PANBB exhibits universal superiority over the widely used methods in both efficiency and robustness, as demonstrated across a benchmark test. To the best of our knowledge, PANBB is the first algorithm that guarantees theoretical convergence for the determinant-constrained optimization. Moreover, the application in the AlCoCrFeNi HEAs also showcases its potential in facilitating calculations of the wide-range EOS and thermodynamical properties for multicomponent alloys.

Acknowledgments

The work of the third author was supported in part by the National Natural Science Foundation of China under Grant No. U23A20537. The work of the fourth author was partially supported by the National Natural Science Foundation of China under Grants No. 12125108, No. 12226008, No. 11991021, No. 11991020, No. 12021001, and No. 12288201, Key Research Program of Frontier Sciences, Chinese Academy of Sciences under Grant No. ZDBS-LY-7022, CAS-Croucher Funding Scheme for Joint Laboratories “CAS AMSS-PolyU Joint Laboratory of Applied Mathematics: Nonlinear Optimization Theory, Algorithms and Applications”. The computations were done on the high-performance computers of State Key Laboratory of Scientific and Engineering Computing, Chinese Academy of Sciences and the clusters at Northwestern Polytechnical University. The authors thank Zhen Yang for the helpful discussions on the MMFP approach.

Convergence Analyses

In what follows, we rigorously establish the convergence of PANBB to equilibrium states regardless of the initial configurations. The analyses depend on the following two reasonable assumptions on the potential energy functional.

Assumption 1. The potential energy functional EE is locally Lipschitz continuously differentiable with respect to RR and AA. The stress tensor Σ\Sigma is continuous with respect to RR and AA.

Assumption 2. The potential energy functional EE is coercive over F\mathcal{F}, namely, E(R,A)→∞E(R,A) \to\infty as ∥R∥2→∞\lVert R\rVert_{2} \to\infty or ∥A∥2→∞\lVert A\rVert_{2} \to\infty with det⁡(A)=V\det(A)=V.

For brevity, we make the following notation

D~latt:=BTF~latt,(12)\widetilde{D}_{\mathrm{latt}} := B^{T}\widetilde{F}_{\mathrm{latt}}, \tag*{(12)}

where F~latt\widetilde{F}_{\mathrm{latt}} is defined in Eq. (3), and define the level set

L:={(R,A):E(R,A)≤E(0), det⁡(A)=V}.(13)\mathcal{L}:=\{(R,A):E(R,A)\le E^{(0)},\ \det(A)=V\}. \tag*{(13)}

In view of Eqs. (6) and (A1), Ainter(k,ℓ)A_{\mathrm{inter}}^{(k,\ell)} can be rewritten as

Ainter(k,ℓ)=A(k)(I+αlatt(k−1,ℓ)D~latt(k)).(14)A_{\mathrm{inter}}^{(k,\ell)}=A^{(k)}\left(I+\alpha_{\mathrm{latt}}^{(k-1,\ell)}\widetilde{D}_{\mathrm{latt}}^{(k)}\right). \tag*{(14)}

Assumptions 1 and 2 directly imply the following lemma.

Lemma 1. Suppose that Assumptions 1 and 2 hold. Then there exists an M>0M>0 such that if (R,A)∈L(R,A)\in\mathcal{L}, all of ∥R∥F\lVert R\rVert_{\mathrm{F}}, ∥A∥F\lVert A\rVert_{\mathrm{F}}, ∥B∥F\lVert B\rVert_{\mathrm{F}}, ∥Fatom∥F\lVert F_{\mathrm{atom}}\rVert_{\mathrm{F}}, and ∥D~latt∥F\lVert\widetilde{D}_{\mathrm{latt}}\rVert_{\mathrm{F}} are upper bounded by MM.

Proof. The existence of a uniform upper bound for ∥R∥F\lVert R\rVert_{\mathrm{F}} and ∥A∥F\lVert A\rVert_{\mathrm{F}} follows directly from Assumption 2 and the definition of the set L\mathcal{L} in Eq. (13). That for ∥B∥F\lVert B\rVert_{\mathrm{F}} is derived from det⁡(A)=V\det(A)=V. The uniform boundedness of ∥Fatom∥F\lVert F_{\mathrm{atom}}\rVert_{\mathrm{F}} and ∥D~latt∥F\lVert\widetilde{D}_{\mathrm{latt}}\rVert_{\mathrm{F}} are then deduced from Eqs. (1), (A1), and Assumption 1. ■

Let αˉatom\bar{\alpha}_{\mathrm{atom}}, αˉlatt\bar{\alpha}_{\mathrm{latt}}, α‾atom\underline{\alpha}_{\mathrm{atom}}, and α‾latt\underline{\alpha}_{\mathrm{latt}} denote the upper and lower bounds for the ABB step sizes in Eq. (10). Based on Assumptions 1 and 2, some positive constants are defined using MM in Lemma 1, δatom\delta_{\mathrm{atom}}, δlatt\delta_{\mathrm{latt}} in Eq. (9) as well as αˉatom\bar{\alpha}_{\mathrm{atom}} and αˉlatt\bar{\alpha}_{\mathrm{latt}} in Eq. (10).

L∇E:=sup⁡{∥∇E(R1,A1)−∇E(R2,A2)∥F∥(R1,A1)−(R2,A2)∥F:∥Ri∥F≤(1+αˉatom)M, ∥Ai∥F≤32M, i=1,2},Llatt:=sup⁡{∣E(R,A1)−E(R,A2)∣∥A1−A2∥F:∥Ri∥F≤(1+αˉatom)M, ∥Ai∥F≤3M, i=1,2},Mˉ:=487M3Llatt,ℓˉ:=⌈max⁡{log⁡L∇Eαˉatom2(1−η)log⁡1δatom,log⁡(L∇E+2Mˉ)αˉlatt2(1−η)log⁡1δlatt,log⁡(2Mαˉlatt)log⁡1δlatt}⌉.(15)\begin{aligned} L_{\nabla E}&:=\sup\left\{\frac{\lVert\nabla E(R_1,A_1)-\nabla E(R_2,A_2)\rVert_{\mathrm{F}}}{\lVert(R_1,A_1)-(R_2,A_2)\rVert_{\mathrm{F}}}:\lVert R_i\rVert_{\mathrm{F}}\le(1+\bar{\alpha}_{\mathrm{atom}})M,\ \lVert A_i\rVert_{\mathrm{F}}\le\frac{3}{2}M,\ i=1,2\right\},\\ L_{\mathrm{latt}}&:=\sup\left\{\frac{\lvert E(R,A_1)-E(R,A_2)\rvert}{\lVert A_1-A_2\rVert_{\mathrm{F}}}:\lVert R_i\rVert_{\mathrm{F}}\le(1+\bar{\alpha}_{\mathrm{atom}})M,\ \lVert A_i\rVert_{\mathrm{F}}\le3M,\ i=1,2\right\},\\ \bar{M}&:=\frac{48}{7}M^3L_{\mathrm{latt}},\qquad \bar{\ell}:=\left\lceil\max\left\{\frac{\log\frac{L_{\nabla E}\bar{\alpha}_{\mathrm{atom}}}{2(1-\eta)}}{\log\frac{1}{\delta_{\mathrm{atom}}}},\frac{\log\frac{(L_{\nabla E}+2\bar{M})\bar{\alpha}_{\mathrm{latt}}}{2(1-\eta)}}{\log\frac{1}{\delta_{\mathrm{latt}}}},\frac{\log(2M\bar{\alpha}_{\mathrm{latt}})}{\log\frac{1}{\delta_{\mathrm{latt}}}}\right\}\right\rceil. \tag*{(15)} \end{aligned}

Lemma 2 below asserts that the LM processes in PANBB are always finite.

Lemma 2. Suppose that Assumptions 1 and 2 hold. Assume that (R(k),A(k))∈L(R^{(k)},A^{(k)})\in\mathcal{L}. Then in the (k+1)(k+1)th iteration, the LM criterion (7) must be satisfied after invoking the backtracking (9) for at most ℓˉ\bar{\ell} times and (R(k+1),A(k+1))∈L(R^{(k+1)},A^{(k+1)})\in\mathcal{L}, where ℓˉ\bar{\ell} is defined in Eq. (15).

Proof. We prove this lemma by showing that there is a sufficient reduction from E(k)E^{(k)} to Etrial(k,ℓ)E_{\mathrm{trial}}^{(k,\ell)} provided that ℓ≥ℓˉ\ell\ge\bar{\ell}, which indicates the fulfillment of the criterion (7). We divide the proof into two steps.

Step 1. Estimate E(Rtrial(k,ℓ),Ainter(k,ℓ))−E(k)E(R_{\mathrm{trial}}^{(k,\ell)},A_{\mathrm{inter}}^{(k,\ell)})-E^{(k)}. It follows from Eqs. (2) and (6) that

⟨Fatom(k),Rtrial(k,ℓ)−R(k)⟩=−1αatom(k−1,ℓ)∥Rtrial(k,ℓ)−R(k)∥F2,(16)\left\langle F_{\mathrm{atom}}^{(k)},R_{\mathrm{trial}}^{(k,\ell)}-R^{(k)}\right\rangle=-\frac{1}{\alpha_{\mathrm{atom}}^{(k-1,\ell)}}\left\|R_{\mathrm{trial}}^{(k,\ell)}-R^{(k)}\right\|_{\mathrm{F}}^2, \tag*{(16)}
⟨Flatt(k),Ainter(k,ℓ)−A(k)⟩=−1αlatt(k−1,ℓ)∥Ainter(k,ℓ)−A(k)∥F2.(17)\left\langle F_{\mathrm{latt}}^{(k)},A_{\mathrm{inter}}^{(k,\ell)}-A^{(k)}\right\rangle=-\frac{1}{\alpha_{\mathrm{latt}}^{(k-1,\ell)}}\left\|A_{\mathrm{inter}}^{(k,\ell)}-A^{(k)}\right\|_{\mathrm{F}}^2. \tag*{(17)}

Since ℓ≥ℓˉ\ell\ge\bar{\ell}, it is not hard to show αlatt(k−1,ℓ)≤1/(2M)\alpha_{\mathrm{latt}}^{(k-1,\ell)}\le1/(2M). From this, αatom(k−1,ℓ)≤αˉatom\alpha_{\mathrm{atom}}^{(k-1,\ell)}\le\bar{\alpha}_{\mathrm{atom}}, and Lemma 1, we can infer αlatt(k−1,ℓ)∥D~latt(k)∥2≤1/2\alpha_{\mathrm{latt}}^{(k-1,\ell)}\|\widetilde{D}_{\mathrm{latt}}^{(k)}\|_2\le1/2 and

∥Rtrial(k,ℓ)∥F≤(1+αˉatom)M,(18)\|R_{\mathrm{trial}}^{(k,\ell)}\|_{\mathrm{F}}\le(1+\bar{\alpha}_{\mathrm{atom}})M, \tag*{(18)}
∥Ainter(k,ℓ)∥F≤(A3)∥A(k)∥F(1+αlatt(k−1,ℓ)∥D~latt(k)∥2)≤32M.(19)\|A_{\mathrm{inter}}^{(k,\ell)}\|_{\mathrm{F}}\overset{\mathrm{(A3)}}{\le}\|A^{(k)}\|_{\mathrm{F}}(1+\alpha_{\mathrm{latt}}^{(k-1,\ell)}\|\widetilde{D}_{\mathrm{latt}}^{(k)}\|_2)\le\frac{3}{2}M. \tag*{(19)}

By the definition of L∇EL_{\nabla E} in Eq. (15), one can derive from Eqs. (A5)–(19) that E(Rtrial(k,ℓ),Ainter(k,ℓ))−E(k)E(R_{\mathrm{trial}}^{(k,\ell)},A_{\mathrm{inter}}^{(k,\ell)})-E^{(k)} is upper bounded by

[L∇E2−1αatom(k−1,ℓ)]∥Rtrial(k,ℓ)−R(k)∥F2+[L∇E2−1αlatt(k−1,ℓ)]∥Ainter(k,ℓ)−A(k)∥F2.(20)\left[\frac{L_{\nabla E}}{2}-\frac{1}{\alpha_{\mathrm{atom}}^{(k-1,\ell)}}\right]\|R_{\mathrm{trial}}^{(k,\ell)}-R^{(k)}\|_{\mathrm{F}}^2+ \left[\frac{L_{\nabla E}}{2}-\frac{1}{\alpha_{\mathrm{latt}}^{(k-1,\ell)}}\right]\|A_{\mathrm{inter}}^{(k,\ell)}-A^{(k)}\|_{\mathrm{F}}^2. \tag*{(20)}

Step 2. Estimate Etrial(k,ℓ)−E(Rtrial(k,ℓ),Ainter(k,ℓ))E_{\mathrm{trial}}^{(k,\ell)}-E(R_{\mathrm{trial}}^{(k,\ell)},A_{\mathrm{inter}}^{(k,\ell)}).

First notice from Eq. (14) that det⁡(Ainter(k,ℓ))=Vdet⁡(I+αlatt(k−1,ℓ)D~latt(k))\det(A_{\mathrm{inter}}^{(k,\ell)})=V\det(I+\alpha_{\mathrm{latt}}^{(k-1,\ell)}\widetilde{D}_{\mathrm{latt}}^{(k)}). Due to αlatt(k−1,ℓ)∥D~latt(k)∥2≤1/2\alpha_{\mathrm{latt}}^{(k-1,\ell)}\|\widetilde{D}_{\mathrm{latt}}^{(k)}\|_2\le1/2 and the fact that the modulus of any eigenvalue is not larger than the maximum singular value, one has

det⁡(Ainter(k,ℓ))∈[18V,278V].(21)\det(A_{\mathrm{inter}}^{(k,\ell)})\in\left[\frac{1}{8}V,\frac{27}{8}V\right]. \tag*{(21)}

By the definition of the scaling operation (see S4 in Figure 1), Eqs. (19) and (21), ∥Atrial(k,ℓ)∥F≤3M\|A_{\mathrm{trial}}^{(k,\ell)}\|_{\mathrm{F}}\le3M. From this,

Etrial(k,ℓ)−E(Rtrial(k,ℓ),Ainter(k,ℓ))≤(A4),(A6)Llatt∥Atrial(k,ℓ)−Ainter(k,ℓ)∥F=Llatt∣det⁡(Ainter(k,ℓ))3−V3∣∥Ainter(k,ℓ)∥Fdet⁡(Ainter(k,ℓ))3.(22)\begin{aligned} E_{\mathrm{trial}}^{(k,\ell)}-E(R_{\mathrm{trial}}^{(k,\ell)},A_{\mathrm{inter}}^{(k,\ell)}) &\overset{\mathrm{(A4),(A6)}}{\le}L_{\mathrm{latt}}\|A_{\mathrm{trial}}^{(k,\ell)}-A_{\mathrm{inter}}^{(k,\ell)}\|_{\mathrm{F}}\\ &=L_{\mathrm{latt}}\left|\sqrt[3]{\det(A_{\mathrm{inter}}^{(k,\ell)})}-\sqrt[3]{V}\right| \frac{\|A_{\mathrm{inter}}^{(k,\ell)}\|_{\mathrm{F}}}{\sqrt[3]{\det(A_{\mathrm{inter}}^{(k,\ell)})}}. \tag*{(22)} \end{aligned}

For one thing, in view of Eqs. (19) and (21),

∥Ainter(k,ℓ)∥Fdet⁡(Ainter(k,ℓ))3≤3MV3.(23)\frac{\|A_{\mathrm{inter}}^{(k,\ell)}\|_{\mathrm{F}}}{\sqrt[3]{\det(A_{\mathrm{inter}}^{(k,\ell)})}}\le\frac{3M}{\sqrt[3]{V}}. \tag*{(23)}

For another, again by Eq. (21),

∣det⁡(Ainter(k,ℓ))3−V3∣=∣det⁡(Ainter(k,ℓ))−V∣det⁡(Ainter(k,ℓ))2/3+det⁡(Ainter(k,ℓ))1/3V1/3+V2/3≤4∣det⁡(Ainter(k,ℓ))−V∣7V2/3.(24)\left|\sqrt[3]{\det(A_{\mathrm{inter}}^{(k,\ell)})}-\sqrt[3]{V}\right| =\frac{|\det(A_{\mathrm{inter}}^{(k,\ell)})-V|}{\det(A_{\mathrm{inter}}^{(k,\ell)})^{2/3}+\det(A_{\mathrm{inter}}^{(k,\ell)})^{1/3}V^{1/3}+V^{2/3}} \le\frac{4|\det(A_{\mathrm{inter}}^{(k,\ell)})-V|}{7V^{2/3}}. \tag*{(24)}

Let λ1,λ2,λ3\lambda_1,\lambda_2,\lambda_3 be the eigenvalues of αlatt(k−1,ℓ)D~latt(k)\alpha_{\mathrm{latt}}^{(k-1,\ell)}\widetilde{D}_{\mathrm{latt}}^{(k)}. Since F~latt(k)\widetilde{F}_{\mathrm{latt}}^{(k)} lies on T(k)\mathcal{T}^{(k)} (see the definition in Eq. (2)), one has ∑i=13λi=0\sum_{i=1}^{3}\lambda_i=0 from Eq. (12). Therefore,

∣det⁡(Ainter(k,ℓ))−V∣=(A3)V∣det⁡(I+αlatt(k−1,ℓ)D~latt(k))−1∣=V∣λ1λ2+λ1λ3+λ2λ3+λ1λ2λ3∣.\begin{aligned} |\det(A_{\mathrm{inter}}^{(k,\ell)})-V| &\overset{\mathrm{(A3)}}{=}V\left|\det(I+\alpha_{\mathrm{latt}}^{(k-1,\ell)}\widetilde{D}_{\mathrm{latt}}^{(k)})-1\right|\\ &=V|\lambda_1\lambda_2+\lambda_1\lambda_3+\lambda_2\lambda_3+\lambda_1\lambda_2\lambda_3|. \end{aligned}

Since max⁡i{∣λi∣}≤αlatt(k−1,ℓ)∥D~latt(k)∥2≤1/2\max_i\{|\lambda_i|\}\le\alpha_{\mathrm{latt}}^{(k-1,\ell)}\|\widetilde{D}_{\mathrm{latt}}^{(k)}\|_2\le1/2,

∣det⁡(Ainter(k,ℓ))−V∣≤4Vmax⁡i{∣λi∣}2≤4V∥αlatt(k−1,ℓ)D~latt(k)∥22=(A3)4V∥B(k)T(Ainter(k,ℓ)−A(k))∥22≤4M2V∥Ainter(k,ℓ)−A(k)∥F2,(25)\begin{aligned} |\det(A_{\mathrm{inter}}^{(k,\ell)})-V| &\le4V\max_i\{|\lambda_i|\}^2\le4V\|\alpha_{\mathrm{latt}}^{(k-1,\ell)}\widetilde{D}_{\mathrm{latt}}^{(k)}\|_2^2\\ &\overset{\mathrm{(A3)}}{=}4V\|B^{(k)\mathsf{T}}(A_{\mathrm{inter}}^{(k,\ell)}-A^{(k)})\|_2^2\\ &\le4M^2V\|A_{\mathrm{inter}}^{(k,\ell)}-A^{(k)}\|_{\mathrm{F}}^2, \tag*{(25)} \end{aligned}

where the last inequality uses Lemma 1. Combining Eq. (15) and Eqs. (22)–(25), one obtains

Etrial(k,ℓ)−E(Rtrial(k,ℓ),Ainter(k,ℓ))≤Mˉ∥Ainter(k,ℓ)−A(k)∥F2.(26)E_{\mathrm{trial}}^{(k,\ell)}-E(R_{\mathrm{trial}}^{(k,\ell)},A_{\mathrm{inter}}^{(k,\ell)})\le\bar{M}\|A_{\mathrm{inter}}^{(k,\ell)}-A^{(k)}\|_{\mathrm{F}}^2. \tag*{(26)}

Finally, putting Eqs. (20) and (26) together, we have

Etrial(k,ℓ)−E(k)≤[L∇E2−1αatom(k−1,ℓ)]∥Rtrial(k,ℓ)−R(k)∥F2+[L∇E2+Mˉ−1αlatt(k−1,ℓ)]∥Ainter(k,ℓ)−A(k)∥F2.E_{\mathrm{trial}}^{(k,\ell)}-E^{(k)}\le \left[\frac{L_{\nabla E}}{2}-\frac{1}{\alpha_{\mathrm{atom}}^{(k-1,\ell)}}\right]\|R_{\mathrm{trial}}^{(k,\ell)}-R^{(k)}\|_{\mathrm{F}}^2+ \left[\frac{L_{\nabla E}}{2}+\bar{M}-\frac{1}{\alpha_{\mathrm{latt}}^{(k-1,\ell)}}\right]\|A_{\mathrm{inter}}^{(k,\ell)}-A^{(k)}\|_{\mathrm{F}}^2.

By virtue of the criterion (7) and E(k)≤Eˉ(k)E^{(k)}\le\bar{E}^{(k)} [23 Lemma 2], the LM process must stop when

{L∇E2−1αatom(k−1,ℓ)≤−ηαatom(k−1,ℓ),L∇E2+Mˉ−1αlatt(k−1,ℓ)≤−ηαlatt(k−1,ℓ),\left\{ \begin{aligned} \frac{L_{\nabla E}}{2}-\frac{1}{\alpha_{\mathrm{atom}}^{(k-1,\ell)}}&\le-\frac{\eta}{\alpha_{\mathrm{atom}}^{(k-1,\ell)}},\\ \frac{L_{\nabla E}}{2}+\bar{M}-\frac{1}{\alpha_{\mathrm{latt}}^{(k-1,\ell)}}&\le-\frac{\eta}{\alpha_{\mathrm{latt}}^{(k-1,\ell)}}, \end{aligned} \right.

which is fulfilled once ℓ\ell reaches ℓˉ\bar{\ell}. Since E(Rtrial(k,ℓ),Ainter(k,ℓ))≤Eˉ(k)≤E(0)E(R_{\mathrm{trial}}^{(k,\ell)},A_{\mathrm{inter}}^{(k,\ell)})\le\bar{E}^{(k)}\le E^{(0)}, (R(k+1),A(k+1))∈L(R^{(k+1)},A^{(k+1)})\in\mathcal{L}.

Leveraging Lemma 2, we are ready to establish the convergence of PANBB.

Theorem 1. Suppose that Assumptions 1 and 2 hold. Let {(R(k),A(k))}\{(R^{(k)},A^{(k)})\} be the configuration sequence generated by PANBB with μk∈[μmin⁡,μmax⁡]⊂(0,1]\mu_k\in[\mu_{\min},\mu_{\max}]\subset(0,1]. Then we have Fatom(k)→0F_{\mathrm{atom}}^{(k)}\to0 and F~latt(k)→0\widetilde{F}_{\mathrm{latt}}^{(k)}\to0 as k→∞k\to\infty. Moreover, there exists at least one limit point of {(R(k),A(k))}\{(R^{(k)},A^{(k)})\} and any limit point (R⋆,A⋆)(R^\star,A^\star) satisfies Fatom⋆=0F_{\mathrm{atom}}^\star=0, Σdev⋆=0\Sigma_{\mathrm{dev}}^\star=0.

Proof. By Lemma 2, the backtracking scheme (9), and αatom(k−1,0)≥αatom(k)\alpha_{\mathrm{atom}}^{(k-1,0)}\ge\alpha_{\mathrm{atom}}^{(k)}, αlatt(k−1,0)≥αlatt(k)\alpha_{\mathrm{latt}}^{(k-1,0)}\ge\alpha_{\mathrm{latt}}^{(k)}, it holds that

αatom(k)≥βatom:=αatom0δatomℓˉ,αlatt(k)≥βlatt:=αlatt0δlattℓˉ.\alpha_{\mathrm{atom}}^{(k)}\ge\beta_{\mathrm{atom}}:=\alpha_{\mathrm{atom}}^{0}\delta_{\mathrm{atom}}^{\bar{\ell}},\qquad \alpha_{\mathrm{latt}}^{(k)}\ge\beta_{\mathrm{latt}}:=\alpha_{\mathrm{latt}}^{0}\delta_{\mathrm{latt}}^{\bar{\ell}}.

Plugging this into the criterion (7), we obtain

E(k+1)≤Eˉ(k)−η[βatom∥Fatom(k)∥F2+βlatt∥F~latt(k)∥F2].(27)E^{(k+1)}\le\bar{E}^{(k)}-\eta\left[\beta_{\mathrm{atom}}\|F_{\mathrm{atom}}^{(k)}\|_{\mathrm{F}}^2+\beta_{\mathrm{latt}}\|\widetilde{F}_{\mathrm{latt}}^{(k)}\|_{\mathrm{F}}^2\right]. \tag*{(27)}

Let Δ(k):=βatom∥Fatom(k)∥F2+βlatt∥F~latt(k)∥F2\Delta^{(k)} := \beta_{\mathrm{atom}}\|F_{\mathrm{atom}}^{(k)}\|_{\mathrm{F}}^2+\beta_{\mathrm{latt}}\|\widetilde{F}_{\mathrm{latt}}^{(k)}\|_{\mathrm{F}}^2. From this, a recursion for the monitoring sequence follows.

E‾(k+1)=(8)E‾(k)+μ(k)q(k)E(k+1)q(k+1)≤(A15)E‾(k)+μ(k)q(k)(E‾(k)−ηΔ(k))q(k+1)=E‾(k)−ημ(k)q(k)q(k+1)Δ(k).(28)\overline{E}^{(k+1)} \overset{(8)}{=} \frac{\overline{E}^{(k)}+\mu^{(k)}q^{(k)}E^{(k+1)}}{q^{(k+1)}} \leq^{(A15)} \frac{\overline{E}^{(k)}+\mu^{(k)}q^{(k)}\left(\overline{E}^{(k)}-\eta\Delta^{(k)}\right)}{q^{(k+1)}} =\overline{E}^{(k)}-\eta\frac{\mu^{(k)}q^{(k)}}{q^{(k+1)}}\Delta^{(k)}. \tag*{(28)}

A byproduct of Eq. (28) is the monotonicity of {E‾(k)}\{\overline{E}^{(k)}\}. Since EE is bounded from below over L\mathcal{L} (by Assumption 1 and Lemma 1) and E(k)≤E‾(k)E^{(k)}\leq\overline{E}^{(k)} for any k≥0k\geq0 [23 Lemma 2], {E‾(k)}\{\overline{E}^{(k)}\} is also bounded from below. Consequently,

∑k=0∞μ(k)q(k)Δ(k)q(k+1)≤(A16)1η∑k=0∞(E‾(k)−E‾(k+1))<+∞.(29)\sum_{k=0}^{\infty}\frac{\mu^{(k)}q^{(k)}\Delta^{(k)}}{q^{(k+1)}}\overset{(A16)}{\leq}\frac{1}{\eta}\sum_{k=0}^{\infty}\left(\overline{E}^{(k)}-\overline{E}^{(k+1)}\right)<+\infty. \tag*{(29)}

Following from Eq. (8) and μmax⁡≥μ(k)≥μmin⁡>0\mu_{\max}\geq\mu^{(k)}\geq\mu_{\min}>0, q(k)≥1q^{(k)}\geq1, we can derive

μ(k)q(k)q(k+1)≥μmin⁡q(k)μ(k)q(k)+1=μmin⁡μ(k)+1/q(k)≥μmin⁡μmax⁡+1,\frac{\mu^{(k)}q^{(k)}}{q^{(k+1)}}\geq\frac{\mu_{\min}q^{(k)}}{\mu^{(k)}q^{(k)}+1}=\frac{\mu_{\min}}{\mu^{(k)}+1/q^{(k)}}\geq\frac{\mu_{\min}}{\mu_{\max}+1},

which, together with Eq. (29), yields Δ(k)→0\Delta^{(k)}\to0 as k→∞k\to\infty. By βatom,βlatt>0\beta_{\mathrm{atom}},\beta_{\mathrm{latt}}>0, the definition of Δ(k)\Delta^{(k)}, and Lemma 1, one has Fatom(k)→0F_{\mathrm{atom}}^{(k)}\to0 and F~latt(k)→0\widetilde{F}_{\mathrm{latt}}^{(k)}\to0 as k→∞k\to\infty.

Let (R⋆,A⋆)(R^\star,A^\star) be a limit point of {(R(k),A(k))}\{(R^{(k)},A^{(k)})\}, whose existence is guaranteed by Lemma 1. Then by the above arguments and Eq. (3), Fatom⋆=0F_{\mathrm{atom}}^\star=0 and

0=F~latt⋆=Flatt⋆−⟨B⋆,Flatt⋆⟩∥B⋆∥F2B⋆=V[Σ⋆B⋆−⟨B⋆,Σ⋆B⋆⟩∥B⋆∥F2B⋆],0=\widetilde{F}_{\mathrm{latt}}^\star=F_{\mathrm{latt}}^\star-\frac{\langle B^\star,F_{\mathrm{latt}}^\star\rangle}{\|B^\star\|_{\mathrm{F}}^2}B^\star =V\left[\Sigma^\star B^\star-\frac{\langle B^\star,\Sigma^\star B^\star\rangle}{\|B^\star\|_{\mathrm{F}}^2}B^\star\right],

which implies that Σ⋆\Sigma^\star is a multiple of II and hence Σdev⋆=0\Sigma_{\mathrm{dev}}^\star=0. ■

Magnetic Disorder Modeling

We elaborate below on the technical details when modeling the magnetic disorder for the PM structures, in particular, how to solve the distribution problem for the spin-up or spin-down sites. We assume that an atomic configuration is available (e.g., given by the SAE method), and moreover introduce some notations:

  • S∈NS\in\mathbb{N}: number of the atomic species (in the AlCoCrFeNi case, equals 5);

  • Smag∈NS_{\mathrm{mag}}\in\mathbb{N}: number of the magnetic atomic species (in the AlCoCrFeNi case, equals 4). For convenience, we assume that species 1,…,Smag1,\ldots,S_{\mathrm{mag}} are magnetic;

  • Ns∈NN_s\in\mathbb{N}: number of the atoms in the unit cell belonging to species ss (in the AlCoCrFeNi case, N1=⋯=N5=32)(s∈{1,…,S})N_1=\cdots=N_5=32) (s\in\{1,\ldots,S\});

  • Nmag∈NN_{\mathrm{mag}}\in\mathbb{N}: number of magnetic atoms in the unit cell (in the AlCoCrFeNi case, equals 64);

  • ℓax∈N\ell_{\mathrm{ax}}\in\mathbb{N}: number of the atomic layers in the unit cell with axis ax∈{a,b,c}\mathrm{ax}\in\{a,b,c\} as normal;

  • Is⊆{1,…,N}I_s\subseteq\{1,\ldots,N\}: indices of the sites occupied by species ss (s∈{1,…,Smag}s\in\{1,\ldots,S_{\mathrm{mag}}\});

  • Iax,l,s⊆{1,…,N}I_{\mathrm{ax},l,s}\subseteq\{1,\ldots,N\}: indices of the sites occupied by species ss located on the llth atomic layer with axis ax\mathrm{ax} as normal (ax∈{a,b,c}\mathrm{ax}\in\{a,b,c\}, l∈{1,…,ℓax}l\in\{1,\ldots,\ell_{\mathrm{ax}}\}, s∈{1,…,Smag}s\in\{1,\ldots,S_{\mathrm{mag}}\});

  • y∈{0,1}Nmagy\in\{0,1\}^{N_{\mathrm{mag}}}: binary indicators for the spin-up configuration, where yn=1y_n=1 (n∈{1,…,Nmag}n\in\{1,\ldots,N_{\mathrm{mag}}\}) means that the nnth atom spins up. Conversely, one may interpret the zero entries as spin-down atoms;

  • y0∈{0,1}Nmagy^0\in\{0,1\}^{N_{\mathrm{mag}}}: binary indicators for the trial spin-up configuration. The meaning of each entry is analogous to that of yy.

With the above notations, we can formulate the spin-up distribution problem as the following binary integer programming:

max⁡y⟨y0,y⟩s.t.⌊Ns2ℓax⌋≤∑n∈Iax,l,syn≤⌈Ns2ℓax⌉,l∈{1,…,ℓax}, ax∈{a,b,c}, s∈{1,…,Smag},∑n∈Isyn=Ns/2,s∈{1,…,Smag},y∈{0,1}Nmag.(30)\begin{aligned} \max_y\quad&\langle y^0,y\rangle\\ \mathrm{s.t.}\quad&\left\lfloor\frac{N_s}{2\ell_{\mathrm{ax}}}\right\rfloor\leq\sum_{n\in I_{\mathrm{ax},l,s}}y_n\leq\left\lceil\frac{N_s}{2\ell_{\mathrm{ax}}}\right\rceil,\\ &l\in\{1,\ldots,\ell_{\mathrm{ax}}\},\ \mathrm{ax}\in\{a,b,c\},\ s\in\{1,\ldots,S_{\mathrm{mag}}\},\\ &\sum_{n\in I_s}y_n=N_s/2,\quad s\in\{1,\ldots,S_{\mathrm{mag}}\},\\ &y\in\{0,1\}^{N_{\mathrm{mag}}}. \tag*{(30)} \end{aligned}

The integer programming (B1) maximizes the similarity between the trial and optimized spin-up structures, represented by the inner product of their corresponding binary indicators. In our experiments, we simply take y0=0y^0=0. The first line of constraints enforces that the spin-up sites of each species are evenly distributed over each atomic layer. The second line of constraints imposes that the numbers of spin-up and spin-down sites for each species are identical. Upon solving the integer programming (B1) (e.g., by off-the-shelf software like MATLAB [52]), one can transform the solution y⋆y^\star to a concrete magnetic configuration.

References

  1. [18]H.-F. Song and H.-F. Liu, Phys. Rev. B 75, 245126 (2007).
  2. [21]J. Shewchuk, An Introduction to the Conjugate Gradient Method Without the Agonizing Pain, Tech. Rep. (Carnegie Mellon University, Pittsburgh, PA, USA, 1994).
  3. [22]P. Pulay, Chem. Phys. Lett. 73, 393 (1980).
  4. [23]Y. Hu, X. Gao, Y. Zhao, X. Liu, and H. Song, Phys. Rev. B 106, 104101 (2022).
  5. [24]H. Song, F. Tian, Q.-M. Hu, L. Vitos, Y. Wang, J. Shen, and N. Chen, Phys. Rev. Mater. 1, 023404 (2017).
  6. [25]J. Wu, Z. Yang, J. Xian, X. Gao, D. Lin, and H. Song, Front. Mater. 7, 590143 (2020).
  7. [26]B. Cheng, F. Zhang, H. Lou, X. Chen, P. K. Liaw, J. Yan, Z. Zeng, Y. Ding, and Q. Zeng, Scripta Mater. 161, 88 (2019).
  8. [27]V. Atalla, Density-Functional Theory and Beyond for Organic Electronic Materials, Ph.D. thesis, Technischen Universität Berlin, Berlin, Germany (2013).DOI
  9. [28]F. Knuth, Strain and Stress: Derivation, Implementation, and Application to Organic Crystals, Ph.D. thesis, Freie Universität Berlin, Berlin, Germany (2015).DOI
  10. [29]J. Nocedal and S. J. Wright, Numerical Optimization, 2nd ed., Springer Series in Operations Research and Financial Engineering (Springer New York, NY, USA, 2006).
  11. [30]J. Barzilai and J. Borwein, IMA J. Numer. Anal. 8, 141 (1988).
  12. [31]Y. Dai and R. Fletcher, Numer. Math. 100, 21 (2005).
  13. [32]R. Fletcher, in Optimization and Control with Applications, Applied Optimization, Vol. 96, edited by L. Qi, K. Teo, and X. Yang (Springer, Boston, MA, USA, 2005) pp. 235–256.
  14. [33]The specific initial step size for the atomic part is also used by CG realized in VASP and cessp. The numerical results of CG and PANBB will become more comparable. A small initial step size is chosen for the lattice part to prevent unexpected breakdown. In simulations, the performance of PANBB are not sensitive to the two values.
  15. [34]We use a small δatom to end the LM process more quickly and help the ensuing ABB step sizes learn local curvatures better. The value of δlatt is not so critical, since usually α(k−1,0)latt is already sufficiently small.
  16. [35]J. Fang, X. Gao, H. Song, and H. Wang, J. Chem. Phys. 144, 244103 (2016).
  17. [36]X. Gao, Z. Mo, J. Fang, H. Song, and H. Wang, Comput. Phys. Commun. 211, 54 (2017).
  18. [37]Y. Zhou, H. Wang, Y. Liu, X. Gao, and H. Song, Phys. Rev. E 97, 033305 (2018).
  19. [38]J. Fang, X. Gao, and H. Song, Commun. Comput. Phys. 26, 156 (2019).
  20. [39]J. P. Perdew, K. Burke, and M. Ernzerhof, Phys. Rev. Lett. 77, 3865 (1996).
  21. [40]F. Jollet, M. Torrent, and N. Holzwarth, Comput. Phys. Commun. 185, 1246 (2014).
  22. [41]H. J. Monkhorst and J. D. Pack, Phys. Rev. B 13, 5188 (1976).
  23. [42]A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, G. Gunter, D. Skinner, G. Ceder, and K. A. Persson, APL Mater. 1, 011002 (2013).
  24. [43]S. S. Borysov, R. M. Geilhufe, and A. V. Balatsky, PLoS ONE 12, e0171501 (2017).
  25. [44]E. Dolan and J. Moré, Math. Program. 91, 201 (2002).
  26. [45]H. Tian, D. Lin, X. Gao, Y. Zhao, and H. Song, J. Chem. Phys. 153, 034101 (2020).
  27. [46]M. Conforti, G. Cornuéjols, and G. Zambelli, Integer Programming, Graduate Texts in Mathematics, Vol. 271 (Springer, Cham, Switzerland, 2014).
  28. [47]B. L. Gyorffy, A. J. Pindor, J. Staunton, G. M. Stocks, and H. Winter, J. Phys. F Met. Phys. 15, 1337 (1985).
  29. [48]F. Birch, Phys. Rev. 71, 809 (1947).
  30. [49]A. Otero-de-la-Roza and V. Luaña, Comput. Phys. Commun. 182, 1708 (2011).
  31. [50]H. Ge, H. Song, J. Shen, and F. Tian, Mater. Chem. Phys 210, 320 (2018).
  32. [51]Y.-F. Kao, S.-K. Chen, T.-J. Chen, P.-C. Chu, J.-W. Yeh, and S.-J. Lin, J. Alloys Compd. 509, 1607 (2011).
  33. [52]The MathWorks Inc., MATLAB version: 9.13.0 (R2022b) (2022).

Paper details

vlnc:W_n2bk33sd
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