betaFeedback

WorksPhysics and Astronomy

Tensor Learning and Compression of N-phonon Interactions

Abstract

Phonon interactions from lattice anharmonicity govern thermal properties and heat transport in materials. These interactions are described by n-th order interatomic force constants (n-IFCs), which can be viewed as high-dimensional tensors correlating the motion of n atoms, or equivalently encoding n-phonon scattering processes in momentum space. Here, we introduce a tensor decomposition to efficiently compress n-IFCs for arbitrary order n. Using tensor learning, we find optimal low-rank approximations of n-IFCs by solving the resulting optimization problem. Our approach reveals the inherent low dimensionality of phonon-phonon interactions and allows compression of the 3 and 4-IFC tensors by factors of up to 103−10410^3-10^4 while retaining high accuracy in calculations of phonon scattering rates and thermal conductivity. Calculations of thermal conductivity using the compressed n-IFCs achieve a speed-up by nearly three orders of magnitude with >98% accuracy relative to the reference uncompressed solution. These calculations include both 3- and 4-phonon scattering and are shown for a diverse range of materials (Si, HgTe, MgO, TiNiSn and monoclinic ZrO2_2). In addition to accelerating state-of-the-art thermal transport calculations, the method shown here paves the way for modeling strongly anharmonic materials and higher-order phonon interactions.

Phonon-phonon (ph-ph) interactions are crucial to understanding thermal transport, lattice dynamics, and structural phase transitions in condensed matter [1]. They originate from the anharmonicity of the lattice potential and can be described by nn-th order interatomic force constant (nn-IFC) tensors, where n≥3n \ge3. Density function theory (DFT) [2] calculations combined with fitting algorithms provide accurate nn-IFC tensors [3, 4, 5, 6], enabling quantitative predictions of thermal transport in materials [7, 8, 9, 10, 11, 12]. However, the high-dimensionality of the IFC tensors obscures the underlying physics and poses significant computational challenges. The complexity of nn-phonon (nn-ph) interactions grows exponentially with order nn, a clear example of the curse of dimensionality. For thermal conductivity calculations, 4-ph interactions require orders of magnitude more computational effort than 3-ph interactions, whereas 5-ph and higher-order ph-ph interactions remain inaccessible.

To overcome this complexity, discovering low-rank approximations of nn-IFC could be game changing. There is growing interest in such dimensionality reduction approaches, including tensor network states for many-body wave functions [13] [14], tensor train for differential equations [15–17] and Feynman diagrams [16, 18], tensor hyper-contraction [19–21] and density fitting [22] for electron interactions in quantum chemistry, and feature optimization for atomic machine learning [23–26]. Recent work has employed singular value decomposition to compress electron-phonon (ee-ph) interactions and greatly speed up first-principles ee-ph calculations [27]. For ph-ph interactions, previous work has taken advantage of crystal symmetry to reduce the number of free parameters and compute the nn-IFCs with fewer DFT force calculations [28, 29]. The compressed sensing technique has also been used to obtain nn-IFCs using sparse solvers [4] [30].

However, in previous work the nn-IFCs are still handled explicitly in full tensor form, Φijk\Phi_{ijk} and Φijkl\Phi_{ijkl} for 3- and 4-IFCs respectively, and a low-rank representation of nn-IFCs is still missing.

In this Letter, we introduce a low-rank tensor ansatz for nn-ph interactions in momentum space based on the CANDECOMP/PARAFAC (CP) decomposition [31], a compression method used in tensor learning [32–34]. The proposed ansatz is a permanent CP (PCP) decomposition, which generalizes the CP decomposition to enforce bosonic statistics. To find an optimal low-rank PCP decomposition, we formulate the tensor learning problem and solve it on GPU hardware. The optimized low-rank PCP tensors achieve large compression factors of 10310^3–10410^4 with minimal compression losses of only a few percent in all materials we study. This result reveals the inherent low-dimensionality of nn-ph interactions and enables a speedup of nearly three orders of magnitude for calculations using nn-IFCs, including phonon relaxation times and thermal conductivity. We also introduce constraints to treat ph-ph interactions for long-wavelength acoustic phonons, leading to nearly lossless predictions of thermal conductivity compared to calculations using full nn-IFC tensors. Finally, we show that the PCP ansatz can uncover dominant modes in nn-ph interactions, providing a valuable tool for understanding microscopic thermal transport mechanisms and formulating accurate minimal models. Beyond ph-ph interactions, the PCP ansatz and corresponding open-source routines developed here provide a blueprint for modeling momentum-dependent tensors, with broad applications in condensed matter physics.

We consider the nn-ph interaction V(n)(Q1,…,Qn)V^{(n)}(\mathbf{Q}_1,\ldots,\mathbf{Q}_n), which describes the scattering amplitude of phonon modes. Each mode is specified by a wave vector qi\mathbf{q}_i and branch index νi\nu_i, collectively denoted as Qi=(νi,qi)\mathbf{Q}_i=(\nu_i,\mathbf{q}_i) for i=1,…,ni=1,\ldots,n. For instance, the 3-ph interaction is given by

V(3)(Q1,Q2,Q3)=δ(∑i=13qi)∑b1,l2b2,l3b3∑α1α2α3Φ0b,l1b1,l2b2,l3b3α1,α2,α3eα1b1Q1eα2b2Q2eα3b3Q3mb1mb2mb3eiq2rl2+iq3rl3,(1)V^{(3)}(\mathbf{Q}_1,\mathbf{Q}_2,\mathbf{Q}_3)=\delta\left(\sum_{i=1}^{3}\mathbf{q}_i\right)\sum_{b_1,l_2b_2,l_3b_3}\sum_{\alpha_1\alpha_2\alpha_3}\frac{\Phi^{\alpha_1,\alpha_2,\alpha_3}_{0b,l_1b_1,l_2b_2,l_3b_3}e^{\mathbf{Q}_1}_{\alpha_1b_1}e^{\mathbf{Q}_2}_{\alpha_2b_2}e^{\mathbf{Q}_3}_{\alpha_3b_3}}{\sqrt{m_{b_1}m_{b_2}m_{b_3}}}e^{i\mathbf{q}_2\mathbf{r}_{l_2}+i\mathbf{q}_3\mathbf{r}_{l_3}}, \tag*{(1)}

where each of the ll, bb, and α\alpha label the primitive cell, atom, and Cartesian coordinate, respectively; eαbQe^{\mathbf{Q}}_{\alpha b} is the displacement eigenvector of phonon mode Q\mathbf{Q}, mbm_b is the mass of atom bb, and rl\mathbf{r}_l is the position of primitive cell ll. Above, Φ0b,l1b1,l2b2α1,α2,α3\Phi^{\alpha_1,\alpha_2,\alpha_3}_{0b,l_1b_1,l_2b_2} is the 3-IFC tensor associated with three atomic displacements, the first displacement fixed at the cell origin. For general nn-ph interactions, analogous nn-IFC tensors Φ(n)\Phi^{(n)} can be defined.

To compress the nn-ph interactions V(n)V^{(n)}, we propose the PCP decomposition as the low-rank ansatz

V~(n)(Q1,…,Qn)=δ(∑i=1nqi)∑ξ=1Nc(n)λξn!∑σ∈SnA1ξ(Qσ1)A2ξ(Qσ2)…Anξ(Qσn),(2)\widetilde{V}^{(n)}(\mathbf{Q}_1,\ldots,\mathbf{Q}_n)=\delta\left(\sum_{i=1}^{n}\mathbf{q}_i\right)\sum_{\xi=1}^{N_c^{(n)}}\frac{\lambda_\xi}{n!}\sum_{\sigma\in S_n}A_1^\xi(\mathbf{Q}_{\sigma_1})A_2^\xi(\mathbf{Q}_{\sigma_2})\ldots A_n^\xi(\mathbf{Q}_{\sigma_n}), \tag*{(2)}

where Nc(n)N_c^{(n)} is the PCP rank, λξ\lambda_\xi the ξ\xi-th PCP singular value, AiξA_i^\xi are the corresponding PCP modes, and SnS_n is the symmetric group of order nn containing all the permutations of (1,…,n)(1,\ldots,n). Note that Eq. (2) correctly preserves the bosonic exchange symmetry and momentum conservation of nn-ph interactions.

The nn-ph interactions in compressed form involve nn momenta (Q1,Q2,…,Qn)(\mathbf{Q}_1,\mathbf{Q}_2,\ldots,\mathbf{Q}_n), each associated with Nc(n)N_c^{(n)} PCP modes Aξ(Q)A_\xi(\mathbf{Q}), which are functions defined on a grid of size NQN_Q. This gives a storage requirement of O(nNc(n)NQ)O(nN_c^{(n)}N_Q) for the nn-ph interactions in compressed form. For example, storing 3-ph interactions V(3)V^{(3)} conventionally requires O(NQ2)O(N_Q^2) resources versus O(3Nc(3)NQ)O(3N_c^{(3)}N_Q) for its PCP compressed counterpart V~(3)\widetilde{V}^{(3)}. This yields a large storage reduction, by a factor of NQ/(3Nc(3))≈103N_Q/(3N_c^{(3)})\approx10^3–10410^4 in a typical calculation.

Since V(n)V^{(n)} is obtained by transforming the nn-IFC Φ(n)\Phi^{(n)} to momentum space, we derive the corresponding low-rank ansatz for 3- and 4-IFCs in real space:

Φ~l1b1,l2b2,l3b3α1,α2,α3=∑ξ=1Nc(3)λξ3!∑σ∈S3∑l∏i=13Aσiξ(li−l,biαi),(3)\widetilde{\Phi}^{\alpha_1,\alpha_2,\alpha_3}_{l_1b_1,l_2b_2,l_3b_3}=\sum_{\xi=1}^{N_c^{(3)}}\frac{\lambda_\xi}{3!}\sum_{\sigma\in S_3}\sum_l\prod_{i=1}^{3}A_{\sigma_i}^\xi(l_i-l,b_i\alpha_i), \tag*{(3)}
Φ~l1b1,l2b2,l3b3,l4b4α1,α2,α3,α4=∑ξ=1Nc(4)λξ4!∑σ∈S4∑l∏i=14Bσiξ(li−l,biαi),\widetilde{\Phi}^{\alpha_1,\alpha_2,\alpha_3,\alpha_4}_{l_1b_1,l_2b_2,l_3b_3,l_4b_4}=\sum_{\xi=1}^{N_c^{(4)}}\frac{\lambda_\xi}{4!}\sum_{\sigma\in S_4}\sum_l\prod_{i=1}^{4}B_{\sigma_i}^\xi(l_i-l,b_i\alpha_i),

where Aiξ(l,bα)A_i^\xi(l,b\alpha) and Biξ(l,bα)B_i^\xi(l,b\alpha) are real-space representations of Aiξ(Q)A_i^\xi(\mathbf{Q}) in Eq. (2) for 3-ph and 4-ph interactions respectively.

To enforce the acoustic sum rule (ASR) [29] on the compressed nn-ph interactions V~(n)\widetilde{V}^{(n)}, we impose ∑l,bAiξ(l,bα)=0\sum_{l,b}A_i^\xi(l,b\alpha)=0, which preserves the ASR in the PCP compression. Derivations of nn-ph interactions and their PCP decomposition are provided in the Supplemental Material (SM) [35]. The computational cost of PCP-compressed Φ~(n)\widetilde{\Phi}^{(n)} and V~(n)\widetilde{V}^{(n)} scales linearly with system size, making it promising for calculations in complex materials with large unit cells.

To find optimal PCP modes AiξA_i^\xi (or BiξB_i^\xi) that best approximate the original IFC tensors Φ(n)\Phi^{(n)} computed with DFT, we minimize the loss function:

L=∥Φ~(n)[A]−Φ(n)∥2,(4)L=\left\|\widetilde{\Phi}^{(n)}[A]-\Phi^{(n)}\right\|^2, \tag*{(4)}

where Φ~(n)\widetilde{\Phi}^{(n)} is the low-rank IFC tensor in Eq. (3), and the norm is defined as ∥x∥=∑ixi2\lVert x\rVert=\sqrt{\sum_i x_i^2} [36]. This is a typical tensor learning problem that lacks closed-form solutions for high-order tensors. In addition, existing optimization algorithms are computationally demanding for large tensors. We address this challenge by developing a computational toolkit, called PhonoN-PCP, implemented using PyTorch [37]. This optimization routine follows a standard “neural network”-like training loop, where Φ~(n)\widetilde{\Phi}^{(n)} is evaluated in a forward pass, and AiξA_i^\xi is the learnable weight trained using automatic differentiation. A typical optimization involves 10310^3–10410^4 trainable parameters, depending on the material. The algorithm runs efficiently on GPUs – the compression process converges in a few minutes (Si and other simple systems) to one hour (ZrO2_2) on a single NVIDIA A100 chip.

We apply our approach to five materials: Si, HgTe, MgO, TiNiSn, and monoclinic ZrO2_2. These materials span a wide range of properties, including nonpolar and polar (or ionic) bonds, high and low crystal symmetry, and different levels of anharmonicity. For each material, we generate force-displacement datasets from DFT force calculations on supercells and then employ Alamode [3] to extract the 3- and 4-IFCs [38]. The symmetry of the nn-IFCs is taken into account in the force constant fitting process [3]. We calculate DFT forces using VASP [39, 40] with PBEsol functional [41]. For polar materials, we calculate Born effective charges using density functional perturbation theory in VASP, and use the Ewald summation to compute the nonanalytic part of the dynamical matrix [42]. Additional computational details, are provided in the SM [35].

To assess the accuracy of PCP for compressing nn-ph interactions, we calculate the error relative to uncompressed nn-IFCs, expressed as the compression loss ∥Φ~(n)−Φ(n)∥/∥Φ(n)∥\lVert\tilde{\Phi}^{(n)}-\Phi^{(n)}\rVert/\lVert\Phi^{(n)}\rVert, as a function of the compression factor γ(n)\gamma^{(n)}, which quantifies parameter reduction in the nn-IFCs:

γ(n)=# of nonzero entries of Φ(n)Nc(n).(5)\gamma^{(n)}=\frac{\#\text{ of nonzero entries of }\Phi^{(n)}}{N_c^{(n)}}. \tag*{(5)}

This error analysis is shown for the 3-ph interactions in Figure 1(a). We achieve compression factors of 10310^3–10410^4 at the 3% compression error threshold. For comparison, a truncated singular value decomposition (tSVD) of 3-ph interactions, inspired by our previous work on ee-ph coupling [27], can only reach compression factors of a few hundred for the same 3% compression loss. Therefore, the PCP approach provides a 10-times more favorable dimensionality reduction than SVD for all materials we study. We attribute the superior performance of PCP to its flexible ansatz combined with the correct permutation symmetry and the explicit parameter optimization during tensor learning, which enable effective low-rank representation of nn-ph interactions. We achieve similarly large compression factors for 4-ph interactions, as shown in Figure 1(b). These results highlight the inherent low-dimensionality of nn-ph interactions and show their accurate low-rank approximation using tensor learning.

Relative compression loss vs. compression factor

Figure 1. Relative compression loss vs. compression factor γ(n)\gamma^{(n)}. (a) Results for 3-IFC tensors comparing PCP (solid lines) and tSVD (dashed lines). (b) Results for 4-IFC tensors. The compression factor γ(n)\gamma^{(n)} is defined in Eq. (5).

Note that crystal symmetry can also reduce the size of the nn-IFCs, but it cannot be used to speed up calculations of phonon relaxation times and thermal conductivity, which require the full nn-IFCs. In addition, the size reduction from symmetry is significantly smaller than the PCP compression factor shown here [35].

Using compressed nn-ph interactions, we compute the thermal conductivity at 300 K for the four materials. Our calculations use the single-mode relaxation time approximation [7] [43, 44] and include both 3- and 4-ph interactions [45–48] with a converged number of scattering processes for 3- and 4-ph scattering rates [49] (see details in SM [35]). The thermal conductivity κ~\tilde{\kappa}, computed with compressed nn-IFC tensors, is compared with the reference value κ\kappa from full (uncompressed) tensors in Figure 2(a) and (Figure 2b). We find that κ~\tilde{\kappa} converges to the reference value for large compression factors, γ(3)≈7×103\gamma^{(3)} \approx7 \times10^{3} and γ(4)≈104\gamma^{(4)} \approx10^{4} for Si, MgO, TiNiSn, and ZrO2_2, and slightly smaller γ(4)=3×103\gamma^{(4)} = 3 \times10^{3} for HgTe, where the 4-ph contribution is more important [9]. Even for these very large compression factors, the approximate thermal conductivity is still within 2% of the reference value obtained with full IFCs [35]. This high accuracy for thermal conductivity is a result of the low compression losses for nn-IFCs.

Ratio of thermal conductivities for compressed 3- and 4-phonon interactions

Figure 2. Ratio of the thermal conductivity κ~\tilde{\kappa}, computed using compressed 3- and 4-ph interactions, to the thermal conductivity κ\kappa computed with uncompressed tensors at 300 K. Results are shown as a function of: (a) 3-ph compression factor, γ(3)\gamma^{(3)}, for fixed γ(4)>103\gamma^{(4)}>10^3 (Nc(4)=48N_c^{(4)}=48), and (b) 4-ph compression factor, γ(4)\gamma^{(4)}, for fixed γ(3)>2×103\gamma^{(3)}>2\times10^3 (Nc(3)=24N_c^{(3)}=24 for all materials except ZrO2_2, and Nc(3)=144N_c^{(3)}=144 for ZrO2_2). The light (dark) shade regions show the 95% (98%) accuracy windows.

We also examine the accuracy of compressed IFC tensors for calculating the microscopic scattering rate for each phonon mode, τνq−1\tau_{\nu\mathbf{q}}^{-1}. In Figure 3(a) and (Figure 3b), the phonon scattering rates from compressed IFC tensors are nearly identical to those from full IFC tensors for each single phonon mode, even in cases where τνq\tau_{\nu\mathbf{q}} varies by up to four orders of magnitude over the phonon spectrum. We calculate the coefficient of determination R2R^2 [50] between approximate and reference τνq−1\tau_{\nu\mathbf{q}}^{-1}, and find values of R2=0.9996R^2 = 0.9996 for Si and R2=0.9986R^2 = 0.9986 for HgTe, further confirming the accuracy of the compressed nn-ph interactions. The approximate scattering rates achieve a similar accuracy in TiNiSn and MgO [35].

Phonon scattering rates for Si and HgTe

Figure 3. Phonon scattering rates combining 3- and 4-ph interactions, shown as a function of phonon energy for (a) Si and (b) HgTe, comparing results for full IFC tensors (blue dots) and compressed IFC tensors (orange circles). In both materials, we show results for compression factors γ(3)=2200\gamma^{(3)}=2200 for 3-ph, and γ(4)=2500\gamma^{(4)}=2500 for 4-ph interactions.

The dimensionality reduction provided by PCP leads to massive cost savings for calculations involving nn-ph interactions, with speed-up proportional to the compression factor. In Figure 4, we compare CPU wall times for calculations of 3-ph and 4-ph scattering rates using full and compressed nn-IFCs with large compression factors (γ(3)>2200\gamma^{(3)} > 2200 and γ(4)>1000\gamma^{(4)} > 1000). For all materials studied here, the use of compressed IFC tensors enables a speed-up of 260–7200 in calculations of thermal conductivity. Calculations of τνq−1\tau_{\nu\mathbf{q}}^{-1} using compressed IFC tensors are so efficient that computing the thermal conductivity with 3- and 4-ph interactions takes less than a minute for each material studied here, including monolithic ZrO2_2 with a 12-atom unit cell. Using compressed IFC tensors, we are able to accurately extrapolate the thermal conductivity to the thermodynamic limit (TDL) by employing progressively denser momentum grids. The importance of extrapolating to the TDL is clear from the case of HgTe, where the TDL-extrapolated κ\kappa is 47% larger than the value obtained with reasonable grid sizes of 16316^3 [35]. The calculation for the largest grid, 2503250^3, requires only 2.5 CPU node-hours. We attribute the slow convergence of the thermal conductivity in HgTe to the important role of 4-ph processes (see analysis in SM [35]).

Comparison of computational cost for 3-ph and 4-ph scattering rates

Figure 4. Comparison of computational cost (CPU wall time) for calculations of phonon scattering rates using compressed and full IFC tensors. Results are shown for (a) 3-ph scattering rates with compression factors γ(3)>2200\gamma^{(3)} > 2200 and (b) 4-ph scattering rates with compression factors γ(4)>1000\gamma^{(4)} > 1000 in all materials. The speedup achieved with compressed tensors is given in figure for each material.

In Figure 4(a), the speed-up increases for increasing unit cell sizes. For large unit cells, such as in ZrO2_2, the speed-up becomes comparable to the compression factor because the cost of computing the nn-ph interactions dominates the total CPU wall time [35]. The significant speed-up achieved by PCP can be understood by comparing the nn-ph interactions and their compressed counterparts. The size of the nn-IFC tensor in real space, Φ(n)\Phi^{(n)}, scales as NRn−1(3Na)nN_R^{n-1}(3N_a)^n, where NRN_R is the number of Wigner-Seitz cells, which is the same as the size of the coarse q\mathbf{q}-point grid, and NaN_a is the number of atoms in the unit cell, so that 3Na3N_a is the number of phonon modes. The summation over these variables in the conventional nn-ph interactions, exemplified by the 3-ph case in Eq. (1), is replaced by a summation over Nc(n)N_c^{(n)} PCP modes for the compressed nn-ph interactions in Eq. (2). This reduces the computational cost of V(n)(Q1,…,Qn)V^{(n)}(\mathbf{Q}_1,\ldots,\mathbf{Q}_n) by a factor of O(3NanNRn−1/Nc(n))\mathcal{O}\left(3N_a^nN_R^{n-1}/N_c^{(n)}\right) (see analysis in SM [35]).

The PCP decomposition also offers interesting ways to analyze nn-ph interactions. We express the nn-ph anharmonic energy E(n)(u)E^{(n)}(\mathbf{u}) in terms of PCP modes [35],

E(n)(u)=∑l∑ξλξn!∏i=1nϕiξ(l,u),(6)E^{(n)}(\mathbf{u}) = \sum_l \sum_\xi\frac{\lambda_\xi}{n!} \prod_{i=1}^{n} \phi_i^\xi(l,\mathbf{u}), \tag*{(6)}
ϕiξ(l,u)=⟨Aiξ(l),u⟩=∑l′bαAσiξ(l′−l,bα)u(l′bα),(7)\begin{aligned} \phi_i^\xi(l,\mathbf{u}) = \langle A_i^\xi(l),\mathbf{u}\rangle\\ &= \sum_{l'b\alpha} A_{\sigma_i}^{\xi}(l'-l,b\alpha)u(l'b\alpha), \tag*{(7)} \end{aligned}

where u(lbα)u(lb\alpha) is the displacement of atom bb in primitive cell ll along the α\alpha direction. Here, the PCP modes [Aiξ(l)]l′bα=Ai,l′−l,bαξ[A_i^{\xi}(l)]_{l'b\alpha}=A_{i,l'-l,b\alpha}^{\xi} are viewed as local vibrational modes, centered in the ll-th primitive cell, which give dominant contributions the nn-th order anharmonic energy E(n)(u)E^{(n)}(u). Therefore, the projection ϕiξ(l,u)\phi_i^{\xi}(l,u) defined in Eq. (7) is a descriptor of the atomic environment that quantifies the similarity between the local atomic displacement uu and the PCP modes Aiξ(l)A_i^{\xi}(l). (Interestingly, the anharmonic energy for PCP decomposition in Eq. (6) has a structure similar to the slave mode [51, 52] and the atomic cluster expansion models [53].)

These PCP modes can be viewed as generalized eigenvectors of the nn-ph interaction tensor, and provide direct physical information about the dominant vibrational patterns. In Figure 5, we visualize the three modes associated with the largest PCP singular value for 3-ph interactions in Si. The PCP-mode triplet {A1ξ=1,A2ξ=1,A3ξ=1}\{A_1^{\xi=1}, A_2^{\xi=1}, A_3^{\xi=1}\} provides the best rank-1 approximation of 3-IFC tensors in Si and is obtained by setting the PCP rank to NC(3)=1N_{\mathrm{C}}^{(3)}=1 in the training process. These three modes are related by a C3C_3 rotation along the [111] crystal direction:

The PCP mode triplet $A_i^{\xi=1}$

Figure 5. The PCP mode triplet Aiξ=1A_i^{\xi=1}, i=1,2,3i=1,2,3 with the largest 3-ph coupling strength in Si. The three modes are related by the C3C_3 rotation symmetry along the [111] direction in Si.

C^3A1ξ=1=A2ξ=1,C^3A2ξ=1=A3ξ=1,C^3A3ξ=1=A1ξ=1.\hat{C}_3 A_1^{\xi=1}=A_2^{\xi=1},\quad\hat{C}_3 A_2^{\xi=1}=A_3^{\xi=1},\quad\hat{C}_3 A_3^{\xi=1}=A_1^{\xi=1}.

Even though we do not explicitly preserve the space-group symmetry when generating the PCP modes, the rank-1 PCP ansatz is capable of learning the C3C_3 symmetry of Si encoded in the uncompressed 3-IFC tensors during the optimization process. In general, the crystal symmetry is preserved in the compressed IFCs, with only a small symmetry loss resulting from compression (see SM [35]).

In summary, we propose a tensor decomposition of nn-ph interactions and show compression of nn-IFC tensors using GPU-accelerated tensor learning. Our PCP decomposition reveals the inherent low-dimensionality of 3- and 4-ph interactions in crystals with generic symmetry and unit cell size, enabling compression factors greater than 10310^3 for minimal compression errors of less than 3%. We achieve corresponding cost savings for calculations of phonon scattering rates and thermal conductivity. The large speed-up makes our method suitable for high-throughput screening of thermal transport in materials and offers a promising pathway to go beyond 3- and 4-ph interactions. The PCP-compressed nn-ph interactions can be combined with CPU and GPU parallelization to accelerate modeling of ph-ph interactions. The PCP decomposition can also uncover dominant atomic environment descriptors, providing valuable information for formulating machine learning atomic force fields. Future work will explore 5-ph and higher nn-ph interactions in strongly anharmonic materials, use PCP to obtain simplified coupled-mode equations for nonlinear phonon processes, and accelerate simulations of ultrafast phonon dynamics.

Acknowledgments

Y.L. thanks Junjie Yang for fruitful discussions. This work was supported by the National Science Foundation under Grant No. OAC-2209262. Y.L. acknowledges support from the Eddleman Fellowship. This research used resources of the National Energy Research Scientific Computing Center, a DOE Office of Science User Facility supported by the Office of Science of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231 using NERSC award NERSC DDR-ERCAP0026831.

Supplemental Material for “Tensor Learning and Compression of N-phonon Interactions”

Yao Luo; Dhruv Mangtani; Shiyu Peng; Jia Yao; Sergei Kliavinek; Marco Bernardi

Supplemental Material for “Tensor Learning and Compression of N-phonon Interactions”

Derivation of nn-ph interactions and their PCP decomposition

The nn-IFC tensor in PCP format is

Φ~l1b1,…,lnbnα1,…,αn=∑ξ=1Nc(n)λξn!∑σ∈Sn∑l∏i=1nAσiξ(li−l,biαi),(8)\widetilde{\Phi}^{\alpha_1,\ldots,\alpha_n}_{l_1b_1,\ldots,l_nb_n} = \sum_{\xi=1}^{N_c^{(n)}} \frac{\lambda_\xi}{n!} \sum_{\sigma\in S_n}\sum_l\prod_{i=1}^{n}A^{\xi}_{\sigma_i}(l_i-l,b_i\alpha_i), \tag*{(8)}

where ll, bb, and α\alpha index the primitive cell, basis atom, and Cartesian coordinate, respectively.

In the following, we show that this expression is consistent with the PCP decomposition of nn-ph interactions, V~(n)\widetilde{V}^{(n)}, in Eq. (2) of main text.

The compressed nn-ph interaction in momentum space, V~(n)\widetilde{V}^{(n)}, is obtained from a basis transformation of Φ~\widetilde{\Phi}:

V~(n)(Q1,…,Qn)=1N∑l1b1,α1eα1b1Q1mb1eiq1rl1⋯∑lnbn,αneαnbnQnmbneiqnrlnΦ~l1b1,…,lnbnα1,…,αn,(9)\widetilde{V}^{(n)}(\mathbf{Q}_1,\ldots,\mathbf{Q}_n) = \frac{1}{N} \sum_{l_1b_1,\alpha_1} \frac{e^{\mathbf{Q}_1}_{\alpha_1b_1}}{\sqrt{m_{b_1}}}e^{i\mathbf{q}_1\mathbf{r}_{l_1}} \cdots \sum_{l_nb_n,\alpha_n} \frac{e^{\mathbf{Q}_n}_{\alpha_nb_n}}{\sqrt{m_{b_n}}}e^{i\mathbf{q}_n\mathbf{r}_{l_n}} \widetilde{\Phi}^{\alpha_1,\ldots,\alpha_n}_{l_1b_1,\ldots,l_nb_n}, \tag*{(9)}

where NN is the number of primitive cells in a Born–von Kármán (BvK) supercell. Substituting Eq. (8) into Eq. (9), and exchanging the order of ∏\prod and ∑\sum, we get

V~(n)(Q1,…,Qn)=∑ξ=1Nc(n)λξn!∑σ∈Sn1N∑l∏i=1n(∑libi,αieαibiQimbieiqirliAσiξ(li−l,biαi)).(10)\widetilde{V}^{(n)}(\mathbf{Q}_1,\ldots,\mathbf{Q}_n) = \sum_{\xi=1}^{N_c^{(n)}}\frac{\lambda_\xi}{n!} \sum_{\sigma\in S_n}\frac{1}{N}\sum_l\prod_{i=1}^{n} \left( \sum_{l_ib_i,\alpha_i} \frac{e^{\mathbf{Q}_i}_{\alpha_ib_i}}{\sqrt{m_{b_i}}} e^{i\mathbf{q}_i\mathbf{r}_{l_i}} A^{\xi}_{\sigma_i}(l_i-l,b_i\alpha_i) \right). \tag*{(10)}

Let us define the PCP modes in momentum space,

Aσiξ(Q)=∑li′bi,αieαibiQimbieiqirli′Aσiξ(li′,biαi).A^{\xi}_{\sigma_i}(\mathbf{Q}) = \sum_{l'_ib_i,\alpha_i} \frac{e^{\mathbf{Q}_i}_{\alpha_ib_i}}{\sqrt{m_{b_i}}} e^{i\mathbf{q}_i\mathbf{r}_{l'_i}} A_{\sigma_i}^{\xi}(l'_i,b_i\alpha_i).

After the change of variable li′=li−ll'_i=l_i-l, we obtain

V~(n)(Q1,…,Qn)=∑ξ=1Nc(n)λξn!∑σ∈Sn1N∑l∏i=1n(eiqirlAσiξ(Qi))(11)\widetilde{V}^{(n)}(\mathbf{Q}_1,\ldots,\mathbf{Q}_n) = \sum_{\xi=1}^{N_c^{(n)}}\frac{\lambda_\xi}{n!} \sum_{\sigma\in S_n}\frac{1}{N}\sum_l\prod_{i=1}^{n} \left(e^{i\mathbf{q}_i\mathbf{r}_l}A^{\xi}_{\sigma_i}(\mathbf{Q}_i)\right) \tag*{(11)}
=∑ξ=1Nc(n)λξn!∑σ∈Sn1N∑lei∑i=1nqirl∏i=1nAσiξ(Qi)(12)= \sum_{\xi=1}^{N_c^{(n)}}\frac{\lambda_\xi}{n!} \sum_{\sigma\in S_n}\frac{1}{N}\sum_l e^{i\sum_{i=1}^{n}\mathbf{q}_i\mathbf{r}_l} \prod_{i=1}^{n}A^{\xi}_{\sigma_i}(\mathbf{Q}_i) \tag*{(12)}
=∑ξ=1Nc(n)λξn!∑σ∈Snδ(∑i=1nqi)∏i=1nAσiξ(Qi).(13)= \sum_{\xi=1}^{N_c^{(n)}}\frac{\lambda_\xi}{n!} \sum_{\sigma\in S_n} \delta\left(\sum_{i=1}^{n}\mathbf{q}_i\right) \prod_{i=1}^{n}A^{\xi}_{\sigma_i}(\mathbf{Q}_i). \tag*{(13)}

Using the property that a permanent is invariant under matrix transpose, we exchange ii and σi\sigma_i in Aσiξ(Qi)A^{\xi}_{\sigma_i}(\mathbf{Q}_i), and obtain the nn-ph interactions in momentum space in PCP format:

V~(n)(Q1,…,Qn)=δ(∑i=1nqi)∑ξ=1Nc(n)λξn!∑σ∈Sn∏i=1nAiξ(Qσi),(14)\widetilde{V}^{(n)}(\mathbf{Q}_1,\ldots,\mathbf{Q}_n) = \delta\left(\sum_{i=1}^{n}\mathbf{q}_i\right) \sum_{\xi=1}^{N_c^{(n)}}\frac{\lambda_\xi}{n!} \sum_{\sigma\in S_n}\prod_{i=1}^{n}A_i^{\xi}(\mathbf{Q}_{\sigma_i}), \tag*{(14)}

which is Eq. (2) in the main text.

Anharmonic energy E(n)E^{(n)} in PCP decomposition

Using the definition of the nn-IFCs, the nnth-order anharmonic energy E(n)(u)E^{(n)}(\mathbf{u}) is:

E(n)(u)=1n!∑l1b1α1⋯∑lnbnαnΦ~l1b1,…,lnbnα1,…,αnu(l1b1α1)u(l2b2α2)…u(lnbnαn).(15)E^{(n)}(\mathbf{u}) = \frac{1}{n!} \sum_{l_{1}b_{1}\alpha_{1}} \cdots\sum_{l_{n}b_{n}\alpha_{n}} \widetilde{\Phi}^{\alpha_{1},\ldots,\alpha_{n}}_{l_{1}b_{1},\ldots,l_{n}b_{n}} u(l_{1}b_{1}\alpha_{1})u(l_{2}b_{2}\alpha_{2})\ldots u(l_{n}b_{n}\alpha_{n}). \tag*{(15)}

Substituting Eq. (8) into the equation above, we obtain

E(n)(u)=1n!∑l1b1α1u(l1b1α1)…∑lnbnαnu(lnbnαn)∑ξ=1Nc(n)λξn!∑σ∈Sn∏i=1nAσiξ(li−l,biαi).(16)E^{(n)}(\mathbf{u}) = \frac{1}{n!} \sum_{l_{1}b_{1}\alpha_{1}} u(l_{1}b_{1}\alpha_{1})\ldots\sum_{l_{n}b_{n}\alpha_{n}} u(l_{n}b_{n}\alpha_{n}) \sum_{\xi=1}^{N_c^{(n)}} \frac{\lambda_{\xi}}{n!} \sum_{\sigma\in S_n} \prod_{i=1}^{n} A_{\sigma_i}^{\xi}(l_i-l,b_i\alpha_i). \tag*{(16)}

Exchanging the order of ∏\prod and ∑\sum in Eq. (10), we rewrite this equation as

E(n)(u)=∑l∑ξ=1Nc(n)λξn!1n!∑σ∈Sn∏i=1n(∑libiαiAσiξ(li−l,biαi)u(libiαi)).(17)E^{(n)}(\mathbf{u}) = \sum_l \sum_{\xi=1}^{N_c^{(n)}} \frac{\lambda_{\xi}}{n!}\frac{1}{n!} \sum_{\sigma\in S_n} \prod_{i=1}^{n} \left(\sum_{l_i b_i\alpha_i} A_{\sigma_i}^{\xi}(l_i-l,b_i\alpha_i)u(l_i b_i\alpha_i)\right). \tag*{(17)}

To simplify this expression, we define the projections appearing in Eq. (7) of main text:

ϕiξ(l,u)=⟨Aiξ(l),u⟩=∑l′bαAσiξ(l′−l,bα)u(l′bα).(18)\phi_i^{\xi}(l,\mathbf{u}) = \langle A_i^{\xi}(l),\mathbf{u}\rangle= \sum_{l'b\alpha} A_{\sigma_i}^{\xi}(l'-l,b\alpha)u(l'b\alpha). \tag*{(18)}

In terms of these quantities, the nnth-order anharmonic energy E(n)(u)E^{(n)}(\mathbf{u}) becomes

E(n)(u)=∑l∑ξ=1Nc(n)λξn!1n!∑σ∈Sn∏i=1nϕσiξ(l,u).(19)E^{(n)}(\mathbf{u}) = \sum_l \sum_{\xi=1}^{N_c^{(n)}} \frac{\lambda_{\xi}}{n!}\frac{1}{n!} \sum_{\sigma\in S_n} \prod_{i=1}^{n} \phi_{\sigma_i}^{\xi}(l,\mathbf{u}). \tag*{(19)}

Since ∏i=1nϕσiξ(l,u)\prod_{i=1}^{n}\phi_{\sigma_i}^{\xi}(l,\mathbf{u}) is independent of σ\sigma, we can further simplify this expression to

E(n)(u)=∑l∑ξ=1Nc(n)λξn!∏i=1nϕiξ(l,u),(20)E^{(n)}(\mathbf{u}) = \sum_l \sum_{\xi=1}^{N_c^{(n)}} \frac{\lambda_{\xi}}{n!} \prod_{i=1}^{n} \phi_i^{\xi}(l,\mathbf{u}), \tag*{(20)}

which is Eq. (6) in the main text.

Computational cost analysis

From a computational viewpoint, Eq. (2) in main text can greatly accelerate calculations of ph-ph interactions. The main bottleneck in such calculations is computing the nn-ph coupling V(n)(Q1,Q2,…,Qn)V^{(n)}(\mathbf{Q}_1,\mathbf{Q}_2,\ldots,\mathbf{Q}_n) for many scattering channels starting from the nn-IFC, Φ(n)\Phi^{(n)}. When using the uncompressed nn-ph interactions in Eq. (1), the computational cost scales as O(NΦNchannel)O(N_{\Phi}N_{\mathrm{channel}}), where NΦN_{\Phi} is the number of non-zero entries in Φ(n)\Phi^{(n)} and NchannelN_{\mathrm{channel}} is the number of active scattering channels in V(n)(Q1,Q2,…,Qn)V^{(n)}(\mathbf{Q}_1,\mathbf{Q}_2,\ldots,\mathbf{Q}_n). In contrast, utilizing the compressed nn-phonon interaction in Eq. (2), the same calculation has a cost that scales as O(Nc(n)Nchannel)\mathcal{O}(N_c^{(n)}N_{\mathrm{channel}}), where Nc(n)N_c^{(n)} is the PCP rank. Therefore, the computational cost savings from using Eq. (1) instead of Eq. (2) is proportional to the compression factor, γ(n)=NΦ/Nc(n)\gamma^{(n)}=N_{\Phi}/N_c^{(n)}. In Figure 4 of main text, we compare calculations using full (uncompressed) and compressed nn-ph interactions. The two sets of calculations are identical except for the method used to calculate V(n)(Q1,Q2,…,Qn)V^{(n)}(\mathbf{Q}_1,\mathbf{Q}_2,\ldots,\mathbf{Q}_n). Therefore, the computational speed-up in Figure 4 is entirely due to the more efficient computation of V(n)(Q1,Q2,…,Qn)V^{(n)}(\mathbf{Q}_1,\mathbf{Q}_2,\ldots,\mathbf{Q}_n) when using compressed nn-ph interactions.

In thermal conductivity calculations, the speed-up analysis is more complex. The total CPU time for computing phonon scattering rates (and thus the thermal conductivity) can be split into two contributions, with respective computational cost:

T1=cost of evaluating the n-phonon interactions V(n)(Q1,Q2,…,Qn),T_1 = \text{cost of evaluating the } n\text{-phonon interactions } V^{(n)}(\mathbf{Q}_1,\mathbf{Q}_2,\ldots,\mathbf{Q}_n),
T2=cost of computing the scattering rates from those interaction tensors.T_2 = \text{cost of computing the scattering rates from those interaction tensors.}

Our PCP method only accelerates T1T_1, reducing it to T1kγn\frac{T_1}{k\gamma_n}, where γn\gamma_n is the compression factor and kk is a small pre-factor (in practice k≈2k \approx2). The cost T2T_2 is left unchanged. Therefore, the speedup for the phonon scattering rate calculation, defined as the ratio of CPU wall times, is

speedup⁡=T1+T2T1/(kγn)+T2.\operatorname{speedup} = \frac{T_1+T_2}{T_1/(k\gamma_n)+T_2}.

In conventional calculations of phonon scattering rates, evaluating nn-phonon interactions is the bottleneck and T2T_2 is negligible. However, since our PCP approach reduces T1T_1 very efficiently, T2T_2 eventually becomes non-negligible. Consequently, there is an upper bound for the speedup:

speedup⁡<T1+T2T2≈T1T2.\operatorname{speedup} < \frac{T_1+T_2}{T_2} \approx\frac{T_1}{T_2}.

For example, in the case of Si, T1/(kγn)T_1/(k\gamma_n) is smaller than T2T_2, so the observed speedup is dominated by the fixed cost T2T_2 and thus it is close to the upper bound:

speedup⁡=T1+T2T1/(kγn)+T2≈T1+T2T2≈260.\operatorname{speedup} = \frac{T_1+T_2}{T_1/(k\gamma_n)+T_2} \approx\frac{T_1+T_2}{T_2} \approx260.

For all cubic materials with small unit cells discussed in the manuscript, the wall time of the calculation is dominated by the fixed cost T2T_2, and thus it is close to the upper bound T1/T2T_1/T_2. For ZrO2_2, which has 12 atoms in the unit cell, we use a much larger PCP rank (Nc=144N_c=144), so that T1/(kγn)>T2T_1/(k\gamma_n)>T_2 and the observed speedup approaches a different limit where the fixed overhead cost T2T_2 becomes irrelevant:

speedup⁡=T1+T2T1/(kγn)+T2≈kγ3≈7200.\operatorname{speedup}=\frac{T_1+T_2}{T_1/(k\gamma_n)+T_2}\approx k\gamma_3\approx7200.

This value reflects the true speedup of our PCP approach, which is close to twice the compression ratio: k≈2k\approx2, γ3≈4000\gamma_3\approx4000, and thus a speedup close to 8000.

DFT calculations and force constants generation

We generate displacement-force datasets using the Alamode package [1]. For all materials studied here, we sample 150 random configurations at 300 K using the harmonic phonon dispersion, and calculate forces for these randomly sampled configurations with the PBEsol functional [2] using VASP [3, 4] with a 2×2×22 \times2 \times2 k-point grid for all supercells. To extract the IFCs, we employ the adaptive LASSO solver implemented in Alamode [1], with hyperparameters optimized by the cross-validation method. For Si, we use a 5×5×55 \times5 \times5 supercell containing 250 atoms, with a unit cell lattice constant of 5.43 Å and a kinetic energy cutoff of 500 eV. For HgTe, we use a 4×4×44 \times4 \times4 supercell containing 128 atoms, with a unit cell lattice constant of 6.52 Å and a kinetic energy cutoff of 300 eV. For MgO, we use a 3×3×33 \times3 \times3 supercell containing 250 atoms, with a unit cell lattice constant of 4.25 Å and a kinetic energy cutoff of 400 eV. For TiNiSn, we use a 2×2×22 \times2 \times2 supercell containing 96 atoms, with a unit cell lattice constant of 5.87 Å and a kinetic energy cutoff of 500 eV. For ZrO2\mathrm{ZrO_2}, we use a 2×2×22 \times2 \times2 supercell containing 96 atoms, with a relaxed unit cell lattice constant of a=4.94a=4.94 Å, b=5.16b=5.16 Å, c=5.08c=5.08 Å and a kinetic energy cutoff of 350 eV.

Phonon scattering rate calculations

We compute the phonon scattering rate as the sum of 3-ph and 4-ph scattering rates [5]. The 3-ph scattering rate τ3ph,Q−1\tau^{-1}_{3\mathrm{ph},\mathbf{Q}}, for a phonon mode Q=νq\mathbf{Q}=\nu\mathbf{q}, reads

τ3ph,Q−1=πℏ4Nq∑q∑Q1∑Q2∣V(3)(Q1,Q2,Q3)∣21ωQωQ1ωQ2[12(1+nQ1+nQ2)δ(ωQ−ωQ1−ωQ2)+(nQ1−nQ2)δ(ωQ+ωQ1−ωQ2)].(21)\tau^{-1}_{3\mathrm{ph},\mathbf{Q}}=\frac{\pi\hbar}{4N_q}\sum_{\mathbf{q}}\sum_{\mathbf{Q}_1}\sum_{\mathbf{Q}_2}\left|V^{(3)}(\mathbf{Q}_1,\mathbf{Q}_2,\mathbf{Q}_3)\right|^2\frac{1}{\omega_{\mathbf{Q}}\omega_{\mathbf{Q}_1}\omega_{\mathbf{Q}_2}} \left[\frac{1}{2}(1+n_{\mathbf{Q}_1}+n_{\mathbf{Q}_2})\delta(\omega_{\mathbf{Q}}-\omega_{\mathbf{Q}_1}-\omega_{\mathbf{Q}_2})+(n_{\mathbf{Q}_1}-n_{\mathbf{Q}_2})\delta(\omega_{\mathbf{Q}}+\omega_{\mathbf{Q}_1}-\omega_{\mathbf{Q}_2})\right]. \tag*{(21)}

The 4-ph scattering rate τ4ph,Q−1\tau^{-1}_{4\mathrm{ph},\mathbf{Q}}, for a phonon mode Q=νq\mathbf{Q}=\nu\mathbf{q}, reads

\tau^{-1}_{4\mathrm{ph},\mathbf{Q}}=\frac{\pi\hbar}{4N_q}\frac{\hbar}{2N_q}\sum_{\mathbf{q}_1}\sum_{\mathbf{q}_2}\sum_{\mathbf{q}_3}\left|V^{(4)}(\mathbf{Q}_1,\mathbf{Q}_2,\mathbf{Q}_3,\mathbf{Q}_4)\right|^2\frac{1}{\omega_{\mathbf{Q}}\omega_{\mathbf{Q}_1}\omega_{\mathbf{Q}_2}\omega_{\mathbf{Q}_3}} \begin{aligned} &\times\left[\frac{1}{6}\frac{n_{\mathbf{Q}_1}n_{\mathbf{Q}_2}n_{\mathbf{Q}_3}}{n_{\mathbf{Q}}}\delta(\omega_{\mathbf{Q}}-\omega_{\mathbf{Q}_1}-\omega_{\mathbf{Q}_2}-\omega_{\mathbf{Q}_3})\\ &+\frac{1}{2}\frac{(1+n_{\mathbf{Q}_1})n_{\mathbf{Q}_2}n_{\mathbf{Q}_3}}{n_{\mathbf{Q}}}\delta(\omega_{\mathbf{Q}}+\omega_{\mathbf{Q}_1}-\omega_{\mathbf{Q}_2}-\omega_{\mathbf{Q}_3})\\ &+\frac{1}{2}\frac{(1+n_{\mathbf{Q}_1})(1+n_{\mathbf{Q}_2})n_{\mathbf{Q}_3}}{n_{\mathbf{Q}}}\delta(\omega_{\mathbf{Q}}+\omega_{\mathbf{Q}_1}+\omega_{\mathbf{Q}_2}-\omega_{\mathbf{Q}_3})\right], \end{aligned} \tag*{(22)}

where NqN_q is the number of q-points in the phonon momentum grid, ωQ\omega_{\mathbf{Q}} is the energy of the phonon mode Q\mathbf{Q} and nQn_{\mathbf{Q}} the corresponding Bose-Einstein thermal occupation. The delta function δ(x)\delta(x) is approximated by a Gaussian function e−(x/ϵ)2e^{-(x/\epsilon)^2}, where ϵ\epsilon is the Gaussian smearing parameter. We set ϵ=0.5\epsilon=0.5 meV, 0.20.2 meV, 0.20.2 meV, 0.20.2 meV, 1.01.0 meV for Si, HgTe, MgO, TiNiSn and ZrO2_2, respectively. We employ the random sampling technique in Ref. [6] to speed up the integration. For Si, HgTe, MgO and TiNiSn, we sample 2×1052\times10^5 scattering processes for both 3- and 4-ph scattering rates for each τQ−1\tau^{-1}_{\mathbf{Q}}, which ensures convergence of the phonon scattering rates. For ZrO2_2, we sample 4×1044\times10^4 scattering processes for 3-ph interactions and 4×1054\times10^5 scattering processes for 4-ph interactions.

Constrained optimization

For acoustic phonons, the ph-ph interactions vanish in the long-wavelength limit due to the acoustic sum rule (ASR). However, the thermal occupation of acoustic phonons can be large at finite temperature and diverge in the long-wavelength limit. This interplay results in acoustic phonon scattering being non-negligible, even though their interaction matrix elements are vanishingly small. We address this problem using a constrained optimization approach inspired by a method we previously developed for compressing ee-ph interactions using SVD [7]. The ASR for the 3-IFCs gives the constraint

∑l1b1Φl1b1,l2b2,l3b3α1,α2,α3=0.(23)\sum_{l_1b_1}\Phi^{\alpha_1,\alpha_2,\alpha_3}_{l_1b_1,l_2b_2,l_3b_3}=0. \tag*{(23)}

To enforce the acoustic sum rule [8], we impose ∑l,bAiξ(l,bα)=0\sum_{l,b}A^\xi_i(l,b\alpha)=0. This ASR leads to the following expansion in the long-wavelength limit (q→0\mathbf{q}\to0),

∑l1b1eiq1rl1Φl1b1,l2b2,l3b3α1,α2,α3≈iq1⋅∑l1b1Φl1b1,l2b2,l3b3α1,α2,α3rl1≈iq1⋅Fl2b2,l3b3α1,α2,α3[Φ].(24)\sum_{l_1b_1}e^{i\mathbf{q}_1\mathbf{r}_{l_1}}\Phi^{\alpha_1,\alpha_2,\alpha_3}_{l_1b_1,l_2b_2,l_3b_3}\approx i\mathbf{q}_1\cdot\sum_{l_1b_1}\Phi^{\alpha_1,\alpha_2,\alpha_3}_{l_1b_1,l_2b_2,l_3b_3}\mathbf{r}_{l_1}\approx i\mathbf{q}_1\cdot\mathbf{F}^{\alpha_1,\alpha_2,\alpha_3}_{l_2b_2,l_3b_3}[\Phi]. \tag*{(24)}

where

Fl2b2,l3b3α1,α2,α3[Φ]=∑l1b1Φl1b1,l2b2,l3b3α1,α2,α3rl1(25)\mathbf{F}^{\alpha_1,\alpha_2,\alpha_3}_{l_2b_2,l_3b_3}[\boldsymbol{\Phi}] = \sum_{l_1b_1} \boldsymbol{\Phi}^{\alpha_1,\alpha_2,\alpha_3}_{l_1b_1,l_2b_2,l_3b_3} r_{l_1} \tag*{(25)}

is the ph-ph deformation potential, which is analogous to the ee-ph deformation potentials discussed in Refs. [9, 10].

To preserve the ph-ph deformation potential F, we introduce a Lagrange multiplier in the loss function, which becomes:

L=∥Φ~[A]−Φ∥2+λF∥F[Φ~[A]]−F[Φ]∥2.(26)L = \left\lVert\widetilde{\boldsymbol{\Phi}}[A] - \boldsymbol{\Phi} \right\rVert^2 + \lambda_F \left\lVert\mathbf{F}[\widetilde{\boldsymbol{\Phi}}[A]] - \mathbf{F}[\boldsymbol{\Phi}] \right\rVert^2 . \tag*{(26)}

In our calculations, we set λF\lambda_F to a large value to numerically enforce the ph-ph deformation potential.

Phonon scattering rates for MgO and TiNiSn

Phonon scattering rates for MgO and TiNiSn, with panels (a) and (c) showing scattering rates versus phonon energy and panels (b) and (d) showing parity plots

Figure S1. Phonon scattering rates as a function of phonon energy for (a) MgO and (c) TiNiSn. The plots compare results computed with full IFC tensors and compressed IFC tensors (we use Nc(3)=24N_c^{(3)} = 24 for MgO and Nc(4)=48N_c^{(4)} = 48 for TiNiSn). Corresponding parity plots comparing results from compressed vs. full IFC tensors for 3+4-phonon scattering rates are given the same materials in panels (b) and (d).

Thermal conductivity results

We calculate the thermal conductivity κ\kappa at 300 K using momentum grids of 40340^3, 30330^3, 20320^3, 20320^3 and 10310^3 for Si, HgTe, MgO, TiNiSn, and ZrO2_2 respectively. In the following, we show the convergence of κ\kappa with respect to the PCP decomposition rank for the 3-ph and 4-ph interactions, respectively, in Table S3 and Table S2. For Nc(3)=24N_c^{(3)} = 24 and Nc(4)=48N_c^{(4)} = 48, the thermal conductivity predicted using compressed IFC tensors is within 98% of the value obtained using full 3- and 4-IFC tensors for all materials studied here.

CrystalsNc(4)=4N_c^{(4)} = 4Nc(4)=8N_c^{(4)} = 8Nc(4)=16N_c^{(4)} = 16Nc(4)=24N_c^{(4)} = 24Nc(4)=32N_c^{(4)} = 32Nc(4)=48N_c^{(4)} = 48Full Φ(3),Φ(4)\Phi^{(3)}, \Phi^{(4)}
Si123.2132.0130.7130.2130.3130.2130.9
HgTe1.792.142.042.032.192.222.24
MgO43.842.942.842.742.742.742.3
TiNiSn15.515.315.015.015.015.015.0

Table S2. Calculated thermal conductivity as in Table S1 above but with results given for different choices of PCP rank for 4-ph interactions while fixing the PCP rank for 3-ph interactions to Nc(3)=24N_c^{(3)} = 24. Bold font indicates thermal conductivities within 98% of the respective reference result obtained from full nn-IFC tensors (rightmost column).

CrystalsNc(3)=4N_c^{(3)} = 4Nc(3)=8N_c^{(3)} = 8Nc(3)=16N_c^{(3)} = 16Nc(3)=24N_c^{(3)} = 24Full Φ(3),Φ(4)\Phi^{(3)}, \Phi^{(4)}
Si116.6131.6130.7130.2130.9
HgTe2.132.132.202.222.24
MgO59.044.742.842.742.3
TiNiSn27.718.715.115.015.0

Table S1. Calculated thermal conductivity, in units of W/(mK), for Si, HgTe, MgO, TiNiSn. We compare reference results from full (uncompressed) nn-IFCs (rightmost column) with calculations using compressed 3-ph and 4-ph interactions with different choices of PCP rank for 3-ph interactions. The PCP rank for 4-ph interactions is fixed to Nc(4)=48N_c^{(4)} = 48. Thermal conductivities that are within 98% of the uncompressed reference result are shown with bold font.

Crystalsbravis lattice∥Φirr(3)∥0\lVert\Phi_{\mathrm{irr}}^{(3)} \rVert_0∥Φfull(3)∥0\lVert\Phi_{\mathrm{full}}^{(3)} \rVert_0PCP rank Nc(3)N_c^{(3)}γ(3)\gamma^{(3)}γsym(3)\gamma_{\mathrm{sym}}^{(3)}
SiCubic17454480242270313
HgTeCubic38354480242270142
MgOCubic21954480242270249
TiNiSnCubic515139230245801270
ZrO2_2Monoclinic13354483600144335836

Table S3. Summary of the compression factors of PCP methods and symmetrization for all the studied materials.

Extrapolation to the thermodynamic limit

Extrapolation of the thermal conductivity for HgTe at 300 K to the thermodynamic limit

Figure S2. Extrapolation of the thermal conductivity for HgTe at 300 K to the thermodynamic limit (TDL). The largest grid size we computed is 2503250^3, which corresponds to the leftmost data point. We extrapolate to the TDL by fitting the thermal conductivity values for the four largest grid sizes using κ(Nq)=κ(∞)−bNq−1/3\kappa(N_q) = \kappa(\infty) - bN_q^{-1/3}, where NqN_q is the number of q-points in the grid, and bb is the slope. Using this procedure, we obtain an extrapolated value of κ(∞)=2.5\kappa(\infty) = 2.5 W/(mK), which is 47% larger than the value using a grid size of 16316^3.

Origin of the slow convergence of thermal conductivity for HgTe

In HgTe, the 4-ph interactions give an important contribution. Without 4-ph scattering, the thermal conductivity is overestimated by around 400%, as has been reported previously [11]. In figure. S3, we calculate the thermal conductivity as a function of grid size for two values of the smearing parameter, 0.2 meV and 0.8 meV, with and without 4-ph interaction. We find that the thermal conductivity without 4-ph interactions is greatly overestimated, as is seen by comparing the y-axis values in the two figure panels. Figure. S3(a) shows the calculation without 4-ph interactions, where the thermal conductivity is nearly converged for a grid of 20320^{3} and is close to the results extrapolated to the thermodynamic limit. This confirms that the 3-ph interactions converge rapidly with the grid size. In Fig. S3(b), in the calculation with 4-ph interactions, the convergence with grid size is slow and persists up to much larger smearing values of 0.8 meV. Therefore, we attribute the slow convergence of the thermal conductivity in HgTe to the importance of the 4-ph processes, which are expected to be less smooth than the 3-ph scattering processes in the Brillouin zone and thus require denser grids to reach convergence.

Convergence of the thermal conductivity for HgTe

Figure S3. Convergence of the thermal conductivity for HgTe at 300 K with (a) 3-ph and (b) 3+4-ph interactions, using two smearing parameters. Blue and orange markers correspond to smearing values of 0.2 meV and 0.8 meV, respectively.

Number of force constants for 3rd order force constants

In this section, we compare the compression factor for our PCP technique, γ(3)\gamma^{(3)} defined in the manuscript, with the size reduction (or “compression”) resulting from symmetry, defined as

γsym(3)=∥Φfull(3)∥0∥Φirr(3)∥0.(27)\gamma_{\mathrm{sym}}^{(3)} = \frac{\lVert\Phi_{\mathrm{full}}^{(3)} \rVert_0}{\lVert\Phi_{\mathrm{irr}}^{(3)} \rVert_0}. \tag*{(27)}

where ∥x∥0\lVert x \rVert_0 indicates the L0L_0 norm, which measures the number of nonzero elements, and Φirr(3)\Phi_{\mathrm{irr}}^{(3)} is the irreducible IFC tensor, which accounts for the reduction of independent entries in the IFC tensor due to symmetry. In Table S3, we compare the PCP and symmetry-derived compression factors for all materials studied in this work, including ZrO2_2. The size reduction deriving from symmetry is of order 30–300 for all materials, and thus much smaller than the PCP compression factors, which are in the range of 2000–6000. In the case of monoclinic ZrO2_2, the symmetry-derived compression factor is relatively small (γsym(3)=36\gamma_{\mathrm{sym}}^{(3)} = 36) due to the low symmetry of ZrO2_2, while the PCP compression is nearly 100 times greater (to be precise, 93 ≈\approx 3358/36 times greater). This means that our tensor learning approach becomes even more favorable for low-symmetry crystals.

Symmetry loss from compression

We quantify the symmetry loss resulting from compression for the ii-th symmetry operation, S^i\hat{S}_i, using

ϵi=∑Q1,Q2,Q3∣V~(3)(Q1,Q2,Q3)−S^i[V~(3)](Q1,Q2,Q3)∣2∑Q1,Q2,Q3∣V~(3)(Q1,Q2,Q3)∣2(28)\epsilon_i = \frac{\sqrt{\sum_{\mathbf{Q}_1,\mathbf{Q}_2,\mathbf{Q}_3}\left|\widetilde{V}^{(3)}(\mathbf{Q}_1,\mathbf{Q}_2,\mathbf{Q}_3)-\hat{S}_i[\widetilde{V}^{(3)}](\mathbf{Q}_1,\mathbf{Q}_2,\mathbf{Q}_3)\right|^2}}{\sqrt{\sum_{\mathbf{Q}_1,\mathbf{Q}_2,\mathbf{Q}_3}\left|\widetilde{V}^{(3)}(\mathbf{Q}_1,\mathbf{Q}_2,\mathbf{Q}_3)\right|^2}} \tag*{(28)}

where S^i[V~(3)]\hat{S}_i[\widetilde{V}^{(3)}] is the compressed 3-ph interaction after applying the symmetry operation S^i\hat{S}_i. The symmetry transformation follows Ref. [12]. In Figure S4, we show ϵi\epsilon_i for each symmetry operation in the point group of Si. In all cases, the symmetry loss resulting from compression is less than 1%.

Relative symmetry loss for the compressed 3-ph interactions

Figure S4. Relative symmetry loss for the compressed 3-ph interactions in Si, using a compression factor γ(3)=2200\gamma^{(3)} = 2200 (Nc(3)=24N_{\mathrm{c}}^{(3)} = 24), given for each symmetry operations in the point group. In all cases, the symmetry loss, ϵi\epsilon_i defined above, is less than 1%. The first symmetry operation is the identity, which is always satisfied exactly. We use SPGLIB [13] to find all the point-group symmetry operations.

References

  1. [2]J. P. Perdew, A. Ruzsinszky, G. I. Csonka, O. A. Vydrov, G. E. Scuseria, L. A. Constantin, X. Zhou, and K. Burke, Restoring the density-gradient expansion for exchange in solids and surfaces, Phys. Rev. Lett. 100, 136406 (2008).pubmed.ncbi.nlm.nih.gov/18517979
  2. [3]G. Kresse and J. Furthmüller, Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set, Phys. Rev. B 54, 11169 (1996).pubmed.ncbi.nlm.nih.gov/9984901
  3. [4]G. Kresse and D. Joubert, From ultrasoft pseudopotentials to the projector augmented-wave method, Phys. Rev. B 59, 1758 (1999).DOI
  4. [5]T. Feng, L. Lindsay, and X. Ruan, Four-phonon scattering significantly reduces intrinsic thermal conductivity of solids, Phys. Rev. B 96, 161201 (2017).DOI
  5. [6]Z. Guo, Z. Han, D. Feng, G. Lin, and X. Ruan, Sampling-accelerated prediction of phonon scattering rates for converged thermal conductivity and radiative properties, Npj Comput. Mater. 10, 31 (2024).DOI
  6. [7]Y. Luo, D. Desai, B. K. Chang, J. Park, and M. Bernardi, Data-driven compression of electron-phonon interactions, Phys. Rev. X 14, 021023 (2024).DOI
  7. [8]K. Esfarjani and H. T. Stokes, Method to extract anharmonic force constants from first principles calculations, Phys. Rev. B 77, 144112 (2008).DOI
  8. [9]J. Bardeen and W. Shockley, Deformation potentials and mobilities in non-polar crystals, Phys. Rev. 80, 72 (1950).DOI
  9. [10]C. Herring and E. Vogt, Transport and deformation-potential theory for many-valley semiconductors with anisotropic scattering, Phys. Rev. 101, 944 (1956).
  10. [11]Y. Xia, V. I. Hegde, K. Pal, X. Hua, D. Gaines, S. Patel, J. He, M. Aykol, and C. Wolverton, High-throughput study of lattice thermal conductivity in binary rocksalt and zinc blende compounds including higher-order anharmonicity, Phys. Rev. X 10, 041029 (2020).DOI
  11. [12]A. A. Maradudin and S. H. Vosko, Symmetry properties of the normal vibrations of a crystal, Rev. Mod. Phys. 40, 1 (1968).DOI
  12. [13]K. S. Atsushi Togo and I. Tanaka, Spglib: a software library for crystal symmetry search, Sci. Technol. Adv. Mater., Meth. 4, 2384822 (2024).DOI

Paper details

vlnc:W_rvf4epar
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