betaFeedback

Reading v1View latest version

WorksPhysics and Astronomy

Theoretical study of high-order harmonic generation in solids based on deep-learning methods

Abstract

High-order harmonic generation (HHG) in solids provides a compact and coherent EUV radiation, which has attracted considerable attention. Many theoretical methods for laser-crystal interactions are developed. However, an approach to treat large-scale amorphous systems with ab-initio accuracy is still lacking because of the enormous computational cost. Here, we incorporate molecular dynamics and tight-binding electron Hamiltonians based on deep-learning methods to solve the problem efficiently. Its accuracy is validated by comparing the HHG spectra of monolayer hexagonal boron nitride (h-BN) with those obtained using existing methods. Then we use it to study the temperature-dependent HHG in bulk Si. We uncover a distinct harmonic-order-dependent thermal response: low-order harmonics are enhanced with increasing temperature, whereas high-order harmonics are suppressed. The former is caused by the modification of the inter-site hopping in the system, while the latter is mainly influenced by the structural disorder induced by the thermal effects, in agreement with experimental observations. This work establishes a scalable computational strategy for investigating HHG in solids and provides new insights into the interplay between lattice fluctuations and ultrafast electron dynamics.

Introduction

Strong-field lightmatter interactions enable the generation and manipulation of non-equilibrium electronic dynamics on the intrinsic electronic timescale. High-order harmonic generation (HHG), as a highly nonlinear coherent radiation process, has emerged as an indispensable tool for probing electronic motion. In gaseous systems, HHG has been successfully applied to areas such as attosecond pulse generation and molecular orbital reconstruction [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]. Compared with gaseous media, solids exhibit much higher particle densities and, in crystalline materials, well-defined electronic band structures [13, 14, 15, 16, 17]. These characteristics enable distinct strong-field electron dynamics, making solid-state HHG a powerful probe of ultrafast electronic processes in condensed matter.

In recent years, experimental and theoretical investigations of solid-state HHG have advanced rapidly. Experimentally, measurements of HHG in semiconductor materials such as ZnO have disclosed intrinsic correlations between harmonic spectral features and underlying electronic band structures [14], demonstrating that HHG can serve as a vital spectroscopic tool for probing electronic structures and ultrafast dynamics in solids. However, most existing studies on solid-state HHG have focused on ideal crystal structures [18, 19, 20] or fixed temperature conditions [21]. In realistic materials, lattice thermal motion is inevitable at finite temperatures. Previous theoretical studies have demonstrated that lattice thermal fluctuations can modify the local electronic environment, potentially affecting electronic transitions [22, 23]. Recent experiments by Mokhtari et al. [24] have further revealed temperature-dependent variations of harmonic intensities in bulk Si, where certain harmonic orders decrease with increasing temperature. These findings highlight the importance of developing computational approaches capable of describing finite-temperature HHG responses under realistic material conditions.

To accurately describe finite-temperature effects, efficient sampling of thermally fluctuating atomic configurations is crucial. In recent years, machine-learning methods have been introduced to study the strong-field physics [25, 26]. Deep-learning potential-based molecular dynamics (DPMD) has provided an effective approach for large-scale atomistic simulations by integrating the high accuracy of first-principles calculations with the computational efficiency of classical molecular dynamics [27, 28]. Among various machine-learning interatomic potentials, neuroevolution potentials (NEP) implemented in the GPUMD package exhibit excellent accuracy and computational efficiency [29, 30, 31], enabling efficient sampling of temperature-dependent atomic configurations in complex material systems.

At present, finite-temperature investigations of solid-state HHG remain heavily constrained by computational methodologies. One-dimensional models, while capable of illustrating specific fundamental physics, fail to capture the complex electronic structures intrinsic to real three-dimensional materials [22]. Time-dependent density functional theory (TDDFT) provides an accurate description of electron dynamics. However, its computational cost becomes prohibitive when applied to systems containing large numbers of atoms or during extensive thermal configuration sampling [23]. On the other hand, conventional tight-binding models offer high computational efficiency but typically rely on empirical parameters [32, 33], rendering them ill-suited for describing complex structural perturbations induced by temperature variations.

Recently, deep-learning tight-binding methods, such as DeePTB [34, 35], have emerged as a compelling approach that offers a new avenue for addressing these challenges. By training on high-precision first-principles density functional theory (DFT) calculations, DeePTB constructs parametrized tight-binding Hamiltonians that rapidly predict electronic structures across diverse structural configurations, effectively bridging the gap between first-principles accuracy and large-scale computational efficiency. Furthermore, when constructed using an orthogonal localized orbital basis, the resulting deep-learning tight-binding Hamiltonians exhibit sparsity and favorable scalability, while providing a compact effective description of the relevant electronic degrees of freedom [36]. Combining DeePTB with the semiconductor Bloch equations (SBE) enables the simulation of electron dynamics driven by strong laser fields at finite temperatures, providing a promising route to systematically elucidate temperature-dependent solid-state HHG mechanisms.

In this work, the theoretical methods are presented in § 2, and the results are shown in § 4. We summarize the paper in Sec. IV.

Theory

In this section, we introduce the methods used in our calculations. Subsection II A describes the Hamiltonian matrix obtained by DeePTB, the others illustrate the time evolution of the system when the laser field is added.

Deep-Learning Tight-Binding Models {#sec:II A}

A distinct advantage of the deep-learning tight-binding model implemented in DeePTB, compared to other deep-learning electronic structure prediction approaches, is that it directly yields the Hamiltonian matrix in the Wannier gauge [34]. The Wannier basis set offers a significantly lower dimensionality than other linear combinations of atomic orbitals basis sets, while possessing orthogonal and sparse Hamiltonian properties [37]. These features create favorable conditions for accelerating subsequent time-evolution calculations.

For a supercell containing NN primitive cells, the Bravais lattice vectors can be expressed as Rn=∑i=13niai\mathbf{R}_n = \sum_{i=1}^{3} n_i\mathbf{a}_i, where ai\mathbf{a}_i represents the primitive lattice vectors. A set of localized Wannier functions is given by ∣Rm⟩|\mathbf{R}m\rangle, where mm indexes the Wannier orbital. The set of Wannier functions forms an orthonormal basis, satisfying ⟨Rn∣0m⟩=δ0Rδmn\langle\mathbf{R}n|0m\rangle= \delta_{0R}\delta_{mn}. Then the Bloch-like functions are defined as

∣ϕmk(W)⟩=1N∑Reik⋅R∣Rm⟩,(1)|\phi_{mk}^{(W)}\rangle= \frac{1}{\sqrt{N}}\sum_{\mathbf{R}}e^{i\mathbf{k}\cdot\mathbf{R}}|\mathbf{R}m\rangle, \tag*{(1)}

and the Hamiltonian and dipole in the Wannier gauge are given by [33, 38]

Hmn,k(W)=⟨ϕmk(W)∣H^0∣ϕnk(W)⟩≡∑Reik⋅R⟨0m∣H^0∣Rn⟩,(2)H_{mn,k}^{(W)} = \langle\phi_{mk}^{(W)}|\hat{H}_0|\phi_{nk}^{(W)}\rangle\equiv\sum_{\mathbf{R}}e^{i\mathbf{k}\cdot\mathbf{R}}\langle0m|\hat{H}_0|\mathbf{R}n\rangle, \tag*{(2)}
Dmn,k(W)=⟨umk(W)∣i∇k∣unk(W)⟩≡∑Reik⋅R⟨0m∣r^∣Rn⟩,(3)D_{mn,k}^{(W)} = \langle u_{mk}^{(W)}|i\nabla_k|u_{nk}^{(W)}\rangle\equiv\sum_{\mathbf{R}}e^{i\mathbf{k}\cdot\mathbf{R}}\langle0m|\hat{\mathbf{r}}|\mathbf{R}n\rangle, \tag*{(3)}

where R\mathbf{R} runs over all lattice vectors, ∣unk(W)⟩|u_{nk}^{(W)}\rangle is the periodic part of the Bloch-like functions, H^0\hat{H}_0 and r^\hat{\mathbf{r}} are the unperturbed Hamiltonian and position operator. The superscript (W) specifies the Wannier gauge. During the band-fitting and prediction process in DeePTB, the centers of the Wannier orbitals do not shift relative to the atomic positions. We can transform the system into the Hamiltonian gauge by diagonalizing H(W)(k)H^{(W)}(\mathbf{k}) as [33, 38]

H(H)(k)=U†(k)H(W)(k)U(k),(4)H^{(H)}(\mathbf{k}) = U^{\dagger}(\mathbf{k})H^{(W)}(\mathbf{k})U(\mathbf{k}), \tag*{(4)}

and the corresponding matrix D(H)(k)D^{(H)}(\mathbf{k}) can be given by

D(H)(k)=U†(k)D(W)(k)U(k)+iU†(k)∂∂kU(k),(5)D^{(H)}(\mathbf{k}) = U^{\dagger}(\mathbf{k})D^{(W)}(\mathbf{k})U(\mathbf{k}) + iU^{\dagger}(\mathbf{k})\frac{\partial}{\partial k}U(\mathbf{k}), \tag*{(5)}

where H(H)H^{(H)} and D(H)D^{(H)} are the Hamiltonian matrix and transition dipole matrix under the Bloch basis, and U(k)U(\mathbf{k}) is the unitary matrix that diagonalizes H(W)H^{(W)}. The corresponding Bloch states in the Hamiltonian gauge are given by

∣ϕmk(H)⟩=∑mUmn(k)∣ϕnk(W)⟩.(6)|\phi_{mk}^{(H)}\rangle= \sum_m U_{mn}(\mathbf{k})|\phi_{nk}^{(W)}\rangle. \tag*{(6)}

In this work, we employ the DeePTB approach to construct the effective electronic Hamiltonian of the system, whose matrix element is ⟨0m∣H^0∣Rn⟩\langle0m|\hat{H}_0|\mathbf{R}n\rangle. Based on the SlaterKoster formalism, DeePTB combines conventional tight-binding theory with neural networks by introducing environment-dependent corrections to the tight-binding parameters. This formulation goes beyond the conventional two-center approximation, enabling a more accurate description of the influence of local structural variations on electronic interactions [36]. Furthermore, DeePTB employs local atomic environment descriptors that are invariant under translations, rotations, and permutations of equivalent atoms [39], while the SlaterKoster formalism ensures the proper transformation of the Hamiltonian under spatial rotations [36, 40]. The tight-binding parameters are optimized using electronic eigenvalues obtained from first-principles calculations, allowing the resulting Hamiltonian to capture the electronic structure in complex local atomic environments while preserving the underlying physical symmetries and orbital-coupling structure.

Density-Matrix-Based SBE

To describe the interaction between an intense laser field and an ideal crystal, we start with the single-particle time-dependent Schrödinger equation (TDSE). Within the electric dipole approximation and velocity gauge [41], the time evolution of electrons driven by the laser field is governed by

i∂ψ(r,t)∂t=H^(t)ψ(r,t)=[12(p+A(t))2+V(r)]ψ(r,t)=[H^0+p⋅A(t)+12A2(t)]ψ(r,t),(7)\begin{aligned} i\frac{\partial\psi(\mathbf{r},t)}{\partial t}=\hat{H}(t)\psi(\mathbf{r},t) \\ &=\left[\frac{1}{2}\left(\mathbf{p}+\mathbf{A}(t)\right)^2+V(\mathbf{r})\right]\psi(\mathbf{r},t) \\ &=\left[\hat{H}_0+\mathbf{p}\cdot\mathbf{A}(t)+\frac{1}{2}\mathbf{A}^2(t)\right]\psi(\mathbf{r},t), \tag*{(7)} \end{aligned}

where A(t)\mathbf{A}(t) is the vector potential of the laser field, V(r)V(\mathbf{r}) represents the periodic potential of the crystal. Using the acceleration theorem for Bloch electrons [42, 43, 44], we construct an adiabatic instantaneous eigenstate of Eq. (7). The Houston states can be defined as

ϕ~nk(r)=e−iA(t)⋅rϕnk(r),(8)\tilde{\phi}_{nk}(\mathbf{r})=e^{-i\mathbf{A}(t)\cdot\mathbf{r}}\phi_{nk}(\mathbf{r}), \tag*{(8)}

which satisfy the time-independent eigenvalue equation:

H^(t)ϕ~nk(r,t)=εnkϕ~nk(r,t).(9)\hat{H}(t)\tilde{\phi}_{nk}(\mathbf{r},t)=\varepsilon_{nk}\tilde{\phi}_{nk}(\mathbf{r},t). \tag*{(9)}

Then, the electronic wave function can be expanded in terms of these states:

ψ(r,t)=∫BZ∑nCnk(t)ϕ~nk(r,t)dk=∫BZ∑nCnk(t)e−iA⋅rϕnk(r)dk.(10)\begin{aligned} \psi(\mathbf{r},t)=\int_{BZ}\sum_n C_{nk}(t)\tilde{\phi}_{nk}(\mathbf{r},t)d\mathbf{k} \\ &=\int_{BZ}\sum_n C_{nk}(t)e^{-i\mathbf{A}\cdot\mathbf{r}}\phi_{nk}(\mathbf{r})d\mathbf{k}. \tag*{(10)} \end{aligned}

The corresponding density matrix elements are defined as

ρnmk(t)=Cm∗k(t)Cnk(t).(11)\rho^k_{nm}(t)=C^{*k}_{m}(t)C^k_{n}(t). \tag*{(11)}

Substituting Eqs. (8)–(11) into the TDSE yields the SBE in the Hamiltonian gauge:

∂ρnm(H)k0(t)∂t=i∑l[ρnl(H)k0(t)Hlm(H)k(t)−Hnl(H)k(t)ρlm(H)k0(t)]+iE(t)⋅∑l[ρnl(H)k0(t)Dlm(H)k(t)−Dnl(H)k(t)ρlm(H)k0(t)].(12)\frac{\partial\rho^{(H)k_0}_{nm}(t)}{\partial t} =i\sum_l\left[\rho^{(H)k_0}_{nl}(t)H^{(H)k(t)}_{lm}-H^{(H)k(t)}_{nl}\rho^{(H)k_0}_{lm}(t)\right] +i\mathbf{E}(t)\cdot\sum_l\left[\rho^{(H)k_0}_{nl}(t)D^{(H)k(t)}_{lm}-D^{(H)k(t)}_{nl}\rho^{(H)k_0}_{lm}(t)\right]. \tag*{(12)}

Then we can get the SBE in Wannier gauge through unitary transformation:

∂ρnm(W)k0(t)∂t=i∑l[ρnl(W)k0(t)Hlm(W)k(t)−Hnl(W)k(t)ρlm(W)k0(t)]+iE(t)⋅∑l[ρnl(W)k0(t)Dlm(W)k(t)−Dnl(W)k(t)ρlm(W)k0(t)],(13)\frac{\partial\rho^{(W)k_0}_{nm}(t)}{\partial t} =i\sum_l\left[\rho^{(W)k_0}_{nl}(t)H^{(W)k(t)}_{lm}-H^{(W)k(t)}_{nl}\rho^{(W)k_0}_{lm}(t)\right] +i\mathbf{E}(t)\cdot\sum_l\left[\rho^{(W)k_0}_{nl}(t)D^{(W)k(t)}_{lm}-D^{(W)k(t)}_{nl}\rho^{(W)k_0}_{lm}(t)\right], \tag*{(13)}

where E(t)=−∂A(t)∂t\mathbf{E}(t)=-\frac{\partial\mathbf{A}(t)}{\partial t} is the electric field of the laser pulse, the crystal momentum evolves as k(t)=k0+A(t)\mathbf{k}(t)=\mathbf{k}_0+\mathbf{A}(t), and k0\mathbf{k}_0 denotes the initial crystal momentum [45, 46, 47].

In SBE, phenomenological dephasing terms are typically introduced to account for dephasing effect induced by electron–electron and electron–phonon scattering, thereby yielding theoretical results in better agreement with experimental observations [15]. Numerically, the dephasing relaxation is handled naturally within the Hamiltonian gauge. Specifically, we propagate the coherent evolution equations in the Wannier gauge [33] while incorporating the dephasing terms in the Hamiltonian gauge. As shown in Eq. (14), the equations of motion in the Hamiltonian gauge are formulated as [32]

∂ρnm(H)k0(t)∂t=i∑l[ρnl(H)k0(t)Hlm(H)k(t)−Hnl(H)k(t)ρlm(H)k0(t)]+iE(t)⋅∑l[ρnl(H)k0(t)Dlm(H)k(t)−Dnl(H)k(t)ρlm(H)k0(t)]−1−δnmT2ρnm(H)k0(t).(14)\frac{\partial\rho^{(H)k_0}_{nm}(t)}{\partial t} =i\sum_l\left[\rho^{(H)k_0}_{nl}(t)H^{(H)k(t)}_{lm}-H^{(H)k(t)}_{nl}\rho^{(H)k_0}_{lm}(t)\right] +i\mathbf{E}(t)\cdot\sum_l\left[\rho^{(H)k_0}_{nl}(t)D^{(H)k(t)}_{lm}-D^{(H)k(t)}_{nl}\rho^{(H)k_0}_{lm}(t)\right] -\frac{1-\delta_{nm}}{T_2}\rho^{(H)k_0}_{nm}(t). \tag*{(14)}

The Crank-Nicolson scheme is used for the time evolution, and the total current is given by

J(t)=−∫BZTr⁡[ρ(W)k0(t)v(W)k0(t)]dk0,(15)\mathbf{J}(t)=-\int_{BZ}\operatorname{Tr}\left[\rho^{(W)k_0}(t)\mathbf{v}^{(W)k_0}(t)\right]d\mathbf{k}_0, \tag*{(15)}

where the elements of the velocity matrix v(W)k0(t)\mathbf{v}^{(W)k_0}(t) which can be calculated from Heisenberg equation of motion is decomposed as

v(W)k0(t)=∂H(W)k(t)∂k(t)−i[D(W)k(t),H(W)k(t)].(16)\mathbf{v}^{(W)k_0}(t)=\frac{\partial H^{(W)k(t)}}{\partial\mathbf{k}(t)}-i\left[D^{(W)k(t)},H^{(W)k(t)}\right]. \tag*{(16)}

However, we find that the matrix elements of the position operator are often neglected in previous works [48, 49]. In this case, Dmn(W)k(t)D^{(W)k(t)}_{mn} in Eq. (3) is zero and Eq. (13) becomes

∂ρnm(W)k0(t)∂t=i∑l[ρnl(W)k0(t)Hlm(W)k(t)−Hnl(W)k(t)ρlm(W)k0(t)],(17)\frac{\partial\rho^{(W)k_0}_{nm}(t)}{\partial t} =i\sum_l\left[\rho^{(W)k_0}_{nl}(t)H^{(W)k(t)}_{lm}-H^{(W)k(t)}_{nl}\rho^{(W)k_0}_{lm}(t)\right], \tag*{(17)}

and the time-dependent current becomes

J(t)=−∫BZTr⁡[ρ(W)k0(t)∂H(W)k(t)∂k(t)]dk0.(18)\mathbf{J}(t)=-\int_{BZ}\operatorname{Tr}\left[\rho^{(W)k_0}(t)\frac{\partial H^{(W)k(t)}}{\partial\mathbf{k}(t)}\right]d\mathbf{k}_0. \tag*{(18)}

The HHG intensity spectrum I(ω)I(\omega) is proportional to the squared magnitude of the Fourier transform of the time derivative of the macroscopic current as

I(ω)∝∣F[J(t)]∣2.(19)I(\omega)\propto\left|\mathcal{F}\left[\mathbf{J}(t)\right]\right|^2. \tag*{(19)}

SBE evolution for occupied states

To reduce the computational cost and memory requirements, the SBE can alternatively be solved by propagating the single-particle states originating from the initially occupied Bloch states, without explicitly evolving the full density matrix. From Eq. (10), the expansion coefficient vector in the Houston basis is defined as

Cλk(H)(t)=(C1k(t)C2k(t)⋮CNk(t)).(20)\mathcal{C}^{(H)}_{\lambda\mathbf{k}}(t)=\begin{pmatrix} C^{\mathbf{k}}_{1}(t) \\ C^{\mathbf{k}}_{2}(t) \\ \vdots\\ C^{\mathbf{k}}_{N}(t) \end{pmatrix}. \tag*{(20)}

Here, NN denotes the total number of energy bands, and λ\lambda labels the electrons participating in the time evolution. According to the Pauli exclusion principle [50, 51] and the identity principle [52], λ\lambda also serves as the corresponding valence-band index.

By applying a unitary transformation, the expansion coefficients can be expressed in the Houston basis under the Wannier gauge as

Cλk0(W)(t)=U†(k(t))Cλk0(H)(t).(21)\mathcal{C}^{(W)}_{\lambda\mathbf{k}_0}(t)=U^{\dagger}(\mathbf{k}(t))\mathcal{C}^{(H)}_{\lambda\mathbf{k}_0}(t). \tag*{(21)}

Substituting Eq. (8), Eq. (20) and Eq. (21) into Eq. (7), we obtain

i∂Cλk0(W)(t)∂t=[Hk(t)(W)+E(t)⋅Dk(t)(W)]Cλk0(W)(t).(22)i\frac{\partial\mathcal{C}^{(W)}_{\lambda\mathbf{k}_0}(t)}{\partial t}=\left[H^{(W)}_{\mathbf{k}(t)}+E(t)\cdot D^{(W)}_{\mathbf{k}(t)}\right]\mathcal{C}^{(W)}_{\lambda\mathbf{k}_0}(t). \tag*{(22)}

It is worth noting that Eq. (22) and Eq. (13) are essentially equivalent. The Crank-Nicolson scheme is used for the time evolution too. Consequently, the electric current can be expressed as

J(t)=−∫BZ∑λ∈NoccCλk0(W)†(t)vk0(W)(t)Cλk0(W)(t) dk0,(23)J(t)=-\int_{\mathrm{BZ}}\sum_{\lambda\in N_{\mathrm{occ}}}\mathcal{C}^{(W)\dagger}_{\lambda\mathbf{k}_0}(t)v^{(W)}_{\mathbf{k}_0}(t)\mathcal{C}^{(W)}_{\lambda\mathbf{k}_0}(t)\,\mathrm{d}\mathbf{k}_0, \tag*{(23)}

where NoccN_{\mathrm{occ}} denotes the number of initially occupied states. As discussed in subsection II B, we ignore the term Dk(t)(W)D^{(W)}_{\mathbf{k}(t)}. Then we can get

J(t)=−∫BZ∑λ∈NoccCλk0(W)†(t)∂Hk(t)(W)∂k(t)Cλk0(W)(t) dk0.(24)J(t)=-\int_{\mathrm{BZ}}\sum_{\lambda\in N_{\mathrm{occ}}}\mathcal{C}^{(W)\dagger}_{\lambda\mathbf{k}_0}(t)\frac{\partial H^{(W)}_{\mathbf{k}(t)}}{\partial\mathbf{k}(t)}\mathcal{C}^{(W)}_{\lambda\mathbf{k}_0}(t)\,\mathrm{d}\mathbf{k}_0. \tag*{(24)}

Results and Discussion

Validation of DeePTB model for solid HHG in hexagonal boron nitride (h-BN)

Prior to performing large-scale simulations at finite temperatures, we first validate the reliability of DeePTB in constructing the electronic Hamiltonian to describe the strong-field response of solids. h-BN is chosen as a benchmark system due to the availability of extensive theoretical studies for comparing its strong-field response characteristics [33, 53]. We performed DFT calculations using ABACUS package with the pseudopotentials and orbitals from APNS to generate the electronic structure

dataset of the h-BN system [54, 55, 56], which was subsequently used to train the DeePTB model. We used the PBE functional for ABACUS calculations [57].

Figure 1 compares the band structure of h-BN obtained by DeePTB with the DFT calculation results. The DeePTB model accurately reproduces the main band features of the system, with particularly good agreement in the region near the Fermi level. This indicates that DeePTB is capable of effectively learning the electronic coupling relationships in the h-BN system, providing a reliable Hamiltonian for subsequent strong-field electron dynamics calculations.

Band structure of h-BN primitive cell given by DeePTB Hamiltonian

Figure 1. Band structure of h-BN primitive cell given by DeePTB Hamiltonian (red lines). The black dots correspond to the band structure calculated by ABACUS. EF means the Fermi energy.

Using the trained Wannier-gauge DeePTB Hamiltonian, the HHG responses along different crystallographic directions were further calculated and compared with previous theoretical results obtained from other approaches. The simulations were performed using a laser pulse with a peak field strength of E0=40E_0=40 MV/cm, a wavelength of 3 μm3\ \mu\mathrm{m}, and a full width at half maximum (FWHM) of 34.2 fs. A 100×100×1100\times100\times1 Monkhorst-Pack k-point grid was adopted, and the k-point sampling was verified to be converged. The pulse envelope was described by a cos⁡2\cos^2 function, and the dephasing time is 2 fs. To obtain clearer harmonic spectra, a Hanning window was applied to the calculated current before Fourier transformation. We use the method described in § 2 B to calculate the current and HHG of h-BN.

As shown in Figure 2, the HHG spectra along different directions driven by a linearly polarized laser are calculated by using the DeePTB-predicted Hamiltonian. They are in excellent agreement with the results obtained using the methods in Ref. [33]. The effective model [33] means the widely used simple tight-binding model in which only pzp_z orbitals are considered. The Wannier90 refers to the results in which the Hamiltonian for calculating the HHG was obtained directly by the Wannier90 software [37]. The DeePTB method reproduces both the main harmonic peak positions and overall spectral characteristics. This agreement demonstrates the reliability of the machine-learning Hamiltonian for solid-state strong-field electron dynamics simulations and establishes its potential for studying more complex large-scale systems.

HHG spectra for a single monolayer of h-BN

Figure 2. HHG spectra computed for a single monolayer of h-BN by the DeePTB (grey lines), the effective model (red lines) [33], and the Wannier90 based Hamiltonian, respectively. In (a) the laser pulse and the harmonic signal are polarized along the Γ\Gamma-M direction. In (b) and (c), the laser pulse is polarized along the Γ\Gamma-K direction, and the harmonic signals parallel and perpendicular to the this direction are presented, respectively. For clarity, the HHG spectra are vertically offset.

Workflow for bulk Si calculations

The workflow adopted in this study for the amorphous system is illustrated in Figure 3. Molecular dynamics simulations of Si unit-cell and supercell configurations at different temperatures are performed by using the GPUMD package [30, 31]. The simulations employed NEP89, a NEP model previously developed and trained in Refs. [29, 55]. High-throughput DFT calculations were subsequently carried out for the temperature sampled unit cell configurations. The DFT results from 50 temperature dependent configurations covering the range of 25–400 K were used as the training dataset to train DeePTB [35]. The trained model was then employed to predict the Hamiltonians of the supercell system at different temperatures. HHG calculations were performed based on these temperature-dependent Hamiltonians.

Workflow for calculating phonon-assisted optical spectra using NEP, GPUMD and DeePTB models

Figure 3. Workflow for calculating phonon-assisted optical spectra using NEP, GPUMD and DeePTB models [35]. NEP and GPUMD generate atomic structures through molecular dynamics simulations. Small-cell structures are then labeled by DFT eigenvalues to train the DeePTB model. Large-cell structures from GPUMD are then used with the trained DeePTB model to predict the Hamiltonians H(R)H(\mathbf{R}). Based on the Hamiltonians, current and HHG can be calculated. The NEP function we used is NEP89 [55]. The steps in the flowchart are indicated by circled numbers.

The details about our workflow are as follows. Initially, molecular dynamics simulations of bulk silicon unit cells are performed using GPUMD with NEP89 potential function over a temperature range of 25–400 K in canonical ensemble (NVT) and isothermalisobaric ensemble (NPT) to generate thermally distributed atomic configurations. The electronic band structures of these configurations are subsequently calculated using the ABACUS package. The resulting atomic configurations and their corresponding electronic band structures are then used to train the DeePTB model. Next, molecular dynamics simulations of 216-atom bulk silicon supercells are carried out using the same NEP89 potential and GPUMD over a temperature range of 100–300 K in NVT to sample thermally distorted supercell configurations at different temperatures. The trained DeePTB model is then employed to predict the corresponding tight-binding Hamiltonians for these configurations. Finally, the time-dependent electronic states are propagated using the Crank–Nicolson scheme, from which the time-dependent electric current and HHG are obtained.

Electronic structure and strong-field response of bulk Si described by DeePTB

Compared with the two-dimensional h-BN, bulk Si exhibits a more complex three-dimensional electronic structure and serves as an important system for studying strong-field responses in semiconductors. We performed DFT calculations using ABACUS with the pseudopotentials and orbitals from APNS to generate the electronic structure dataset of bulk Si system [54, 55, 56], which was subsequently used to train the DeePTB model. We used the PBE functional for ABACUS calculations [57]. First, the trained model described above was used to reproduce the band structure of the Si primary cell at 0 K, and the DeePTB results were compared with the DFT calculations by ABACUS [54] as presented in Figure 4. The DeePTB Hamiltonian accurately reproduces the main electronic structure features of Si, demonstrating its capability for describing the electronic properties of the system.

Band structure of bulk Si primitive cell by different methods

Figure 4. Band structure of bulk Si primitive cell by different methods. The red lines represent the band structure given by DeePTB, while the black dots correspond to the band structure calculated using ABACUS. EFE_\mathrm{F} means the Fermi energy.

Subsequently, as shown in Figure 5, the HHG response of the Si unit cell was calculated using the Hamiltonian by DeePTB and compared with previous TDDFT calculations by the software Octopus and periodic-model calculations in Ref. [58]. The real-space periodic model describes the motion of electrons in solids by introducing an effective periodic potential that reflects the periodicity of the crystal lattice. The simulations were performed using a laser pulse with a peak field strength of E0=10E_0 = 10 MV/cm, a wavelength of λ=2\lambda= 2 µm, and a FWHM of 30 fs. The vector potential of the laser field is given by

HHG spectra of bulk Si unit cell

Figure 5. HHG of bulk Si unit cell. Grey line is HHG calculated based on DeePTB, red line is calculated by the periodic model [58], and blue line is calculated by Octopus [58]. For clarity, the HHG spectra are vertically offset.

A(t)=E0ωsin⁡2(πt2tFWHM)sin⁡(ωt).(25)\mathbf{A}(t) = \frac{\mathbf{E}_0}{\omega}\sin^2\left(\frac{\pi t}{2t_\mathrm{FWHM}}\right)\sin(\omega t). \tag*{(25)}

The laser was incident normal to the crystal surface. A 32×32×3232 \times32 \times32 Monkhorst–Pack kk-point grid was adopted. In Subsections III C and III D, we use the method mentioned in Subsection II C to calculate the laser-induced current and HHG. The HHG spectra obtained using the DeePTB Hamiltonian are in good agreement with previous theoretical results [58]. Moreover, compared with the simple periodic model, the DeePTB results show closer agreement with the first-principles calculations.

Notably, this calculation requires approximately 8 hours on a single A100 GPU, demonstrating that the deep-learning tight-binding approach provides a favorable balance between accuracy and computational efficiency. This capability enables its application to large-scale systems in subsequent studies.

Finite-temperature HHG response of large-scale Si supercells

After establishing the reliability of the deep-learning tight-binding model for a unit cell, we further investigate the HHG response of large-scale Si systems under finite-temperature conditions. To assess the accuracy of DeePTB for Si supercells of different sizes, we compare its predicted band structures with those from DFT calculations at a fixed temperature of 300 K. The comparison is based on the distribution of the error energy and the mean absolute error (MAE). The error energy is defined as

ΔE=EnkDFT−EnkDeePTB,(26)\Delta E = E_{nk}^{DFT} - E_{nk}^{DeePTB}, \tag*{(26)}

and MAE is defined as

MAE=1NkNband∑k=1Nk∑n=1Nband∣EnkDFT−EnkDeePTB∣,(27)\mathrm{MAE} = \frac{1}{N_kN_{\mathrm{band}}}\sum_{k=1}^{N_k}\sum_{n=1}^{N_{\mathrm{band}}}\left|E_{nk}^{DFT}-E_{nk}^{DeePTB}\right|, \tag*{(27)}

where NkN_k is the number of kk-points per band, NbandN_{\mathrm{band}} is the number of bands we mainly considered. EnkDeePTBE_{nk}^{DeePTB} and EnkDFTE_{nk}^{DFT} denote the energies of the nnth band and kkth kk point predicted by DeePTB and calculated by ABACUS, respectively. MAE can quantitatively demonstrate the predictive accuracy. Typically, MAE values lower than 50 meV indicate that the error is negligible compared with the typical energy scale of electronic bands [59]. As shown in Figure 6, the MAE remains nearly

Error histograms for bulk Si supercells containing 8, 64, and 216 atoms

Figure 6. Generalization and accuracy of DeePTB at different scales of bulk Si supercell. (a-c) The error histogram of bands between the results by DeePTB and the corresponding DFT calculations with a cell size of 8, 64, and 216 atoms, respectively. We count the energy differences for a number of bands equal to four times the number of atoms, sampling 302 kk-points for each band.

unchanged when the system size increases, proving that DeePTB captures local atomic environment effects and can be generalized from small cells to large supercells. DeePTB maintains high predictive accuracy for Si supercells with various numbers of atoms, indicating that the method can effectively capture the variation of electronic structures induced by thermal fluctuations.

Considering the balance between computational accuracy and efficiency, we select a Si supercell containing 216 atoms for the temperature-dependent HHG calculations. We employ laser pulses with a peak electric field strength of E0=15E_0 = 15 MV/cm, a wavelength of λ=3.2\lambda= 3.2 μ\mum, an FWHM of 60 fs, and a cos⁡2\cos^2 pulse envelope. The laser pulse is incident perpendicular to the crystal surface. The time-dependent current responses at different temperatures are calculated using the SBE based on occupied-state propagation. A 11 ×\times 11 ×\times 11 MonkhorstPack kk-point grid was adopted, and five atomic configurations are sampled for each temperature. The results have been verified to achieve convergence. And this calculation requires approximately 4 days on a single A100 GPU for one supercell structure.

Figure 7 presents the time-dependent current responses of Si supercells at different temperatures. With increasing temperature, the electronic dynamics driven by the strong laser field exhibit pronounced changes. The amplitude of the current decreases as the temperature increases, demonstrating that thermal structural fluctuations can effectively modulate the strong-field electron dynamics. We further perform a Fourier transform of the time-dependent current to obtain the HHG spectra at different temperatures. A Hanning window is applied to the current signal to improve the clarity of the harmonic spectra. As shown in Figure 8, low-order and high-order harmonics exhibit distinct temperature-dependent behaviors. For the third- and fifth-order harmonics, the harmonic yield gradually increases with increasing temperature. In contrast, the harmonic yield of higher-order harmonics (ninth order and above) decreases as the temperature increases. The temperature-dependent behavior of different harmonic orders exhibits a transition around the seventh harmonic order. This is in good qualitative agreement with the experimental measurements of HHG in Si for the crystalline and amorphized states [24, 60], further indicating the validity of our method.

Laser-induced currents of a 216-atom bulk Si supercell at various temperatures

Figure 7. Laser-induced currents of a 216-atom bulk Si supercell at various temperatures. The gray, red, blue, green, and purple lines correspond to the current responses of the bulk Si supercell at 100 K, 150 K, 200 K, 250 K, and 300 K, respectively.

Calculated high harmonic spectra for different temperatures

Figure 8. Calculated high harmonic spectra for different temperatures based on DeePTB.

To quantitatively analyze the temperature dependence of different harmonic orders, we calculate the variation of the harmonic yield for each order as a function of temperature. To further illustrate this trend, Figure 9 and Figure 10 show the normalized temperature dependence of the high-order and low-order harmonic yields, respectively, with respect to the results obtained at 100 K.

Temperature dependence of the harmonic yield for the 9th, 11th, 13th, and 15th harmonics

Figure 9. Temperature dependence of the harmonic yield for the 9th (a), 11th (b), 13th (c), and 15th (d) harmonics. The points are simulation results at 100 K, 150 K, 200 K, 250 K, and 300 K (Use 100 K as the baseline value of 1), respectively.

Temperature dependence of the harmonic yield for the 3rd, 5th, and 7th harmonics

Figure 10. Temperature dependence of the harmonic yield for the 3rd (a), 5th (b), 7th (c) harmonics. The points are simulation result at 100 K, 150 K, 200 K, 250 K, and 300 K (Use 100 K as the baseline value of 1), respectively.

We discuss the high order harmonics first. The results presented in Figure 9 are in good agreement with the previously reported experimental results [24]. As the temperature increases, thermal atomic displacements become more pronounced, leading to enhanced structural fluctuations in the Si supercell. The radial distribu-

@page 7/13 starts_mid=1 ends_mid=1 tion functions shown in Figure 11 reveal the temperature-dependent structural variations of bulk Si. With increasing temperature, the peaks of the radial distribution functions become broader and less pronounced, indicating a wider distribution of SiSi bond lengths caused by enhanced atomic vibrations. Meanwhile, the attenuation of higher-order peaks demonstrates a reduction of long-range structural correlations due to thermal disorder. These features directly evidence that elevated temperatures introduce stronger lattice fluctuations in the Si supercell. These thermal structural fluctuations modify the electronic Hamiltonian and consequently influence the strong-field-driven electron dynamics.

Radial distribution functions of bulk Si at different temperatures

Figure 11. The radial distribution functions of bulk Si at 100 K (grey line), 150 K (red line), 200 K (blue line), 250 K (green line), 300 K (purple line), respectively.

Because bulk Si has a small thermal expansion coefficient, lattice expansion can be safely neglected within the temperature range of 100–400 K under ambient pressure [61, 62, 63]. Therefore, the temperature dependence of harmonic yields mainly originates from thermally induced structural fluctuations rather than lattice expansion effects [35]. Because lattice expansion is negligible in this regime, the variation in HHG yields can be attributed mainly to temperature-induced atomic rearrangement. In finite-temperature solids, lattice vibrations influence electronic dynamics through electron-phonon coupling. In this work, instead of introducing an additional empirical electronphonon scattering model, we explicitly generate thermally perturbed crystal structures through finitetemperature molecular dynamics simulations and construct temperature-dependent electronic Hamiltonians based on these structures. The high-order harmonics are mainly generated from long-range electron transport. These thermally perturbed structural disorder induces scattering which changes the coherence of electron-hole pairs during their excursion in laser-induced dynamics [24]. Thus the higher-order harmonics become weak as temperature increases. In addition, we discuss the low-order harmonics. They are in the perturbative regime. The excursion of corresponding electron-hole pairs may be negligible. To gain deeper insight into the microscopic origin of the temperature-dependent enhancement of low-order harmonics, we further investigate the evolution of interatomic hopping amplitudes within the Si supercell. As illustrated in Figure 12, the hopping amplitudes exhibit an overall enhancement with increasing

Histograms of interatomic orbital hopping energy distributions at 100 K, 150 K, 200 K, 250 K and 300 K

Figure 12. Histograms of interatomic orbital hopping energy distributions in a 216 atoms bulk Si supercell at 100 K(a), 150 K(b), 200 K(c), 250 K(d) and 300 K(e), respectively. The atomic index labels all Si atoms in the supercell, and the displayed hopping energy is the maximum hopping energy associated with each atom as the central site.

temperature. This behavior is also closely related to the thermally induced structural fluctuations revealed by Figure 11. As discussed above, the peaks of the radial distribution functions become progressively broad with increasing temperature, while the characteristic SiSi bond lengths remain nearly unchanged. This observation suggests that thermal excitation primarily increases the fluctuations of local atomic configurations rather than producing a substantial change in the average crystal structure. Such fluctuations continuously modify the spatial separation and relative orientation of neighboring atomic orbitals, thereby altering their overlap and the corresponding electronic hopping amplitudes. Consequently, even in the absence of appreciable lattice expansion or contraction, enhanced thermal vibrations can produce noticeable modifications of the local electronic coupling within the supercell.

The microscopic origin of this behavior can be further understood from the nonlinear dependence of interatomic hopping on bond length [64]. Within the tight-binding framework, hopping amplitudes describe the coupling between neighboring atomic orbitals and are highly sensitive to their spatial overlap. In general, decreasing the interatomic distance increases the orbital overlap and tends to enhance the magnitude of the corresponding hopping integral, whereas increasing the distance produces the opposite effect. Importantly, these two contributions are not necessarily symmetric because the hopping amplitude depends nonlinearly on the interatomic separation [65, 66]. Because hopping magnitudes respond nonlinearly to the changes in bond length, the enhancement associated with shortened bonds can outweigh the reduction caused by elongated bonds. As temperature increases, the broadening of the SiSi bond-length distribution increases the occurrence of both compressed and elongated local configurations. The resulting asymmetric response of the hopping amplitudes may therefore lead to an increase in their statistical average, even when the characteristic bond length remains nearly constant. In addition, thermally induced variations in local bond orientations can modify the directional overlap and hybridization of Si orbitals, further contributing to the redistribution of electronic coupling strengths. These effects provide a plausible microscopic explanation for the observed temperature-dependent enhancement of interatomic hopping amplitudes.

Importantly, the hopping amplitudes determine the coupling between atomic orbitals and contribute to the electronic dispersion and the associated velocity and current operators. Consequently, temperature-induced modifications of hopping amplitudes can affect the response of electrons to an external laser field [67]. In the strong-field regime, changes in local electronic coupling can affect carrier dynamics, electronhole coherence, and elated quantities. Such modifications can alter the efficiency of harmonic emission. Therefore, the observed enhancement of low-order harmonic yields may be associated with the nonlinear modulation of interatomic electronic hopping by lattice thermal fluctuations. This interpretation suggests that the temperature dependence of low-order HHG is sensitive not only to changes in electronic energy levels but also to the microscopic redistribution of orbital coupling induced by thermal atomic motion.

In contrast, high-order harmonics rely on phase-coherent electron motion over extended spatial regions under strong-field excitation. Although thermal structural fluctuations enhance local orbital coupling, they also introduce temperature-dependent variations into the electronic Hamiltonian and perturb the phase coherence of long-range electron dynamics. This dephasing effect induced by the structural disorder reduces the efficiency of high-order harmonic, resulting in a temperature dependence opposite to that of low-order harmonics.

These results demonstrate that lattice thermal fluctuations at finite temperatures do not simply enhance or suppress HHG. Instead, they induce a harmonic-order-dependent redistribution of the HHG response by simultaneously modifying local electronic coupling and strong-field-driven coherent electron dynamics. The enhanced local coupling promotes low-order harmonic emission, whereas the modification of phase-coherent electron motion reduces the efficiency of high-order harmonics, leading to distinct temperature responses across different harmonic orders.

Summary

Lattice thermal fluctuations are ubiquitous in solids, yet their role in strong-field HHG remains difficult to investigate in realistic large-scale systems. Here we develop an approach that combines deep learning molecular dynamics, the DeePTB electronic Hamiltonian, and semiconductor Bloch equations. Calculations for h-BN show that the DeePTB Hamiltonian reproduces the reference HHG spectra with high accuracy. We then apply the method to 216-atom bulk Si supercells using an occupied-state propagation scheme to study HHG in thermally fluctuating lattices.

We find that lattice temperature modifies the harmonic response in an order-selective manner. The intensities of low-order harmonics increase with temperature, whereas those of high-order harmonics decrease, with a crossover near the seventh harmonic. By analyzing the temperature-dependent electronic Hamiltonians, we trace the enhancement of low-order harmonics to thermally induced changes in local orbital couplings and the resulting increase in hopping amplitudes. For the higher harmonic orders, thermal structural fluctuations instead suppress the coherence of the strong-field electron dynamics in a different regime, leading to a contrasting temperature dependence.

These results show that lattice thermal fluctuations can selectively reshape solid-state HHG rather than simply suppressing the harmonic response. The combination of machine-learning electronic Hamiltonians and SBE provides an efficient route for studying temperature-dependent strong-field electron dynamics in large and structurally disordered solid systems.

Acknowledgments

We thank Prof. S. Gräfe and Dr. M. Thümmler for providing the data obtained with Octopus and the periodic model. We thank Prof. T.T. Luu, Dr. J.C. Zhang, Prof. M.H. Chen, Prof. Q.Q. Gu, Dr. Z. Zhouyin, Prof. Z.Y. Fan, Prof. J.X. Chen, and Dr. Z.W. Ding for helpful discussions. This work is supported by the Hubei Provincial Natural Science Foundation of China (Grant No. 2024AFA029), the National Natural Science Foundation of China (Grant Nos. 12304383, and 12674429), and the Hubei Province Postdoctoral Pioneer Support Program (Grant No. 2025HBBSHXF011).

References

  1. [1]P. B. Corkum and F. Krausz, Attosecond science, Nature Physics 3, 381387 (2007).
  2. [2]F. Krausz and M. Ivanov, Attosecond physics, Reviews of Modern Physics 81, 163234 (2009).
  3. [3]T. Popmintchev, M.-C. Chen, P. Arpin, M. M. Murnane, and H. C. Kapteyn, The attosecond nonlinear optics of bright coherent x-ray generation, Nature Photonics 4, 822832 (2010).DOI
  4. [4]F. Silva, S. M. Teichmann, S. L. Cousin, M. Hemmer, and J. Biegert, Spatiotemporal isolation of attosecond soft x-ray pulses in the water window, Nature Communications 6, 6611 (2015).DOI
  5. [5]M. Chini, K. Zhao, and Z. Chang, The generation, characterization and applications of broadband isolated attosecond pulses, Nature Photonics 8, 178186 (2014).DOI
  6. [6]D. Shafir, H. Soifer, B. D. Bruner, M. Dagan, Y. Mairesse, S. Patchkovskii, M. Y. Ivanov, O. Smirnova, and N. Dudovich, Resolving the time when an electron exits a tunnelling barrier, Nature 485, 343346 (2012).DOI
  7. [7]J. Zhao and M. Lein, Determination of ionization and tunneling times in high-order harmonic generation, Physical Review Letters 111, 043901 (2013).pubmed.ncbi.nlm.nih.gov/23931368
  8. [8]S. Yue, Y. Li, S. Xue, H. Du, and M. M. Lein, Ionization and recombination times of the long trajectory in high-order harmonic generation, Physical Review A 106, 023117 (2022).DOI
  9. [9]S. Baker, J. S. Robinson, C. A. Haworth, H. Teng, R. A. Smith, C. C. Chiril, M. Lein, J. W. G. Tisch, and J. P. Marangos, Probing proton dynamics in molecules on an attosecond time scale, Science 312, 424427 (2006).pubmed.ncbi.nlm.nih.gov/16513942
  10. [10]M. Lein, Attosecond probing of vibrational dynamics with high-harmonic generation, Physical Review Letters 94, 053004 (2005).pubmed.ncbi.nlm.nih.gov/15783636
  11. [11]L. He, Q. Zhang, P. Lan, W. Cao, X. Zhu, C. Zhai, F. Wang, W. Shi, M. Li, X.-B. Bian, P. Lu, and A. D. Bandrauk, Monitoring ultrafast vibrational dynamics of isotopic molecules with frequency modulation of high-order harmonics, Nature Communications 9, 1108 (2018).DOI
  12. [12]J. Itatani, J. Levesque, D. Zeidler, H. Niikura, H. Pépin, J. C. Kieffer, P. B. Corkum, and D. M. Villeneuve, Tomographic imaging of molecular orbitals, Nature 432, 867871 (2004).pubmed.ncbi.nlm.nih.gov/15602553
  13. [13]G. Ndabashimiye, S. Ghimire, M. Wu, D. A. Browne, K. J. Schafer, M. B. Gaarde, and D. A. Reis, Solid-state harmonics beyond the atomic limit, Nature 534, 520523DOI
  14. [14]S. Ghimire and D. A. Reis, High-harmonic generation from solids, Nature Photonics 15, 1016 (2019).DOI
  15. [15]G. Vampa, C. P. McDonald, G. Orlando, D. . Klug, P. . Corkum, and T. Brabec, Theoretical analysis of high-harmonic generation in solids, Physical Review Letters 113, 073901 (2014).pubmed.ncbi.nlm.nih.gov/25170708
  16. [16]K. L. Ishikawa, Y. Shinohara, T. Sato, and T. Otobe, Strong-field electron dynamics in solids, in Ultrafast Elec-tronic and Structural Dynamics, edited by K. Y. Ueda (Springer Nature Singapore, Singapore, 2024) p. 119154.arxiv.org/abs/2003.14090
  17. [17]Y. S. You, Y. Yin, Y. Wu, A. Chew, X. Ren, F. Zhuang, S. Gholam-Mirzaei, M. Chini, Z. Chang, and S. Ghimire, High-harmonic generation in amorphous solids, Nature Communications 8, 724 (2017).DOI
  18. [18]N. Tancogne-Dejean, O. D. Mü cke, F. X. Kärtner, and A. Rubio, Impact of the electronic band structure in high-harmonic generation spectra of solids, Physical Review Letters 118, 087403 (2017).
  19. [19]J. Li, S. Fu, H. Wang, X. Zhang, B. Ding, B. Hu, and H. Du, Limitations of the single-active-electron approximation in quantum simulations of solid high-order harmonic generation, Phys. Rev. A 98, 043409 (2018).DOI
  20. [20]T.-Y. Du, Z. Guan, X.-X. Zhou, and X.-B. Bian, Enhanced high-order harmonic generation from periodic potentials in inhomogeneous laser fields, Physical Review A 94, 023419 (2016).DOI
  21. [21]P. B. Allen and V. Heine, Theory of the temperature dependence of electronic band structures, Journal of Physics C: Solid State Physics 9, 2305 (1976).DOI
  22. [22]T.-Y. Du and C. Ma, Temperature-induced dephasing in high-order harmonic generation from solids, Physical Review A 105, 053125 (2022).DOI
  23. [23]D. Freeman, A. Kheifets, S. Yamada, A. Yamada, and K. Yabana, High-order harmonic generation in semiconductors driven at near- and mid-infrared wavelengths, Physical Review B 106, 075202 (2022).DOI
  24. [24]S. Mokhtari, V. Jelic, D. N. Purschke, S. Gholam-Mirzaei, K. M. Kowalczyk, D. A. Reis, T. J. Hammond, D. M. Villeneuve, A. Staude, F. Légaré, and G. Vampa, Phonon-driven decoherence of high-harmonic generation in the solid-state (2026), arXiv:2604.07543.arxiv.org/abs/2604.07543
  25. [25]X. Liu, G. Zhang, J. Li, G. Shi, M. Zhou, B. Huang, Y. Tang, X. Song, and W. Yang, Deep learning for feynman’s path integral in strong-field time-dependent dy- namics, Phys. Rev. Lett. 124, 113202 (2020).
  26. [26]S. Kumar Giri, U. Saalmann, and J. M. Rost, Purifying electron spectra from noisy pulses with machine learning using synthetic hamilton matrices, Phys. Rev. Lett. 124, 113201 (2020).
  27. [27]L. Zhang, J. Han, H. Wang, R. Car, and W. E, Deep potential molecular dynamics: A scalable model with the accuracy of quantum mechanics, Phys. Rev. Lett. 120, 143001 (2018).DOI
  28. [28]J. Behler and M. Parrinello, Generalized neural-network representation of high-dimensional potential-energy surfaces, Phys. Rev. Lett. 98, 146401 (2007).pubmed.ncbi.nlm.nih.gov/17501293
  29. [29]Z. Fan, Y. Wang, P. Ying, K. Song, J. Wang, Y. Wang, Z. Zeng, K. Xu, E. Lindgren, J. M. Rahman, A. J. Gabourie, J. Liu, H. Dong, J. Wu, Y. Chen, Z. Zhong, J. Sun, P. Erhart, Y. Su, and T. Ala-Nissila, Gpumd: A package for constructing accurate machine-learned potentials and performing highly efficient atomistic simulations, The Journal of Chemical Physics 157, 114801 (2022).DOI
  30. [30]Z. Fan, Z. Zeng, C. Zhang, Y. Wang, K. Song, H. Dong, Y. Chen, and T. Ala-Nissila, Neuroevolution atomistic machine learning potentials: Combining high accuracy and low cost in atomistic simulations and application to heat transport, Physical Review B 104, 104309 (2021).DOI
  31. [31]K. Song, R. Zhao, J. Liu, Y. Wang, E. Lindgren, Y. Wang, S. Chen, K. Xu, T. Liang, P. Ying, N. Xu, Z. Zhao, J. Shi, J. Wang, S. Lyu, Z. Zeng, S. Liang, H. Dong, L. Sun, Y. Chen, Z. Zhang, W. Guo, P. Qian, J. Sun, P. Erhart, T. Ala-Nissila, Y. Su, and Z. Fan, General-purpose machine-learned potential for 16 elemental metals and their alloys, Nature Communications 15, 10208 (2024).DOI
  32. [32]Y. Zhang, M. Zhang, W. Yang, H. Yu, M. S. Si, S. Xue, and H. Du, Defects of the nearest-neighbor tight-binding model in the study of solid harmonics, Physical Review A 108, 043508 (2023).DOI
  33. [33]R. E. F. Silva, F. Martín, and M. Ivanov, High harmonic generation in crystals using maximally localized wannier functions, Physical Review B 100, 195201 (2019).DOI
  34. [34]Q. Gu, Z. Zhouyin, S. K. Pandey, P. Zhang, L. Zhang, and W. E, Deep learning tight-binding approach for large-scale electronic simulations at finite temperatures with ab initio accuracy, Nature Communications 15, 6772 (2024).DOI
  35. [35]S. Yang, S. K. Pandey, Z. Zhouyin, and Q. Gu, Deep neural network for phonon-assisted optical spectra of semiconductors at finite temperatures, npj Computational Materials 12, 268 (2026).DOI
  36. [36]J. Zou, Z. Zhouyin, S. K. Pandey, and Q. Gu, Review of machine learning tight-binding models: Route to accurate and scalable electronic simulations, Chinese Physics B 35, 017101 (2026).DOI
  37. [37]A. A. Mostofi, J. R. Yates, Y.-S. Lee, I. Souza, D. Vanderbilt, and N. Marzari, wannier90: A tool for obtaining maximally-localised wannier functions, Computer Physics Communications 178, 685 (2008).DOI
  38. [38]X. Wang, J. R. Yates, I. Souza, and D. Vanderbilt, Ab initio calculation of the anomalous hall conductivity by wannier interpolation, Physical Review B 74, 195118 (2006).DOI
  39. [39]Y. Li, X. Zhang, M. Liu, and L. Shen, A critical review of machine learning interatomic potentials and hamiltonian, Journal of Materials Informatics 5, 10.20517/jmi.2025.17 (2025).DOI
  40. [40]D. A. Papaconstantopoulos and M. J. Mehl, The slaterkoster tight-binding method: a computationally efficient and accurate approach, Journal of Physics: Condensed Matter 15, R413 (2003).DOI
  41. [41]M. Korbman, S. Yu Kruchinin, and V. S. Yakovlev, Quantum beats in the polarization response of a dielectric to intense few-cycle laser pulses, New Journal of Physics 15, 013006 (2013).DOI
  42. [42]J. B. Krieger and G. J. Iafrate, Time evolution of bloch electrons in a homogeneous electric field, Physical Review B 33, 54945500 (1986).pubmed.ncbi.nlm.nih.gov/9939055
  43. [43]W. V. Houston, Acceleration of electrons in a crystal lattice, Physical Review 57, 184186 (1940).DOI
  44. [44]J. B. Krieger and G. J. Iafrate, Quantum transport for bloch electrons in a spatially homogeneous electric field, Physical Review B 35, 96449658 (1987).pubmed.ncbi.nlm.nih.gov/9941391
  45. [45]R. Peierls, Zur theorie des diamagnetismus von leitungselektronen, Zeitschrift für Physik 80, 763791 (1933).DOI
  46. [46]M. Graf and P. Vogl, Electromagnetic fields and dielectric response in empirical tight-binding theory, Phys. Rev. B 51, 4940 (1995).pubmed.ncbi.nlm.nih.gov/9979365
  47. [47]M. Schüler, J. A. Marks, Y. Murakami, C. Jia, and T. P. Devereaux, Gauge invariance of light-matter interactions in first-principle tight-binding models, Phys. Rev. B 103, 155409 (2021).DOI
  48. [48]D. Kim, Y. Lee, A. Chacón, and D.-E. Kim, Effect of interlayer coupling and symmetry on high-order harmonic generation from monolayer and bilayer hexagonal boron nitride, Symmetry 14, 10.3390/sym141001084 (2022).arxiv.org/abs/2112.10117
  49. [49]X.-Y. Wu, H. Liang, X.-S. Kong, Q. Qong, and L.-Y. Peng, Multiscale numerical tool for studying nonlinear dynamics in solids induced by strong laser pulses, Phys. Rev. E 105, 055306 (2022).DOI
  50. [50]C. P. C. V. Bernardo, R. A. M. Lameirinhas, J. P. de Melo Cunha, and J. P. N. Torres, A revision of the semiconductor theory from history to applications, Discover Applied Sciences 6, 316 (2024).DOI
  51. [51]K. Rupp, C. Jungemann, S. M. Hong, M. Bina, T. Grasser, and A. Juengel, A review of recent advances in the spherical harmonics expansion method for semiconductor device simulation, JOURNAL OF COMPUTATIONAL ELECTRONICS 15, 939 (2016).
  52. [52]E. Bocquillon, V. Freulon, J.-M. Berroir, P. Degiovanni, B. Plaçais, A. Cavanna, Y. Jin, and G. Fève, Coherence and indistinguishability of single electrons emitted by independent sources, Science 339, 1054 (2013).
  53. [53]N. Tancogne-Dejean and A. Rubio, Atomic-like high-harmonic generation from two-dimensional materials, Science Advances 4, eaao5207 (2018).
  54. [54]W. Zhou, D. Zheng, Q. Liu, D. Lu, Y. Liu, P. Lin, Y. Huang, X. Peng, J. J. Bao, C. Cai, Z. Jin, J. Wu, H. Zhang, G. Jin, Y. Ji, Z. Shen, X. Li, L. Sun, Y. Cao, M. Sun, J. Liu, T. Chen, R. Liu, Y. Li, H. Han, X. Liang, T. Bao, Z. Deng, T. Liu, N. Chen, H. Ren, X. Zhang, Z. Liu, Y. Fu, M. Liu, Z. Li, T. Wen, Z. Tang, Y. Xu, W. Duan, X. Wang, Q. Gu, F.-Z. Dai, Q. Zheng, Y. Zhong, H. Xiang, X. Gong, J. Zhao, Y. Zhang, Q. Ou, H. Jiang, S. Liu, B. Xu, S. Xu, X. Ren, L. He, L. Zhang, and M. Chen, A package for the ai era, The Journal of Chemical Physics 163, 192501 (2025).arxiv.org/abs/2501.08697
  55. [55]T. Liang, K. Xu, E. Lindgren, Z. Chen, R. Zhao, J. Liu, E. Berger, B. Tang, B. Zhang, Y. Wang, K. Song, P. Ying, N. Xu, H. Dong, S. Chen, P. Erhart, Z. Fan, T. Ala-Nissila, and J. Xu, Nep89: universal neuroevolution potential for inorganic and organic materials across 89 elements, Nature Computational Science 6, 789801 (2026).DOI
  56. [56]P. Li, X. Liu, M. Chen, P. Lin, X. Ren, L. Lin, C. Yang, and L. He, Large-scale ab initio simulations based on systematically improvable atomic basis, Computational Materials Science 112, 503517 (2016).arxiv.org/abs/1503.00097
  57. [57]J. P. Perdew, K. Burke, and M. Ernzerhof, Generalized gradient approximation made simple, Phys. Rev. Lett. 77, 3865 (1996).pubmed.ncbi.nlm.nih.gov/10062328
  58. [58]M. Thümmler, A. Croy, U. Peschel, and S. Gräfe, Modeling high-order harmonic generation in quantum dots using a real-space tight-binding approach, The Journal of Chemical Physics 164, 094114 (2026).
  59. [59]L. Yuting, Y. Jiong, and X. Jinyang, Machine learning empowered electronic structure calculations: Progress, challenges, and prospects, Acta Physica Sinica 75 (2026).DOI
  60. [60]D. N. Purschke, D. Vick, A. Cárdenas, N. Haram, P. Bastani, S. Gholam-Mirzaei, S. Mokhtari, V. Jelic, J. Chen, J. Canlas, J. Tordiff, M. W. Rahman, A. Y. Naumov, D. M. Villeneuve, A. Staudte, M. Salomons, R. E. F. Silva, . Jiménez-Galán, and G. Vampa, Giant enhancement of attosecond tunnel ionization competes with disorder-driven decoherence in silicon (2025), arXiv:2511.14678.
  61. [61]C. A. Swenson, Recommended values for the thermal expansion of silicon from 0 to 1000 k, Journal of Physical and Chemical Reference Data 12, 179182 (1983).DOI
  62. [62]K. G. Lyon, G. L. Salinger, C. A. Swenson, and G. K. White, Linear thermal expansion measurements on silicon from 6 to 340 k, Journal of Applied Physics 48, 865868 (1977).DOI
  63. [63]Y. Okada and Y. Tokumaru, Precise determination of lattice parameter and thermal expansion coefficient of silicon between 300 and 1500 k, Journal of Applied Physics 56, 314320 (1984).
  64. [64]J. L. Mercer and M. Y. Chou, Tight-binding total-energy models for silicon and germanium, Phys. Rev. B 47, 9366 (1993).pubmed.ncbi.nlm.nih.gov/10005004
  65. [65]V. M. Pereira, A. H. Castro Neto, and N. M. R. Peres, Tight-binding approach to uniaxial strain in graphene, Phys. Rev. B 80, 045401 (2009).DOI
  66. [66]P. Moon and M. Koshino, Optical absorption in twisted bilayer graphene, Phys. Rev. B 87, 205404 (2013).DOI
  67. [67]O. Neufeld, J. Zhang, U. De Giovannini, H. Hübener, and A. Rubio, Probing phonon dynamics with multidimensional high harmonic carrier-envelope-phase spectroscopy, Proc Natl Acad Sci U S A 119, e2204219119 (2022).DOI

Paper details

vlnc:W_epqpbksb
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