Pith. sign in

REVIEW 5 major objections 6 minor 64 references

Deep Neural Network for Phonon-Assisted Optical Spectra in Semiconductors

T0 review · 5 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Pairing two deep-learning models reproduces measured silicon and gallium arsenide light absorption over five orders of magnitude.

desk verdict Promising ML workflow for phonon-assisted optical spectra, but the finite-temperature renormalization claim is weakened by classical MD sampling that omits zero-point motion. read the letter →

arxiv 2502.00798 v2 pith:EGCO6IZK submitted 2025-02-02 cond-mat.mtrl-sci cs.LG

Qiangqiang Gu , Shishir Kumar Pandey , Zhanghao Zhouyin This is my paper · ORCID
classification cond-mat.mtrl-scics.LG
keywords phonon-assistedopticalabsorptiondeeplearningtight-bindingpotentialmoleculardynamicsWilliams-Laxtheorytemperature-dependentbandgapsilicongalliumarsenideelectron-phononcoupling
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that pairing a deep-learning interatomic potential (DeePMD) with a deep-learning tight-binding Hamiltonian model (DeePTB) makes it practical to compute phonon-assisted optical absorption spectra at finite temperatures using expensive functionals like HSE and SCAN. The key move is to replace the usual phonon normal-mode sampling with molecular-dynamics snapshots and to evaluate the single-particle optical response for each snapshot with a learned tight-binding Hamiltonian, trained on small-cell DFT eigenvalues but applied to supercells of up to 4096 atoms. On silicon and gallium arsenide, the resulting absorption coefficients track experiment from roughly $10^{1}$ to $10^{6}$ $cm^{-1}$, recovering the indirect absorption tail below the direct gap that is invisible in clamped-atom calculations. If the claim holds, temperature-dependent optical properties of semiconductors cease to require costly large-supercell hybrid-functional calculations, opening a route to high-throughput screening of optoelectronic materials.

What carries the argument

Two deep-learning components carry the argument: DeePMD, a deep neural-network interatomic potential that generates finite-temperature atomic configurations by molecular dynamics, and DeePTB, a deep-learning tight-binding model that maps each atomic configuration X to a tight-binding Hamiltonian H(R;X), trained on DFT (HSE or SCAN) eigenvalues of small cells. The optical response is built from Eq. (1) for each snapshot, the Williams-Lax thermal average approximated as a simple average over MD snapshots in Eq. (3), and the momentum matrix element evaluated from the k-derivative of the learned Hamiltonian in Eq. (4). The transferability of DeePTB from small training cells to 4096-atom displaced supercells is what makes the large-supercell average computationally affordable.

What would settle it

Run HSE (for Si) or SCAN (for GaAs) on one configuration from each sampled temperature from the 8x8x8 supercell, compare the resulting band-edge eigenvalues and the absorption spectrum with the DeePTB prediction, and check whether the difference is below the experimental tolerance claimed (roughly the bandgap renormalization of tens of meV). If the learned Hamiltonian drifts with displacement or cell size, the agreement with experiment and the claimed supercell convergence would not survive direct DFT validation.

Watch

Extended reading notes

Core claim

The central discovery is that the Williams-Lax ensemble average of the dielectric function, which formally requires sampling thermally populated nuclear configurations, can be approximated by a small number of classical molecular-dynamics snapshots provided the supercell is large enough, and that a deep-learning tight-binding model can deliver the Hamiltonian and momentum operator for each snapshot at ab initio quality. Concretely, for Si the authors train DeePTB on HSE eigenvalues of small cells, run DeePMD dynamics on an 8x8x8 cell, and reproduce the measured 300 K absorption coefficient, including the 1.1-3.3 eV indirect region, over five orders of magnitude; supercell convergence is reached by L=4 with deviations between snapshots becoming negligible. For GaAs, a SCAN-trained DeePTB model with a disclosed 1.0 eV scissor shift captures the phonon-assisted absorption below the direct gap and the redshift with respect to clamped-atom spectra. The temperature dependence of the Si indirect gap from 100 to 400 K follows the experimental narrowing trend.

Load-bearing premise

The whole calculation assumes that a tight-binding model trained on small-cell DFT eigenvalues keeps its accuracy when applied to 4096-atom supercells whose atoms are displaced by thermal motion at 100-400 K; the paper does not directly compare DeePTB predictions with DFT on those large, displaced cells.

Editorial extensions

If this is right

  • For Si and GaAs, absorption coefficients computed from MD snapshots agree with experiment over five orders of magnitude, including the indirect absorption region.
  • Supercell convergence is reached at L=4 for Si, and for L=4 and L=8 a single MD snapshot already represents the ensemble average, enabling one-shot spectral calculations without specially designed displacement patterns.
  • The Si indirect band gap from 100 to 400 K reproduces the experimentally observed temperature-induced narrowing.
  • Because DeePTB is trained on HSE/SCAN eigenvalues, large-supercell spectra can be obtained with these expensive functionals at a fraction of the cost, which was previously inaccessible.
  • The approach is generic and can be extended to alloys, nanostructures, and other phonon-assisted phenomena.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Going beyond the paper, a natural testable extension is to compare DeePTB-predicted eigenvalues and spectra against direct HSE/SCAN calculations on the same large supercell; the paper does not show this validation, so the supercell-size convergence could in principle mask a systematic model error.
  • Going beyond the paper, at 100 K classical MD sampling may underestimate quantum nuclear zero-point and tunneling effects; using path-integral or stochastic self-consistent harmonic sampling could reveal whether the low-temperature agreement is robust.
  • Going beyond the paper, the same workflow could be pointed at temperature-dependent mobilities or thermoelectric coefficients, since the learned Hamiltonian and its k-derivative supply both band velocities and electron-phonon coupling information.
  • Going beyond the paper, the scissor correction for GaAs and Gaussian broadenings are empirical knobs; quantifying how much of the five-order-of-magnitude agreement depends on them would clarify the predictive content of the method.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper presents a computational workflow that combines deep-learning interatomic potentials (DeePMD) and deep-learning tight-binding Hamiltonians (DeePTB) to compute finite-temperature phonon-assisted optical absorption spectra within the Williams–Lax ensemble-averaging framework. The method is demonstrated on silicon (indirect gap) and gallium arsenide (direct gap), using HSE for Si and SCAN with a 1.0 eV scissor correction for GaAs. The authors report absorption spectra over five orders of magnitude, supercell-size convergence up to 4096 atoms, single-snapshot statistical convergence, and temperature-dependent indirect band gaps for Si from 100 K to 400 K, with the central claim that the workflow enables large-scale simulations of temperature-dependent optical properties with advanced exchange-correlation functionals.

Significance. If the central claims hold, the workflow is a valuable contribution: it offers a practical route to phonon-assisted optical spectra with hybrid functionals on supercells far beyond direct DFT, demonstrates that single MD snapshots can represent the ensemble average for large cells, and makes the DeePTB code publicly available. The demonstrated agreement with experiment at 300 K for Si and GaAs, including the indirect absorption tail, is encouraging. However, the significance is moderated by two unvalidated assumptions: classical MD sampling in place of the quantum nuclear average, and transferability of DeePTB from small training cells to large displaced supercells. These assumptions bear directly on the claimed accuracy of temperature-dependent bandgap renormalization.

major comments (5)
  1. [II.A, Eqs. (2)–(3), and Fig. 4(b)] The manuscript replaces the quantum Williams–Lax average of Eq. (2) with a classical MD ensemble average, Eq. (3), without discussing the validity of this replacement at the temperatures studied. For Si at 100 K, optical phonons have ħω ~ 60 meV, far exceeding kBT ~ 8.6 meV, so classical sampling severely underestimates the zero-point contribution to the mean-square displacements and hence to phonon-induced bandgap renormalization. The consistent offset between the calculated and experimental temperature-dependent gaps in Fig. 4(b) is attributed to 'inherent limitations of the DFT calculations used for training,' but the DeePTB model is trained on HSE eigenvalues that already place the Si gap near the experimental value, so this attribution is questionable. A more natural explanation is the missing zero-point renormalization. As a result, the abstract's claim that the method 'accurately captures phonon-induced bandgap renormalization' is supported only for the thermal slope, not for the absolute magnitude. To strengthen this point, the authors should either include quantum zero-point effects (e.g., stochastic sampling of quantum phonon normal modes) as a benchmark, or explicitly restrict the claim to the thermal slope and quantify the zero-point offset.
  2. [II.A and Fig. 1] The DeePTB model is trained on eigenvalues from small-cell DFT calculations and then applied to 4096-atom supercells and MD-displaced configurations, yet no direct validation of DeePTB eigenvalues, band gaps, or spectra against DFT on a large supercell or on displaced configurations is provided. The statement in Section II.A that 'the trained DeePTB model enables efficient and accurate prediction of the TB Hamiltonian H(R; X) for every sampled configuration X in MD simulations' is therefore unsupported. If the model error grows with system size or displacement amplitude, the observed supercell convergence and agreement with experiment could be coincidental. The authors should present error metrics (e.g., eigenvalue or bandgap errors) as a function of supercell size and displacement amplitude, or at least compare DeePTB predictions with DFT on a moderately sized supercell with finite-temperature displacements.
  3. [Abstract and Section II.B (GaAs)] The abstract states that the method is 'demonstrated on silicon and gallium arsenide across temperature 100–400 K,' but temperature-dependent results are presented only for silicon (Fig. 4); GaAs is shown only at a single temperature of 300 K (Fig. 5). This is an overstatement of the demonstrated scope. The authors should either add temperature-dependent GaAs spectra or revise the abstract to limit the temperature-range claim to silicon, noting that GaAs is demonstrated at one temperature.
  4. [Section II.B (GaAs paragraph)] For GaAs, the DeePTB model is 'adjusted with a 1.0 eV scissor correction to align with experimental band gaps,' an empirical parameter fitted to experiment. This is disclosed, but it qualifies the otherwise parameter-free narrative of ab initio fidelity in the abstract and introduction. Since the central demonstration for GaAs relies on this empirical scissor, the authors should quantify how sensitive the phonon-assisted tail and the thermal slope are to the scissor value, and should make clear that the GaAs demonstration is not a fully parameter-free prediction from the SCAN functional alone.
  5. [Section II.B and supplementary materials] The paper repeatedly defers data preparation, model training, and MD simulation details to 'the supplementary materials (SM),' but no SM is available with the arXiv version. For a methods paper, essential reproducibility information such as training set sizes, descriptor parameters, network architectures, DeePTB fitting procedure, MD thermostat settings, and equilibration protocols is missing. Without these details, the results cannot be reproduced or independently assessed. The authors should include at least a summary of these parameters in the main text or an appendix and make the training data and scripts available.
minor comments (6)
  1. [Title/abstract] The article title appears inconsistently: the running header reads 'Deep Neural Network for Phonon-Assisted Optical Spectra of Semiconductors at finite temperatures' while the abstract header reads 'Deep Neural Network for Phonon-Assisted Optical Spectra in Semiconductors.' Please unify the title.
  2. [Fig. 3 caption and Section II.B] In the convergence discussion, the statement that 'a single configuration can statistically represent the ensemble average' is demonstrated only for Si at 300 K, with five snapshots. Please note in the text that this conclusion is empirical and may depend on temperature, material, and the choice of Gaussian broadening.
  3. [Eq. (1) and following text] The notation for the plasma frequency ωp = 4πNee²/me is nonstandard (the factor 4π is usually absorbed in Gaussian units, and the expression as written has units of frequency²); please clarify the unit system or correct the prefactor.
  4. [Reference [48]] In Section I, the text cites 'M. Zacharias et al. [48]' for the one-shot approach, but reference [48] is a two-author paper (Zacharias and Giustino). Please correct the citation to 'M. Zacharias and F. Giustino'.
  5. [Fig. 4(a)] The Tauc plots show linear fits over 'an energy range spanning nearly 1 eV'; for clarity, please specify the exact fitting range used to extract the indirect gap in Fig. 4(b).
  6. [Code availability] The availability statement lists only the DeePTB package; for full reproducibility, please also link the DeePMD package, the trained model files, and any scripts used to generate the spectra.

Circularity Check

1 steps flagged · score 6.0 of 10

GaAs direct-gap onset is forced by a scissor shift fitted to experiment; the Si phonon-assisted spectra and temperature dependence remain independent, so circularity is partial.

  1. fitted input called prediction [Section II.B, GaAs paragraph introducing Fig. 5 and the discussion following Fig. 5]
    "The DeePTB model, trained on SCAN functional eigenvalues of conventional unit cell configurations, was adjusted with a 1.0 eV scissor correction to align with experimental band gaps. ... Our results are in good agreement with experimental data over a broad energy range, accurately capturing the absorption onset at the direct gap and subtle high-energy features."

    The 1.0 eV scissor shift is fitted to the experimental GaAs band gap, so the calculated direct absorption onset is placed at the experimental gap by construction. The paper then presents the agreement of the onset with experiment as a successful capture of the direct absorption process. This is a fitted input presented as a prediction for the gap position. The phonon-assisted tail below the direct gap and the redshift from MD sampling are not forced by the scissor, so the circularity is limited to the direct-onset part of the claimed agreement.

full rationale

The core silicon derivation is not circular: DeePMD/DeePTB models are trained on DFT energies, forces, and HSE eigenvalues, while the absorption spectra, indirect tail, and temperature dependence emerge from MD ensemble averaging via Eqs. (1), (3), and (4) and are compared with experiment. The classical replacement of the quantum Williams-Lax expectation in Eq. (3) is an approximation and a correctness risk, not a definitional identity, and the absence of zero-point sampling at 100 K is a physical limitation rather than a circular reduction. The reliance on the authors' DeePTB package is supported by a separately published, code-available method and by the present independent experimental comparisons, so it is not counted as load-bearing self-citation. The single constructional circularity is the disclosed GaAs scissor correction: fitting the static gap to experiment and then claiming accurate prediction of the direct absorption onset verifies the fit. Because the Si results and the sub-gap GaAs phonon-assisted tail retain independent predictive content, the circularity is partial, giving a score of 6 rather than higher.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The method's output is a convolution of trained ML models and known many-body approximations. No new physical entities are introduced. The main uncharged premises are the classical treatment of nuclei and the transferability of the DeePTB model from small training cells to 4096-atom MD snapshots; both are asserted rather than validated against a direct ab initio reference.

free parameters (2)
  • GaAs scissor correction = 1.0 eV
    Applied to the SCAN-trained DeePTB eigenvalues to align the direct gap with experiment; this shift controls the absorption onset and therefore contributes to the claimed agreement.
  • Gaussian broadening width = 30 meV (Si), 50 meV (GaAs)
    Ad hoc broadening chosen for the spectra; different widths change line shapes and near-edge absorption magnitude, and no systematic sensitivity analysis is given.
assumptions (6)
  • domain assumption Williams-Lax quantum thermal average (Eq. 2) is the correct theory for temperature-dependent epsilon2 in the Born-Oppenheimer approximation.
    Invoked in Section II.A as the foundation for Eq. 2; standard many-body result, but its validity for these systems is assumed.
  • ad hoc to paper Classical MD sampling with DeePMD at 100-400 K reproduces the quantum nuclear average in Eq. 2.
    Eq. 3 replaces the quantum trace with time averages over classical trajectories; no zero-point or quantum-nuclear error analysis is provided, especially at 100 K.
  • ad hoc to paper DeePTB Hamiltonian trained on small-cell DFT eigenvalues transfers accurately to 4096-atom supercells and MD-displaced configurations.
    Core of the workflow in Fig. 1; no direct comparison of DeePTB eigenvalues or spectra with DFT on a large supercell is shown.
  • domain assumption Single-particle, dipole-approximation epsilon2 (Eq. 1) without excitonic effects is accurate enough for the reported spectral range.
    Used throughout; the authors acknowledge excitonic effects are omitted and cause underestimation of the E1 peak and GaAs absorption.
  • domain assumption DeePMD potential reproduces DFT energies and forces sufficiently for unbiased MD sampling.
    Standard assumption for ML interatomic potentials; the paper does not report validation metrics in the main text.
  • domain assumption Tauc relation [omega^2 epsilon2]^(1/2) proportional to (hbar omega - Delta_T) holds near the absorption onset for extracting indirect gaps.
    Used for Fig. 4(a) to extract the temperature-dependent gap; requires parabolic bands and constant matrix elements.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Neural Network for Phonon-Assisted Optical Spectra in Semiconductors." pith.science (2026). https://pith.science/paper/EGCO6IZK

@misc{pith2026250200798,
  author       = {Pith},
  title        = {Pith review of: Deep Neural Network for Phonon-Assisted Optical Spectra in Semiconductors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EGCO6IZK}},
  note         = {Machine review of arXiv:2502.00798}
}
read the original abstract

Ab initio based accurate simulation of phonon-assisted optical spectra of semiconductors at finite temperatures remains a formidable challenge, as it requires large supercells for phonon sampling and computationally expensive high-accuracy exchange-correlation (XC) functionals. In this work, we present an efficient approach that combines deep learning tight-binding and potential models to address this challenge with ab initio fidelity. By leveraging molecular dynamics for atomic configuration sampling and deep learning-enabled rapid Hamiltonian evaluation, our approach enables large-scale simulations of temperature-dependent optical properties using advanced XC functionals (HSE, SCAN). Demonstrated on silicon and gallium arsenide across temperature 100-400 K, the method accurately captures phonon-induced bandgap renormalization and indirect/direct absorption processes which are in excellent agreement with experimental findings over five orders of magnitude. This work establishes a pathway for high-throughput investigation of electron-phonon coupled phenomena in complex materials, overcoming traditional computational limitations arising from large supercell used with computationally expensive XC-functionals.

Figures

Figures reproduced from arXiv: 2502.00798 by the authors.

Figure 1
Figure 1. FIG. 1. Workflow for calculating phonon-assisted optical [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Absorption coefficient of silicon (Si) at 300 K. Blue [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. Convergence of supercell size and number of snapshots (a) The optical absorption spectra of Si were calculated [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: FIG. 4. Temperature-renormalized indirect band gaps of Si. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5. Absorption coefficient of GaAs at 300 K. Blue dashed [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

64 extracted references · 55 canonical work pages

  1. [1]

    T. B. Blank, S. D. Brown, A. W. Calhoun, and D. J. Doren, Neural network models of potential energy sur- faces, The Journal of Chemical Physics 103, 4129 (1995)

    work page 1995

  2. [2]

    Behler and M

    J. Behler and M. Parrinello, Generalized Neural-Network Representation of High-Dimensional Potential-Energy Surfaces, Phys. Rev. Lett. 98, 146401 (2007)

    2007

  3. [3]

    A. P. Bart´ ok, M. C. Payne, R. Kondor, and G. Cs´ anyi, Gaussian Approximation Potentials: The Accuracy of Quantum Mechanics, without the Electrons, Phys. Rev. Lett. 104, 136403 (2010)

    2010

  4. [4]

    Thompson, L

    A. Thompson, L. Swiler, C. Trott, S. Foiles, and G. Tucker, Spectral neighbor analysis method for auto- mated generation of quantum-accurate interatomic po- tentials, Journal of Computational Physics 285, 316 (2015)

    work page 2015

  5. [5]

    A. V. Shapeev, Moment tensor potentials: A class of sys- tematically improvable interatomic potentials, Multiscale Modeling & Simulation 14, 1153 (2016)

    2016

  6. [6]

    Zhang, J

    L. Zhang, J. Han, H. Wang, R. Car, and W. E, Deep potential molecular dynamics: A scalable model with the accuracy of quantum mechanics, Physical Review Letters 120, 143001 (2018)

    2018

  7. [7]

    K. T. Sch¨ utt, H. E. Sauceda, P.-J. Kindermans, A. Tkatchenko, and K.-R. M¨ uller, SchNet – A deep learn- ing architecture for molecules and materials, The Journal of Chemical Physics 148, 241722 (2018)

    work page 2018

  8. [8]

    Chmiela, H

    S. Chmiela, H. E. Sauceda, K.-R. M¨ uller, and A. Tkatchenko, Towards exact molecular dynamics sim- ulations with machine-learned force fields, Nature Com- munications 9, 3887 (2018)

    work page 2018

Show all 64 references
  1. [9]

    Drautz, Atomic cluster expansion for accurate and transferable interatomic potentials, Physical Review B 99, 014104 (2019)

    R. Drautz, Atomic cluster expansion for accurate and transferable interatomic potentials, Physical Review B 99, 014104 (2019)

    2019

  2. [10]

    O. T. Unke and M. Meuwly, PhysNet: A Neural Net- work for Predicting Energies, Forces, Dipole Moments, and Partial Charges, Journal of Chemical Theory and Computation 15, 3678 (2019)

    2019

  3. [11]

    J. P. Mailoa, M. Kornbluth, S. Batzner, G. Samsonidze, S. T. Lam, J. Vandermause, C. Ablitt, N. Molinari, and B. Kozinsky, A fast neural network approach for di- rect covariant forces prediction in complex multi-element extended systems, Nature Machine Intelligence 1, 471 (2019)

    2019

  4. [12]

    Gasteiger, J

    J. Gasteiger, J. Groß, and S. G¨ unnemann, Directional message passing for molecular graphs, in International Conference on Learning Representations, (ICLR) (2020)

    2020

  5. [13]

    C. W. Park, M. Kornbluth, J. Vandermause, C. Wolver- ton, B. Kozinsky, and J. P. Mailoa, Accurate and scalable graph neural network force field and molecular dynamics with direct force architecture, npj Computational Mate- rials 7, 1 (2021)

    2021

  6. [14]

    Y. Xie, J. Vandermause, L. Sun, A. Cepellotti, and B. Kozinsky, Bayesian force fields from active learn- ing for simulation of inter-dimensional transformation of stanene, npj Computational Materials 7, 1 (2021)

    2021

  7. [15]

    Batzner, A

    S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky, E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials, Nature Communications 13, 2453 (2022)

    2022

  8. [16]

    Musaelian, S

    A. Musaelian, S. Batzner, A. Johansson, L. Sun, C. J. Owen, M. Kornbluth, and B. Kozinsky, Learning local equivariant representations for large-scale atomistic dy- namics, Nature Communications 14, 579 (2023)

    2023

  9. [17]

    Q. Gu, L. Zhang, and J. Feng, Neural network repre- sentation of electronic structure from ab initio molecular dynamics, Science Bulletin 67, 29 (2022)

    2022

  10. [18]

    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 Communications15, 6772 (2024)

    2024

  11. [19]

    Z. Wang, S. Ye, H. Wang, J. He, Q. Huang, and S. Chang, Machine learning method for tight-binding Hamiltonian parameterization from ab-initio band struc- ture, npj Comput. Mater. 7, 11 (2021)

    2021

  12. [20]

    G. Fan, A. McSloy, B. Aradi, C.-Y. Yam, and T. Frauen- heim, Obtaining Electronic Properties of Molecules through Combining Density Functional Tight Binding with Machine Learning, The Journal of Physical Chem- istry Letters 13, 10132 (2022)

    2022

  13. [21]

    W. Sun, G. Fan, T. van der Heide, A. McSloy, T. Frauen- heim, and B. Aradi, Machine Learning Enhanced DFTB Method for Periodic Systems: Learning from Electronic Density of States, Journal of Chemical Theory and Com- putation 19, 3877 (2023)

    2023

  14. [22]

    McSloy, G

    A. McSloy, G. Fan, W. Sun, C. H¨ olzer, M. Friede, S. Ehlert, N.-E. Sch¨ utte, S. Grimme, T. Frauenheim, and B. Aradi, TBMaLT, a flexible toolkit for combining tight- binding and machine learning, The Journal of Chemical Physics 158, 034801 (2023)

    2023

  15. [23]

    Hegde and R

    G. Hegde and R. C. Bowen, Machine-learned approxima- tions to Density Functional Theory Hamiltonians, Scien- tific Reports 7, 42669 (2017)

    2017

  16. [24]

    K. T. Sch¨ utt, M. Gastegger, A. Tkatchenko, K.-R. M¨ uller, and R. J. Maurer, Unifying machine learning and quantum chemistry with a deep neural network for molec- ular wavefunctions, Nat. Commun. 10, 5024 (2019)

    2019

  17. [25]

    Nigam, M

    J. Nigam, M. J. Willatt, and M. Ceriotti, Equivariant representations for molecular Hamiltonians and N-center atomic-scale properties, J. Chem. Phys. 156, 014115 (2022)

    2022

  18. [26]

    Zhang, B

    L. Zhang, B. Onat, G. Dusson, A. McSloy, G. Anand, R. J. Maurer, C. Ortner, and J. R. Kermode, Equivari- ant analytical mapping of first principles Hamiltonians to accurate and transferable materials models, npj Comput. Mater. 8, 1 (2022). 7

    2022

  19. [27]

    H. Li, Z. Wang, N. Zou, M. Ye, R. Xu, X. Gong, W. Duan, and Y. Xu, Deep-learning density functional theory Hamiltonian for efficient ab initio electronic- structure calculation, Nat. Comput. Sci. 2, 367 (2022)

    2022

  20. [28]

    X. Gong, H. Li, N. Zou, R. Xu, W. Duan, and Y. Xu, General framework for E(3)-equivariant neural network representation of density functional theory hamiltonian, Nature Communications 14, 2848 (2023)

    2023

  21. [29]

    Zhong, H

    Y. Zhong, H. Yu, M. Su, X. Gong, and H. Xiang, Trans- ferable equivariant graph neural networks for the Hamil- tonians of molecules and solids, npj Computational Ma- terials 9, 1 (2023)

    2023

  22. [30]

    Zhouyin, Z

    Z. Zhouyin, Z. Gan, S. K. Pandey, L. Zhang, and Q. Gu, Learning local equivariant representations for quantum operators (2024), arXiv:2407.06053

    2024 arXiv

  23. [31]

    X. Wang, H. Wang, Q. Luo, and J. Yang, Structural and electrocatalytic properties of copper clusters: A study via deep learning and first principles, The Journal of Chem- ical Physics 157, 074304 (2022)

    2022

  24. [32]

    T. Hu, L. Xu, F. Dai, G. Zhou, F. Fu, X. Wang, L. Li, X. Ai, and S. Xu, Impact of amorphous LiF coating layers on cathode-electrolyte interfaces in solid-state batteries, Advanced Functional Materials 2024, 2402993

    2024

  25. [33]

    K. L. Liu, R. L. Xiao, Y. Ruan, and B. Wei, Ac- tive learning prediction and experimental confirmation of atomic structure and thermophysical properties for liq- uid Hf76 W24 refractory alloy, Phys. Rev. E 108, 055310 (2023)

    2023

  26. [34]

    Majewski, A

    M. Majewski, A. P´ erez, P. Th¨ olke, S. Doerr, N. E. Char- ron, T. Giorgino, B. E. Husic, C. Clementi, F. No´ e, and G. De Fabritiis, Machine learning coarse-grained po- tentials of protein thermodynamics, Nature Communica- tions 14, 5739 (2023)

    2023

  27. [35]

    G. M. Sommers, M. F. C. Andrade, L. Zhang, H. Wang, and R. Car, Raman spectrum and polarizability of liq- uid water from deep neural networks, Physical Chemistry Chemical Physics 22, 10592 (2020)

    2020

  28. [36]

    N. Xu, P. Rosander, C. Sch¨ afer, E. Lindgren, N. ¨Osterbacka, M. Fang, W. Chen, Y. He, Z. Fan, and P. Erhart, Tensorial Properties via the Neuroevolution Potential Framework: Fast Simulation of Infrared and Raman Spectra, Journal of Chemical Theory and Com- putation 20, 3273 (2024)

    2024

  29. [37]

    Zhang, M

    L. Zhang, M. Chen, X. Wu, H. Wang, W. E, and R. Car, Deep neural network for the dielectric response of insu- lators, Physical Review B 102, 041121 (2020)

    2020

  30. [38]

    Gastegger, J

    M. Gastegger, J. Behler, and P. Marquetand, Machine learning molecular dynamics for the simulation of in- frared spectra, Chemical Science 8, 6924 (2017)

    2017

  31. [39]

    Garc ´ ıa-Risue˜ no, E

    P. Garc ´ ıa-Risue˜ no, E. Armengol, J. M. Garc ´ ıa-Cerda˜ na, `Angel Garc ´ ıa-Lastra, and D. Carrasco-Busturia, Elec- tron–vibrational renormalization in fullerenes through ab initio and machine learning methods, Phys. Chem. Chem. Phys. 26, 20310 (2024)

    2024

  32. [40]

    Haldar, Q

    A. Haldar, Q. Clark, M. Zacharias, F. Giustino, and S. Sharifzadeh, Machine learning electron-phonon in- teractions in two-dimensional semiconducting materi- als: The case of zero-point renormalization, Phys. Rev. Mater. 8, L101001 (2024)

    2024

  33. [41]

    L. H. Hall, J. Bardeen, and F. J. Blatt, Infrared absorp- tion spectrum of germanium, Phys. Rev. 95, 559 (1954)

    1954

  34. [42]

    Noffsinger, E

    J. Noffsinger, E. Kioupakis, C. G. Van de Walle, S. G. Louie, and M. L. Cohen, Phonon-Assisted Optical Ab- sorption in Silicon from First Principles, Physical Review Letters 108, 167402 (2012)

    2012

  35. [43]

    Marini, Ab initio finite-temperature excitons, Physical Review Letters 101, 106405 (2008)

    A. Marini, Ab initio finite-temperature excitons, Physical Review Letters 101, 106405 (2008)

    2008

  36. [44]

    Giustino, S

    F. Giustino, S. G. Louie, and M. L. Cohen, Electron- phonon renormalization of the direct band gap of dia- mond, Physical Review Letters 105, 265501 (2010)

    2010

  37. [45]

    Giustino, Electron-phonon interactions from first prin- ciples, Reviews of Modern Physics 89, 015003 (2017)

    F. Giustino, Electron-phonon interactions from first prin- ciples, Reviews of Modern Physics 89, 015003 (2017)

    2017

  38. [46]

    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)

    1976

  39. [47]

    Zacharias, C

    M. Zacharias, C. E. Patrick, and F. Giustino, Stochastic Approach to Phonon-Assisted Optical Absorption, Phys- ical Review Letters 115, 177401 (2015)

    2015

  40. [48]

    Zacharias and F

    M. Zacharias and F. Giustino, One-shot calculation of temperature-dependent optical spectra and phonon- induced band-gap renormalization, Physical Review B 94, 075125 (2016)

    2016

  41. [49]

    F. E. Williams, Theoretical low temperature spectra of the thallium activated potassium chloride phosphor, Phys. Rev. 82, 281 (1951)

    1951

  42. [50]

    Lax, The franck-condon principle and its application to crystals, The Journal of Chemical Physics 20, 1752 (1952)

    M. Lax, The franck-condon principle and its application to crystals, The Journal of Chemical Physics 20, 1752 (1952)

    1952

  43. [51]

    J. Heyd, G. E. Scuseria, and M. Ernzerhof, Hybrid func- tionals based on a screened coulomb potential, The Jour- nal of Chemical Physics 118, 8207 (2003)

    2003

  44. [52]

    J. Sun, A. Ruzsinszky, and J. P. Perdew, Strongly con- strained and appropriately normed semilocal density functional, Phys. Rev. Lett. 115, 036402 (2015)

    2015

  45. [53]

    Y. Kang, H. Peelaers, K. Krishnaswamy, and C. G. Van de Walle, First-principles study of direct and indirect op- tical absorption in BaSnO3, Applied Physics Letters112, 062106 (2018)

    2018

  46. [54]

    Errea, M

    I. Errea, M. Calandra, and F. Mauri, First-principles theory of anharmonicity and the inverse isotope effect in superconducting palladium-hydride compounds, Phys. Rev. Lett. 111, 177002 (2013)

    2013

  47. [55]

    Errea, M

    I. Errea, M. Calandra, and F. Mauri, Anharmonic free energies and phonon dispersions from the stochastic self- consistent harmonic approximation: Application to plat- inum and palladium hydrides, Phys. Rev. B 89, 064302 (2014)

    2014

  48. [56]

    Bianco, I

    R. Bianco, I. Errea, L. Paulatto, M. Calandra, and F. Mauri, Second-order structural phase transitions, free energy curvature, and temperature-dependent anhar- monic phonons in the self-consistent harmonic approx- imation: Theory and stochastic implementation, Phys. Rev. B 96,...

    2017

  49. [57]

    Monacelli, R

    L. Monacelli, R. Bianco, M. Cherubini, M. Calandra, I. Errea, and F. Mauri, The stochastic self-consistent har- monic approximation: calculating vibrational properties of materials with full quantum and anharmonic effects, Journal of Physics: Condensed Matter 33, 363001 (2021)

    2021

  50. [58]

    M. A. Green and M. J. Keevers, Optical properties of intrinsic silicon at 300 K, Progress in Photovoltaics: Re- search and Applications 3, 189 (1995)

    1995

  51. [59]

    L. X. Benedict, E. L. Shirley, and R. B. Bohn, Theory of optical absorption in Diamond, Si, Ge, and GaAs, Phys- ical Review B 57, R9385 (1998)

    1998

  52. [60]

    V. Alex, S. Finkbeiner, and J. Weber, Temperature de- pendence of the indirect energy gap in crystalline silicon, Journal of Applied Physics 79, 6943 (1996). 8

    1996

  53. [61]

    Bassani and G

    G. Bassani and G. Parravicini, Electronic States and Op- tical Transitions in Solids , Vol. 8 (Elsevier Science & Technology, 1975)

    1975

  54. [62]

    J. Tauc, R. Grigorovici, and A. Vancu, Optical properties and electronic structure of amorphous germanium, Phys. Status Solidi B 15, 627 (1966)

    1966

  55. [63]

    M. D. Sturge, Optical absorption of gallium arsenide be- tween 0.6 and 2.75 eV, Physical Review 127, 768 (1962)

    1962

  56. [64]

    D. E. Aspnes and A. A. Studna, Dielectric functions and optical parameters of si, ge, GaP, GaAs, GaSb, InP, InAs, and InSb from 1.5 to 6.0 eV, Physical Review B 27, 985 (1983)

    1983

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.