Scholarlink

<p>Simulating phonon-assisted optical spectra in semiconductors at finite temperatures presents a classic accuracy-efficiency dilemma: thermodynamic configurational sampling requires large supercells, while accurate electronic properties demand expensive exchange-correlation (XC) functionals. In this work, we present a machine learning workflow that integrates deep learning potentials (DeePMD) with tight-binding (DeePTB) models to overcome this bottleneck. This approach enables non-perturbative treatment of finite-temperature structural sampling and configuration-dependent electronic evaluation within the Born-Oppenheimer (BO) approximation. By confining expensive Heyd-Scuseria-Ernzerhof (HSE) calculations to small cells and transferring the trained Hamiltonian to supercells of up to 4096 atoms, we achieve large-scale simulations of temperature-dependent optical properties without empirical corrections. We validate this framework by modeling phonon-induced bandgap renormalization and direct/indirect absorption in mainly two classes of materials: non-polar Si, weakly-polar GaAs, and GaP (100-400 K). Our findings are in excellent agreement with experiments. This method enables high-fidelity ab initio simulation of electron-ion-coupled phenomena in this wide class of materials at previously inaccessible scales.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep neural network for phonon-assisted optical spectra of semiconductors at finite temperatures

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月20日

查看全文
  • Shengguo Yang,
  • Shishir Kumar Pandey,
  • Zhanghao Zhouyin,
  • Qiangqiang Gu

摘要

Simulating phonon-assisted optical spectra in semiconductors at finite temperatures presents a classic accuracy-efficiency dilemma: thermodynamic configurational sampling requires large supercells, while accurate electronic properties demand expensive exchange-correlation (XC) functionals. In this work, we present a machine learning workflow that integrates deep learning potentials (DeePMD) with tight-binding (DeePTB) models to overcome this bottleneck. This approach enables non-perturbative treatment of finite-temperature structural sampling and configuration-dependent electronic evaluation within the Born-Oppenheimer (BO) approximation. By confining expensive Heyd-Scuseria-Ernzerhof (HSE) calculations to small cells and transferring the trained Hamiltonian to supercells of up to 4096 atoms, we achieve large-scale simulations of temperature-dependent optical properties without empirical corrections. We validate this framework by modeling phonon-induced bandgap renormalization and direct/indirect absorption in mainly two classes of materials: non-polar Si, weakly-polar GaAs, and GaP (100-400 K). Our findings are in excellent agreement with experiments. This method enables high-fidelity ab initio simulation of electron-ion-coupled phenomena in this wide class of materials at previously inaccessible scales.