This letter presents a general-purpose fully-atomistic method to simulate phase change memory (PCM), by combining density functional theory (DFT) and deep neural network (DNN). Its maximum calculation error of atomic forces is about 10⁻¹ eV/Å, which is 1–2 orders of magnitude more accurate than state-of-art artificial neural network (ANN) in PCM literature (over 10¹ eV/Å). Its simulation time, ${t}_{\text {s}}$ , scales linearly with the number of atoms ${n}_{\text {a}}$ ( ${t}_{\text {s}}\propto n_{\text {a}}$ ), which is more efficient than DFT ( ${t}_{\text {s}}\propto {n}_{\text {a}}^{3}$ ) widely used to model PCM, leading to approximately 2, 4, 6 orders of magnitude reduction of modeling time when ${n}_{\text {a}}\approx {10}^{{1}}$ , 10², 10³, for instance. Its efficiency and accuracy may be useful to develop next-generation atomistic modeling tools to enable in-depth optimization of PCM.

中文翻译：

---

用于相变记忆准确有效原子建模的深度神经网络

这封信通过结合密度泛函理论 (DFT) 和深度神经网络 (DNN)，提出了一种通用的全原子方法来模拟相变记忆 (PCM)。其原子力的最大计算误差约为 10 ^-1 eV/Å，比 PCM 文献中最先进的人工神经网络 (ANN)（超过 10 ¹ eV/Å）精确 1-2 个数量级。它的模拟时间， ${t}_{\text {s}}$ , 与原子数成线性比例 ${n}_{\text {a}}$ ( ${t}_{\text {s}}\propto n_{\text {a}}$ )，这比 DFT ( ${t}_{\text {s}}\propto {n}_{\text {a}}^{3}$ ) 广泛用于对 PCM 建模，导致建模时间减少大约 2、4、6 个数量级，当 ${n}_{\text {a}}\大约 {10}^{{1}}$ , 10 ² , 10 ³，例如。它的效率和准确性可能有助于开发下一代原子建模工具，以实现 PCM 的深入优化。