Deep Neural Network for Accurate and Efficient Atomistic Modeling of Phase Change Memory

Insights into the Heterogeneous Nuclei of an Ultrafast‐Crystallizing Glassy Solid

“…To assess the structural dynamics of individual atoms during phase transitions, say, belonging either to the crystalline or liquid clusters, a variant of bond order parameter $$q_{\mathrm{4}}^{{\mathrm{dot}}}$$ for crystallinity was usually employed. [ 36–38 ] The liquid‐crystalline boundary was defined as the threshold value of ≈0.3 for $$q_{\mathrm{4}}^{{\mathrm{dot}}}$$ (Figure 4d) (see the Experimental Section for details). We denoted the averaged $$q_{\mathrm{4}}^{{\mathrm{dot}}}$$ value over all the atoms in the specific clusters as $$Q_{\mathrm{4}}^{{\mathrm{dot}}}$$.…”

“…In addition, it is well known that the periodic boundary conditions have to be applied in AIMD simulations, which unavoidably introduce artificial size effects. We anticipate the potential small-size issue [45] for our AIMD simulations can be solved by employing large-scale and high-efficiency atomistic modeling based on machine-learning techniques with ab initio accuracy, [38,46] which provides a powerful approach to enable more in-depth analysis on the phase-transition mechanisms of the SST PCMs.…”

CMOS Scaling for the 5 nm Node and Beyond: Device, Process and Technology

“…This has been demonstrated in a recent study, in which the neurons act as the tight-binding (TB) matrix elements in the Hamiltonian parameterization of the TB model for energy band calculation [124]. From calculation methodology to case studies, several studies have been reported in which machine-learningaugmented DFT works efficiently with high accuracy in atomistic modeling for devices, including the prediction of atomic force in phase change memory [125], the calculation of potential energy surface in SiGe alloys [126], and the simulation of surface reconstruction of the Si ( 111)-(7 × 7) surface [127].…”

Machine learning interatomic potential: Bridge the gap between small-scale models and realistic device-scale simulations

“…Notably, the NN potential is also applicable to slightly altered GeTe alloys but faces limitations with off-stoichiometric compositions, such as Ge 0.15 Te 0.85 , due to its inability to accurately model long Te-Te chains. 181 , 182 , 183 Additionally, using the GAP framework, another research 184 introduces an MLIP for the single ternary PCMs compounds of Ge 2 Sb 2 Te 5 , which enabled the creation of a detailed 7200-atom model, providing insights into the material’s structure, and facilitated the generation of smaller models for in-depth chemical bonding studies, as shown in Figure 5 B. Furthermore, without extra datasets, they expand the above model to six Sb-Te alloy PCMs, 185 , 186 with compositions ranging from 2:3 to 4:1, revealing that all exhibit similar local structural motifs of defective octahedra.…”