Atomistic simulations of Ge-rich GeSbTe alloys for phase change memories

Prototypical phase change compounds, typically based on GeSbTe (GST) alloys, display a crystallization temperature not suitable for embedded Phase Change Memories (ePCM) of interest for applications in the automotive sector. The search for an alternative material is thus a very active research field. Ge-rich GST alloys are emerging as promising materials for ePCM thanks to the higher thermal stability of their amorphous phase. Upon crystallization, Ge-rich GST alloys undergo a phase separation into Ge and other GST alloys. The segregation phenomena enhance the crystallization temperature (Tx), but it comes also with several drawbacks such as a high cell-to-cell variability and a drift of the electrical resistance with time in the set state. The details regarding the decomposition process are largely unknown and are a matter of debate. During my PhD studies, I investigated the phase separation by means of high-throughput Density Functional theory (DFT) calculations based on thermodynamical analysis. We computed the formation free energy of all GST alloys in the central part of the ternary phase diagram modelled in the rocksalt metastable phase, which is the phase relevant to the operation of the memory. Then, we computed all possible decomposition reactions for each GST alloy. We summarized all our thermochemical data in one descriptor called "decomposition propensity", which measures the tendency of an alloy to undergo phase separation. I also studied the structural properties of the amorphous phase of Ge-rich GST alloys as a function of the Ge content. We found that by increasing the Ge content the local structure of the amorphous phase becomes more and more dissimilar from the crystalline phase which might hinder the crystallization kinetic. These results suggest a possible strategy to minimize the phase separation (low decomposition propensity) and still keep high Tx (crystallization might be hindered due to the dissimilarity). Aside the thermodynamic analysis discussed above, we should however address kinetics effects that could be modelled for instance by molecular dynamics (MD) simulations. To this end, one should enlarge the scope of DFT framework by developing a Neural Network interatomic potential (NNIP) by fitting a large DFT database. This scheme allows to perform large-scale simulations with a close to DFT accuracy and the speed of classical force fields. As a first step towards the generation of NNIP for Ge-rich GST alloys, we developed a NNIP for Ge2Sb2Te5 compound (the prototypical GST compound) which was used to directly simulate the crystallization process by MD.