One of the most computationally expensive parts is convergence, which depends on the number of configurations.
We already exploit a lot of stochastic cleaning by removing the SSCHA forces and the average residual force of the ensemble from the forces.
However, in some highly anharmonic and low-symmetry cases, the number of required configurations to converge is extremely high (in some cases over hundreds of thousands, and maybe in glasses or proteins can easily exceed the millions).
In this case, we need to investigate solutions (even approximate ones), like a cutoff on high-rank interactions, or other tricks that reduce the stochastic noise (maybe check whether the Sobol can be improved or other clever stochastic techniques).
This is the last performance-killing issue of the SSCHA, if we manage to implement the interpolation correctly.
One of the most computationally expensive parts is convergence, which depends on the number of configurations.
We already exploit a lot of stochastic cleaning by removing the SSCHA forces and the average residual force of the ensemble from the forces.
However, in some highly anharmonic and low-symmetry cases, the number of required configurations to converge is extremely high (in some cases over hundreds of thousands, and maybe in glasses or proteins can easily exceed the millions).
In this case, we need to investigate solutions (even approximate ones), like a cutoff on high-rank interactions, or other tricks that reduce the stochastic noise (maybe check whether the Sobol can be improved or other clever stochastic techniques).
This is the last performance-killing issue of the SSCHA, if we manage to implement the interpolation correctly.