You might recall that in a past Open ForceField meeting Christopher mentioned how soft core potentials could improve the phase space overlap when converting a LJ 12-6 to a Mie 16-6. I don't think Christopher meant that we literally use a soft core potential (in fact he might have misunderstood when I said Mie lambda-6 since the soft core has a lambda term as well) but I think the concept of only gradually turning on the Mie 16 repulsive barrier could be beneficial.
I have played around with this idea a bit but I am considering giving it more attention. Essentially, the reason why we get poor phase space overlap for Potoff and TraPPE is because the closest interacting pairs sampled with TraPPE are too close for Potoff and, thus, result in very high energies and forces. (The inverse scenario of Potoff as the reference and TraPPE as the MBAR reweighting results in too few close interactions).
The idea that I have had is that at some point we switch the Potoff potential back to the TraPPE potential. The point at which Potoff would switch to TraPPE is unclear but I have several ideas. For example, at the point where the energies of TraPPE are greater than those of Potoff we use the TraPPE potential instead in our "rerun". In other words, u = u_Potoff if u_Potoff < u_TraPPE else u = u_TraPPE. This is probably not the best approach, especially when sigma changes as well but it is a start. What this would do is allow the closest interactions to not yield completely unrealistic energies and pressures. Here is what it would look like (see "Potoff, TraPPE-core"):
Notice that I have included the traditional soft core for TraPPE and Potoff (with lambda_soft_core=0.9, alpha =0.5, m=1,n=1). These do not really get us anywhere useful since they will likely result in very poor estimates of VLE due to the poor dispersive tails.
The theoretical basis behind switching the potential back to the reference at some distance can be visualized by the following figure.
This figure plots the running sum of the "true" integrals of the pressure from the simulated PCFs for the TraPPE and Potoff force fields at various temperatures. Note that at a distance (slightly less than 0.38) the two running sums are equal for a given temperature. This suggests that if we use the TraPPE potential up until that point of intersection and the Potoff afterwards we would get an accurate prediction of the Potoff model using MBAR.
I was hoping that we might be able to determine the point where this intersection takes place a priori by making use of the zeroth order PCF. Specifically, I solved for the distance when the running sum of P of TraPPE and Potoff obtained using the zeroth order PCF are equal. This is what the "P0_equal" lines represent. Notice that they are temperature dependent but not strongly enough. The distance where the two integrals are equal is a good estimate of the intersection at low T but gets worse at high T.
@mrshirts
You might recall that in a past Open ForceField meeting Christopher mentioned how soft core potentials could improve the phase space overlap when converting a LJ 12-6 to a Mie 16-6. I don't think Christopher meant that we literally use a soft core potential (in fact he might have misunderstood when I said Mie lambda-6 since the soft core has a lambda term as well) but I think the concept of only gradually turning on the Mie 16 repulsive barrier could be beneficial.
I have played around with this idea a bit but I am considering giving it more attention. Essentially, the reason why we get poor phase space overlap for Potoff and TraPPE is because the closest interacting pairs sampled with TraPPE are too close for Potoff and, thus, result in very high energies and forces. (The inverse scenario of Potoff as the reference and TraPPE as the MBAR reweighting results in too few close interactions).
The idea that I have had is that at some point we switch the Potoff potential back to the TraPPE potential. The point at which Potoff would switch to TraPPE is unclear but I have several ideas. For example, at the point where the energies of TraPPE are greater than those of Potoff we use the TraPPE potential instead in our "rerun". In other words, u = u_Potoff if u_Potoff < u_TraPPE else u = u_TraPPE. This is probably not the best approach, especially when sigma changes as well but it is a start. What this would do is allow the closest interactions to not yield completely unrealistic energies and pressures. Here is what it would look like (see "Potoff, TraPPE-core"):
Notice that I have included the traditional soft core for TraPPE and Potoff (with lambda_soft_core=0.9, alpha =0.5, m=1,n=1). These do not really get us anywhere useful since they will likely result in very poor estimates of VLE due to the poor dispersive tails.
The theoretical basis behind switching the potential back to the reference at some distance can be visualized by the following figure.
This figure plots the running sum of the "true" integrals of the pressure from the simulated PCFs for the TraPPE and Potoff force fields at various temperatures. Note that at a distance (slightly less than 0.38) the two running sums are equal for a given temperature. This suggests that if we use the TraPPE potential up until that point of intersection and the Potoff afterwards we would get an accurate prediction of the Potoff model using MBAR.
I was hoping that we might be able to determine the point where this intersection takes place a priori by making use of the zeroth order PCF. Specifically, I solved for the distance when the running sum of P of TraPPE and Potoff obtained using the zeroth order PCF are equal. This is what the "P0_equal" lines represent. Notice that they are temperature dependent but not strongly enough. The distance where the two integrals are equal is a good estimate of the intersection at low T but gets worse at high T.