I wanted to know if there is any way I can calculate the cummulative grain size distribution from geomtric mean and standard deviation, eventaully d15 and d85 of the mixture.
Without additional information there is no way to identify definitively a cumulative distribution or even the general shape from only a variant of the first two moments (for you: G_mean, SD). If you're willing to make strong assumptions about the distributional shape, you'd have better luck.
Indeed, we can find not full curve but it seems D15 and D85 can be obtained but I am not sure if the proposed relation can be used in my case.
I came across some literature where some relationships between percentage finer (D15.9 and D84.1) and statistical parameters (geometric mean; Dg and and standard deviation; σg) are proposed assuming grain size distribution follows log normal distribution (often natural sediments can be approximated to log normal distribution). The relationships are as follows:
Dg=(D84.1*D15.9)^1/2
σg=log(D84.1*D15.9)^1/2
I am confused about applicability of these relations because when I compare the calculated values with actual D15.9 and D84.1 (directly extracted from cummulative grain size curve), they are slightly different. The actual grain size distribution ranges from 2mm (min) to 32mm (max) but calculated D15.9 is 1.63mm (smaller than 2mm), which is not possible.
My calculations were based on log base 2, is it right?
That aside, it's always possible that the performance of the estimator you've found will be less than perfect, for a number of reasons. One famous quote comes to mind, "All models are wrong, but some are useful."