And modelling it as an exponential, which seems somehow more sensible, somehow, we have a fit in a logarithmic y scale
y = 4.74 + 0.00526 x - 3.63E-05 x2 + 1E08 x3
where y = loge( - SNR_dB ). Just for those odd people amongst us who like such things.
Since the second-order and higher coefficients are so small relative to the first order then we can say that the original SNR graph in dB approximates pretty well with with an e−x curve.