The form of the predictor is irrelevant to the choice of family which describes the conditional distribution of the response.
There are a number of families on might use for a continuous positive random variable.
I'd probably start with considering a Gamma family (variance proportional to mean squared), possibly with a log-link (your own subject matter knowledge should form a better basis on which to choose a link though).
There's also Inverse Gaussian (variance proportional to mean cubed) in most implementations of GLMs, and in some packages you can use the Tweedie family (power variance function -- which includes those earlier two as special cases).
Another alternative that's sometimes used with concentrations is the lognormal; with positive data one might take logs and then fit a (possibly linear) least squares regression model (since the assumed conditional distribution would be normal after taking logs).