Data transformation when residuals vs fitted is linear




I am doing a linear mixed-effects model with response variable that is non-negative but include some zeros. And my predictor variables are age and other fatorial variables.

I checked my assumptions with a graph of residuals vs fitted values and I got a very nice linear relation... So I am trying some transformation to get the proper non-linearity graph. I tried log(x+1) with the response variable, I tried Log(age). I also tried the square root with response variable... and nothing changed.

Is there a magic transformation for that kind of problem ?