Coefficient of determination is not a good way to judge the accuracy / reliability of a non-linear model. The coefficient of determination describes the amount of variation in the RV that can be attributed to the amount of variation in the EV, assuming a linear association.
If applying a transformation still gives a non-linear model, but the r^2 has increased, this does not necessarily mean that the new non-linear model is better than the old one. Strictly speaking, you can only use increasing r^2 to judge effectiveness of a transformation if the associations are linear.
But, as mentioned above, if you suspect you've got some dodgy data on a SAC, and are being asked to apply linear regression techniques to non-linear data, you should consult your teacher. They may not be able to discuss any specifics with you, but it could be considered when marking / moderating SACs.