A new neural network method for reconstructing the depth profile of thermally inhomogeneous materials from a photothermal spectrum is described in detail. The possibilities and limits of this kind of photothermal depth profiling are evaluated. The influence of the network architecture and several profile and signal parameters on the reconstruction quality is discussed. Statistical predictions are given for the accuracy of the reconstructed profiles for different levels of Gaussian noise in the signal and at different depths. ©1996 American Institute of Physics.