A neural network was used to classify laser welds based upon their acoustical signatures. During laser welding of metal, a plasma (a region of ionized gas) is present above the surface of the material. Characteristics of the process cause this plasma to fluctuate over time, generating an airborne acoustical signal. In this work, the signals from laser welds were detected by a microphone, digitized in16‐stime slices and Fourier transformed. The 2048‐point power spectrum of the transformed signal was input to backpropagation networks of variable size. It was possible to train the backpropagation networks to classify laser welds as full or partial penetration based upon their acoustical signature (in a full penetration weld, the molten zone fully penetrates through the thickness of the metal being welded). Furthermore, it was also found that backpropagation networks with two hidden layers could learn to classify weld signals more efficiently than single hidden layer networks. The ability to classify welds as full or partial penetration will be useful in real‐time process control algorithms to be developed in future work.