Full waveform inversion (FWI) is a powerful tool to reconstruct the material distribution based on sparsely measured data obtained by wave propagation. FWI has improved in the recent years with the integration of Neural Networks (NN). One possibility is supervised learning, where the NN is used to map measured signals directly to the material distribution of the domain under investigation. However, such a purely supervised approach can lead to wrong predictions for signals outside the training data set. We present a transfer learning framework to improve the convergence of NN-based FWI. It roots in the idea that NNs may favourably be combined with classical FWI as in [1] utilizing an adjoint formulation. Instead of starting the inversion process with randomly initialised weights, we perform a pre-training of the neural network in a supervised manner. This approach of transfer-learning presented in [2] already provides reliable predictions for data outside the training set and reduces the number of iterations required for convergence. In this talk, we demonstrate that further improvements are possible by using gradients calculated from the classical adjoint method to train the NN which maps to the corresponding material distribution. Subsequently, we use the pre-trained network as a predictor in the spirit of transfer learning to provide a better starting point for the inversion process. REFERENCES [1] L. Herrmann, T. Bürchner, F. Dietrich, and S. Kollmannsberger, “On the use of neural networks for full waveform inversion,” Comput. Methods Appl. Mech. Eng., vol. 415, p. 116278, Oct. 2023, doi: 10.1016/j.cma.2023.116278. [2] S. Kollmannsberger, D. Singh, and L. Herrmann, “Transfer Learning Enhanced Full Waveform Inversion,” in 2023 IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM), IEEE, Jun. 2023, pp. 866–871. doi: 10.1109/AIM46323.2023.10196158.