Machine learning (ML) approach for dispersion curve inversion using mixture density networks (MDN) based on Keil and Wassermann (2023). The ML approach presented here allows the simultaneous estimation of layer numbers, layer depth and a complete probability distribution of the S-wave velocity structure in the upper 100 m. This is achieved by a two-step ML approach, where 1) a regular NN classifies the number of layers within the upper 100 m of the subsurface and 2) individual trained mixture density networks output the depth estimates together with a fully probabilistic solution of the S-wave velocity structure. We trained the model to distinguish structures with 2 - 7 subsurface layers. The trained classification NN and the individual MDNs are located in the folder ./trained_models. With the jupyter notebook Prediction.ipynb the dispersion curve inversion can be performed using the already trained ML models. With the jupyter notebooks Training-MDN.ipynb and Training-classification.ipynb the models can be trained on new data. The code for the set-up of the MDN is based on Earp et al. (2020). More details and updates on the code can be found on: https://github.com/SabrinaKeil/MDN_Inversion

{"references": ["Earp, S., Curtis, A., Zhang, X., & Hansteen, F. (2020). Probabilistic neural network tomography across Grane field (North Sea) from surface wave dispersion data. Geophysical Journal International, 223(3), 1741-1757.", "Keil, S. and Wassermann, J. (2023). Dispersion curve inversion using mixture density networks. Geophysical Journal International.(Submitted)"]}