Workshop Description

Deep learning (DL) for seismic processing is a very active field of research. Currently most of the investigations consist in learning to reproduce the results obtained by classical seismic processing algorithms or workflows, i.e. to predict the processed image from the “initial” image using deep neural networks (DNNs). However, DL for seismic processing encounters specific challenges:

• Many state-of-the-art algorithms and workflows available for seismic processing are physics-based. They often use regularisations, and user-defined parameters can be tuned with the geophysicist's domain expertise in an interactive fashion before the algorithm is applied to the entire dataset. Such fine control (that allows adapting the processing method to the peculiarities of the data) is harder to enforce in a DL framework. Also, training (the most computationally expensive part of a ML inference problem) must be run to convergence before the goodness of the trained model can be evaluated. Moreover, operations performed in DNNs are not straightforward to both interpret physically and QC, which is uncommon in seismic processing. The corresponding processes (algorithms together with user’s tuning) are highly non-linear and being able to outperform them represents a challenge for DL.  
• Seismic data is very different from natural images. Raw seismic traces are time-series data with multidimensional location that are seldom regularly sampled. They are signed and oscillatory, with a frequency bandwidth (typically 2.5-150 Hz) that must be preserved for further processing tasks. They are made of coherent events and the features of interest are, in addition to the kinematics, the variations in amplitude, phase and event shape (wavelet). The amplitudes of the seismic volumes are important to preserve, which represents a challenge for DNNs.

Seismic particularities necessitate adapting the existing DL methods for industrial use in seismic processing. Especially, four important challenges should be overcome:

1. Intelligently defining good training data (both quantity and diversity of the data are critical).
2. Ensuring the best DNN training (ultimately give robustness to noisy input data, estimates of uncertainty...). 
3. Devising efficient training strategies. Use of transfer learning, automatic monitoring during training, and adaptive hyper-parameter adjustment to reduce turn-around time.
4. Model interpretability. Physics based interpretations of DNN and its components, resisting the urge to look at them as “black boxes” (fine-tuning DNN, interpretation of DNN architectures, analysis of feature maps…)

The workshop topics will be of interest to new and experienced practitioners/researchers in geophysics as well as data scientists developing deep learning workflows for geophysics.  

Workshop Programme

Morning Session
09:30	Welcome and introductions
Session 1: Data processing (same domain)
10:00	Some learnings on Deep Learning for seismic processing E. Vershuur
10:25	Seismic-based Deep Learning with Recurrent Inference Machines I. Vasconcelos
10:50	Coffee break
11:05	Machine-learning seismic processing tasks by fine tuning a pre-trained Attention-based neural network: StorSeismic T. Alkhalifah
11:30	Seismic interpolation with DL M. Fernandez
11:55	Seismic denoise with deep neural network and its application to North Sea datasets P. Zhao
12:20	General discussion
12:45	Lunch
Session 2: Velocity model building and imaging (different domains)
14:15	Wave-equation based inversion with amortized variational Bayesian inference F. J. Herrmann
14:40	Physics assisted deep learning for full waveform inversion M. Sen
15:05	Coffee break
15:20	First Steps in the Application of Physics-Informed Neural Networks to Full Waveform Inversion S. L. De Souza
15:45	Physics-coupled deep learning inversion: geophysical data applications D. Colombo
16:10	Rapid Replication of Geophysical Processing Best Practices with Machine Learning H. Rijnja
16:35	Self-Supervised Learning approach to increase the SNR of low frequencies in Seismic Data, J.-P. Mascomère
17:00	General Discussion - Conclusions

---