DASF: A high-performance and scalable framework for large seismic datasets

Machine learning (ML) techniques have become important tools for seismic data analysis. However, the sheer size of seismic datasets and lack of proper tools still impose challenges when exploring ML techniques on large datasets, thus limiting many researchers to just explore small datasets. In this work, we present DASF, a framework designed to facilitate the exploration of ML techniques with large seismic datasets. DASF provides support for reading traditional and high-performance seismic dataset formats, computing wellestablished seismic attributes, and training classical and deep machine learning models. To showcase the performance and scalability of DASF, we present three case studies: the computation of seismic attributes, seismic facies classification with unsupervised learning, and seismic attribute estimation with deep learning models. As demonstrated by our results, DASF is capable of quickly computing a variety of seismic attributes that can be combined to enhance ML pipelines for seismic processing. It is also capable of processing larger datasets as the number of computing nodes increases. Finally, we show that DASF can leverage GPUs to speed the execution up by a factor of 10 to 50 for attribute calculation and ML pipelines.