Encoded time-series model training with U-Net running on Wafer Scale Engine

U-Net-based models, traditionally used for image segmentation, can be adapted for time series by integrating attention mechanisms, by replacing 2D convolutions with linear complexity, or by transforming univariate time series data into 2D representations. In this work, the U-Net architecture was used for regression model training, where time series data were encoded in 2D images and compute acceleration of the training was achieved through massive workflow parallelism on Cerebras wafer scale engine. Geological reservoir modeling as case study demonstrated the challenges of processing dissimilar input-output image pairs and the effects of the model complexity and training time on image "ghosting" and the overall model quality.