Nonlinear interpolated Variational Autoencoder for generalized fluid content estimation

Generalizing machine learning models for petroleum applications, especially in scenarios with limited and less varied training data compared to real-world conditions, remains a persistent challenge. This study introduces a novel method combining interpolation mixup with a Variational Autoencoder (VAE) and adaptable interpolation loss for downstream regression tasks. By implementing this approach, we generate high-quality interpolated samples, yielding accurate estimations. Experimental validation on a real-world industrial dataset focused on fluid content measurement demonstrates the superior performance of our method compared to other interpolation and regularization techniques. Our approach achieves over a 15% improvement on generalized out-of-distribution datasets, offering crucial insights for fluid content estimation and practical implications for industrial applications. • DAS data enables accurate identification of fluid content in pipelines. • ML struggles to generalize when test data differs from training data. • Generative modeling aids water breakthrough and ratio estimation in DAS analysis. • Neural networks assess generative models for quality in downstream tasks.