Virtual flow meters, which estimate flow rates without physical sensors, are often built from either mechanistic or data-driven models. A new preprint on arXiv argues that in sectors like energy, aquaculture, and chemical processing, both model types are frequently held back by poor-quality training data. Such data obscures the underlying physical relationships the models are meant to capture, leading to degraded performance.

The authors propose a text-informed training data filtering approach. Instead of relying solely on numeric sensor readings, the method uses textual information associated with the data to identify and remove low-quality samples before training. This is intended to ensure the remaining dataset better reflects the true physical behavior of the system.

By cleaning the training data upfront, the resulting virtual flow meters should be more reliable. The preprint does not provide specific performance figures in the abstract, but it points to a promising direction: using descriptive text as a guide for data curation, a step that is often overlooked in favor of more complex modeling techniques.