All you need is noise — from feature selection to explainable industrial AI

Modern chemical plants record thousands of sensor tags, yet only a small fraction meaningfully influence yield, quality, or throughput. Identifying those key drivers is often more difficult than building the predictive model itself. In this work, we show that appending one or more Synthetic Noise Features (SNFs), non-informative random variables known a priori, provide a simple reference for judging variable relevance. We show the impact of this model agnostic step across three workflows. In supervised learning, noise features establish an automatic cutoff for the feature importance, guide model regularization and signal when the dataset itself lacks predictive information. In unsupervised learning, they provide an unbiased threshold preventing spurious anomalies and latent dimensions. Finally, we demonstrate the applicability of this approach to small datasets typical of experimental work and Design of Experiments (DoE), including Definitive Screening, Response Surface, and space-filling designs, as well as active learning using Bayesian optimization. By turning nothing but noise into a quantitative benchmark, SNFs offer an immediately deployable safeguard against overfitting and misplaced experimental effort in data-driven chemical engineering.