Abstract Soil health research exhibits a century‐long recurring pattern of using soil properties as simple indexes in statistical analyses, pedotransfer functions, and machine learning models, often without sufficient connection to physical mechanisms governing soil systems. A classic physics example illustrates the problem: using one radius instead of both principal radii required by the Young–Laplace equation reduces capillary pressure prediction accuracy from r 2 = 1.00 to 0.14, with errors exceeding an order of magnitude. The use of soil properties as indices for comparative benchmarking (e.g., Soil Health Gap) represents proper application distinct from using similar indices for mechanistic process inference. Common soil health indicators, including plant available water (historically defined at −33 and −1500 kPa despite lacking mechanistic validity), bulk density, and biological measures, exhibit substantial temporal variability that undermines predictive utility. Bulk density can change by up to 33% within months in both tilled and no‐till systems, altering saturated hydraulic conductivity by one to two orders of magnitude. Machine learning approaches often lack mechanistic constraints, with only 24% of reviewed studies explicitly addressing uncertainty. Emerging digital twin technologies risk becoming elaborate monitoring systems rather than mechanistic simulation tools. Sensitivity analyses examining how temporal parameter changes affect predictions remain rare despite their critical importance; yet, models incorporating dynamic parameter updates consistently outperform static approaches. We advocate for research priorities that emphasize mechanistic understanding over correlative pattern‐finding in well‐studied systems, reserve index‐based approaches for comparative benchmarking, incorporate temporal dynamics through sensitivity analyses, and recognize soil properties as dynamic expressions requiring mechanistic characterization for reliable prediction.
Daigh et al. (Sun,) studied this question.