Here’s an example of how AI science could work. If you ask machine learning to review a data set from a system and predict the future behavior of the system, it can often do a good job, but the mathematical approach it is using to do that is mostly opaque. It is also divorced from any sense of the system structure, its interaction with the physical universe, and the physical laws governing it. But you can ask the computer to constrain itself within those known physical laws, and then it may be able to provide you insights on the structure and physical processes inside the system. So at that point, you and/or the computer should be able to form scientific hypotheses and test them against the data. This example from Water Resources Research is about soil moisture.
Coupling Intuitive Physics Into Deep Learning for Soil Moisture Flow Processes Learning
Soil water flow processes in the unsaturated zone support ecosystems and regulate water, energy, and biogeochemical cycles. Recently, deep learning (DL) approaches have significantly advanced soil moisture (SM) prediction tasks yet still challenging to interpret. It’s difficult to peer into the internal reasoning procedures of algorithms, let alone associate them with specific physical processes. Thus, DL alone is unlikely to satisfy soil hydrological modeling needs and cannot advance process understanding. Here, we present DPL-S (deep process learning for SM dynamics) approach, which couples intuitive physics into deep learning architecture as structural guidance to facilitate comprehensive surrogate modeling of soil water flow. DPL-S discretizes the SM state evolution into multiple sub-process effects (e.g., gravity, matric potential) at the intuitive physics level and abstracts them into format-specific and learnable tensors. By cascading state-action matrices in a differentiable end-to-end framework and enforcing penalties for physical inconsistencies, DPL-S enables a profound understanding of physical functions and scenes of soil water flow. Comprehensive numerical experiments including layered soil conditions and tests with in situ observations, demonstrate that it achieves reliable SM profile reconstruction with predictive performance comparable to the state-of-the-art DL model on supervised items. The internal inference of DPL-S is fully transparent and the tensor representations achieve strict physical realism under limited water content supervision, thus enabling continuous predictions like physical models during the testing period. The model’s flexibility, generalization, noise resistance, and large-sample diverse data synergies are also evaluated. This work represents a solid step toward learning hydrophysical processes from large data sets.
