Hybrid AI permafrost modelling

Deep learning is an approach capable of extracting spatio-temporal features automatically while processing large amounts of data through complex structures. Structures that, for example, are able to learn from past patterns and share with the future if strong correlation is found. It could be assumed AI models only need to be built and gather enough data to find links between input and outputs. However, this approach cannot ensure that predictions would respect the laws of physics, e.g. due to extrapolation or observational biases. Restricting models by introducing physics can add strong theoretical guidelines alongside observations.
In the context of permafrost models, data observations are lacking (e.g. wind speed or humidity) and models lose the possibility of spatial extrapolation. In this case, simplification of the physics is an usual procedure. Such that the problem is solved by an approximate solution that still captures broad spatio-temporal features while responding to more accessible predictors (e.g. surface air temperature or air pressure). More specifically, permafrost present-day thermal state is the consequence of past climate conditions that induced long-term variations of deep reservoirs of organic carbon and ground ice. Reproducing permafrost evolution at century to millennia scale requires models to operate with limited and highly uncertain information about thermal and hydrological ground properties.
In need of both data and physical constraints, climate models themselves could be used as data generators. Here, CryoGrid Lite, a simplified version of the permafrost model CryoGrid 3, is used to simulate ground thermal regime and ice balance. Daily data of air temperature, pressure and geothermal flux run CryoGrid Lite to simulate the evolution of the thermal state of permafrost and active layer thickness over many centuries for the Canadian Arctic permafrost region. This dataset, generated by CryoGrid Lite, trains a neural network model to emulate its behavior. Physics equations governing the original model are also introduced into the objective function to penalize the network training when outputs exceed a tolerance range. This approach restricts outputs to the knowledge provided by CryoGrid Lite, enhancing physical reliability of forecasts. This is in contrast to the traditional 'black-box' structure of neural networks, which usually rely on minimizing errors with respect to observations. An assessment of the impact of including such additional constraints is provided. 

This study explores an hybrid approach between coupling physical process models with the flexibility of data-driven machine learning. The inclusion of physics within AI structures could improve their performance in permafrost modeling, while overcoming the reliability challenge that hampers its adoption in geoscience.