Is Spatial AI Special?

The application of machine learning to geology has attracted a lot of attention, but questions remain about which methods are suitable and how they can be applied. Compared to these conventional applications of machine learning, geological applications provide a number of challenges. Geological observations and data are dependent on their spatial and/or temporal context, the number of data points is generally small, and in many cases, the data need to be accompanied by a measure of uncertainty. The current set of established machine learning methods struggles with the limitations outlined above. As an example, Convolutional Neural Networks (CNNs) can make use of spatial neighbourhood relations, but require very large training datasets and cannot report uncertainty. Random Forest, another popular method, can report uncertainty, can be trained with modest-sized datasets, but does not take spatial relationships into account. Generally, machine learning models cannot be used to extrapolate beyond the parameter space of their training data. Model explainability is being able to examine the manner a model decision was reached and is becoming a requirement in sectors such banking and medicine. Not all algorithms offer explainability, such as CNNs and deep-learning methods. The same requirements for explainable geological applications exist, and will likely be enforced as best-practice, if not policy, for high-exposure decisions in resource definition and mineral processing. Various feature importance methods and surrogate modelling can assist, but the workflow needs to be able to accommodate these sometimes complex with commensurate training for the decision-maker. This paper reviews a selection of these limitations and current methods for their solution. Further examination of geoscience ML applications is presented with the aim of ensuring physical realism in the resultant models as a much needed sanity check.