GHOST IN THE MACHINE: USING ARTIFICIAL INTELLIGENCE, TRAINED ON GROUNDWATER MODELS, FOR RAPID DECISION SUPPORT (Invited Presentation)

Recent advances in artificial intelligence are providing unprecedented opportunities to learn from and model data for applications in all aspects of society. Many applications leverage huge datasets to train machine learning (ML) models for predictions enabling advances such as computer vision and real-time language translation. Importantly, these techniques focus almost solely on correlation and are not skilled at discerning causal connection. Once trained, ML models make predictions extremely quickly. In natural science applications, two challenges are frequent paucity of data and a lack of certainty to establish causal connections. On the other hand, models that simulate physics and natural processes have strong causal connections but often at great computational expense which can make them cumbersome for decision support. To bridge both the need for solid causal connections and the low computational burden needed for decision support applications, a bridge can be achieved through metamodels. In these cases, ML techniques are used to train on datasets generated by physics simulators achieving the rapid results with causal connections assured by the design of the underlying physics simulators. I will highlight several examples with groundwater models forecasting the source of water to wells and groundwater age. Issues of transferability to non-modeled regions are also explored.