How to Limit the Epistemic Failure of Machine Learning Models?

The intelligible understanding of natural phenomena such as earthquakes is one of the main epistemic aims of science. Its very aims are shaped by technological discoveries that can change the cognitive fabric of a research field. Artificial intelligence, of which machine learning (ML) is one of the fundamental pillars, is the cutting-edge technology that promises the greatest scientific breakthroughs. In particular, great hopes are placed in ML
models as a source of inspiration for the formulation of new concepts or ideas, thanks to their ability to represent data at different levels of abstraction inaccessible to humans alone.
However, the opacity of ML models is a major obstacle to their explanatory potential. Although efforts have recently been made to develop ML interpretability methods that condense the complexity of ML models into human-understandable descriptions of how they work and make decisions, their epistemic success remains highly controversial. Because they are based on approximations of ML models, these methods can generate misleading explanations that are overfitted to human intuition and give an illusory sense of scientific understanding.
In this study, we address the question of how to limit the epistemic failure of ML models. To answer it, we use the example of an ML model trained to provide insights into how to better forecast newly emerging earthquakes associated with the expansion of hydrocarbon production in the Delaware Basin, West Texas. Through this example, we show that by changing our conception of explanation models derived from interpretability methods,
i.e. idealised scientific models rather than simple rationalisations, we open up the possibility of revealing promising hypotheses that would otherwise have been ignored. Analysis of our interpreted ML model unveiled a meaningful linear relationship between stress perturbation distribution values derived from ML decision rules and earthquake probability, which could be further explored to forecast induced seismicity in the basin and beyond. This observation also helped to validate the ML model for a subsequent causal approach to the factors underlying earthquakes.