Résumé
We propose a novel system for automated scientific knowledge discovery at scale, referred to as Discovery Engine (DE). Traditional data analysis methods rely on human assumptions and linear patterns, while performant machine learning models find complex patterns but remain opaque. Our system overcomes these limitations by combining machine learning’s pattern recognition ability with the SOTA interpretability methods [1, 2] that broadly aim at shedding light on decision making of machine learning models. This enables human-understandable insights from complex data that would otherwise remain hidden.While existing work shows AI’s potential for scientific discovery, automation across diverse datasets remains unexplored. The exception to this are Large Language Model (LLM) driven discovery pipelines that aim to go from data to insight directly [3, 4]. However, in practice due to data confidentiality and/or context length issues, it is not trivial to expose the LLM to the entire dataset. This makes it very difficult to judge what part of the discovery is in fact specifically related to the dataset, rather than the LLM’s knowledge based on its training data. This is on top of LLMshortcomings in mathematical tasks in general [5, 6]. DE addresses these limitations, making AI-driven discovery data-driven, automated, systematic and reproducible.Belowwe provide an overviewof DE components, followed by two case studies. For the case study that has ground truth patterns, we make a comparison with the ground truth, and results from a pure LLM pipeline. In the second case study we validate the patterns by exposing them to the academic expert who has collected the dataset for the study.