The combination of natural language processing and machine learning is transforming materials discovery paradigm through scientific literature extraction and intelligent materials discovery. However, batch effects, systematic inter-laboratory differences, represent critical confounding factors that compromise the prediction of reliable structure-activity relationships. We developed an automated knowledge discovery framework which enables a full-pipeline automation from PDFs to causal mechanism discovery. The framework combines hybrid error correction, visual paragraph screening, and large language model (LLM) extraction to convert unstructured literature into high-quality databases while incorporating batch effect control and causal analysis. Application to CO₂ hydrogenation to methanol literature revealed that batch effects account for most performance variance and analysis without batch effect control yielded physically unsound conclusions. Through causal inference methods including directed acyclic graphs (DAGs), DOI fixed-effects with cluster interactions, double machine learning, and augmented inverse probability weighting, batch effects and confounders are minimized, and reliable catalytic correlations for several typical catalysts of CO₂ hydrogenation to methanol are obtained. Our findings demonstrate that batch effect control is essential for literature-based knowledge extraction, providing a methodological paradigm for heterogeneous data analysis in data-driven catalyst design.
Xu et al. (Fri,) studied this question.