Plastic deformation in van der Waals (vdW) crystals is essential for flexible and wearable electronics. However, its evaluation typically requires destructive testing or costly first-principles calculations. To address this, we present a causality-driven, interpretable machine learning framework that extracts accurate plasticity descriptors from easily accessible material features, avoiding empirical assumptions and high computational costs. By leveraging causal discovery, we clarify the roles of interlayer and intralayer interactions in plasticity and identify key features that govern plastic behavior. A causality-constrained symbolic regression method then generates a compact, physically meaningful descriptor with 96% accuracy in distinguishing plastic from brittle vdW materials. This descriptor enables rapid screen of large material databases and identifies 29 previously unrecognized semiconductors with superior plasticity. This framework provides a generalizable approach to convert empirical data into actionable materials knowledge, accelerating functional materials design.

Developing high-performance catalysts using traditional trial-and-error methods is generally time consuming and inefficient. Here, by combining machine learning techniques and first-principle calculations, we are able to discover novel graphene-supported single-atom catalysts for nitrogen reduction reaction in a rapid way. Successfully, 45 promising catalysts with highly efficient catalytic performance are screened out from 1626 candidates. Furthermore, based on the optimal feature sets, new catalytic descriptors are constructed via symbolic regression, which can be directly used to predict single-atom catalysts with good accuracy and good generalizability. This study not only provides dozens of promising catalysts and new descriptors for nitrogen reduction reaction but also offers a potential way for rapid screening of new electrocatalysts.