Dresden 2026 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz

AKPIK 6: AI Methods for Physics and Materials Science

AKPIK 6.2: Talk

Thursday, March 12, 2026, 17:00–17:15, BEY/0127

Causal--Physical Descriptor Discovery for Interpretable Materials Informatics — •Kanchan Sarkar and Axel Gross — Institute of Theoretical Chemistry, Ulm University, 89069 Ulm, Germany

Linking data-driven models to physically grounded behavior remains a key challenge in materials informatics. Data Nexus Vista (DNV1) is a causally informed framework that integrates domain knowledge with machine learning to identify interpretable descriptors. It provides standardized, configurable workflows spanning data preprocessing, feature construction, model training, and interpretation. DNV1 combines established feature-importance and descriptor-design tools with causal analyses, including counterfactual interventions, allowing for direct assessment of how model predictions respond to controlled changes. The framework supports both graphical and programmatic interfaces and provides descriptor-tracing utilities that map model features to physically meaningful variables. All workflows and outputs adhere to FAIR principles, ensuring reproducibility and transparency. The framework has been tested on multiple datasets and demonstrated with DFT-computed spinel cathodes. Rather than limiting ML models with fixed descriptors, DNV1 allows descriptors to emerge through causal interrogation of the data-physics nexus.

Keywords: Materials Informatics; Descriptor Discovery; Interpretable Modeling; Open-Source Methodology; Causal Feature Attribution