Modern science generates enormous amounts of data, but many artificial intelligence (AI) tools function as so-called "black boxes," whose internal workings and decision-making processes are often difficult to understand. As a result, the interpretation of predictions is limited, and in certain cases, models may even overlook basic laws or factors. To advance science, we need explainable artificial intelligence that respects fundamental physical principles, enables the integration of multimodal data (image, text, numerical), and effectively combines experimental knowledge with findings from the literature. Such an approach can significantly accelerate development in many fields, from medicine and biotechnology to materials and environmental sciences.

The project is organized into four work packages: (A) explainable machine learning (including neural-symbolic methods, prediction explanation, literature trend tracking), (B) multimodal basic models (health, oncology, materials), (C) automated modeling with equation and differential equation discovery, (D) ontologies and databases for FAIR/open science and AutoML/AutoOPT. Specific examples are envisaged: pre-screening of mammograms and prognosis for primary brain tumors, saRNA and drug design, modeling of Lake Bled, electrocatalysis, and online "observatories" of scientific trends. 

The goal is to develop explainable, multimodal, and semantically grounded approaches based on artificial intelligence, test them in selected scientific domains, and establish an open infrastructure (models, data, ontologies) for repeatable, explainable, and transferable artificial intelligence in science. The project is led by a Slovenian consortium: Jožef Stefan Institute, University of Ljubljana, National Institute of Chemistry, University of Maribor, National Institute of Biology.