A Framework for AI-Ready Data in Earth and Environmental Science

Machine learning is becoming central to Earth and environmental research. Yet the FAIR principles, while an essential foundation, do not cover several requirements that arise when data are used to train and evaluate AI models. This talk presents a framework of 25 criteria for assessing and improving the AI-readiness of research data in distributed data infrastructures. It was developed by a multidisciplinary group from seven Helmholtz centres, representing data providers, infrastructure developers, data managers and AI practitioners. The criteria cover five categories: data preparation, data quality, documentation, access, and targets for supervised learning. Eleven are identical to FAIR criteria; fourteen add AI-specific requirements such as treatment of missing data, bias and uncertainty documentation, human operator decisions, consent for AI use, and documented quality control of labels. Each criterion was co-designed by working through the question it answers for ML users, why it matters, who benefits, and how it is put into practice. Seven use cases show that the importance of individual criteria strongly depends on the application. We therefore argue that AI-readiness is primarily a matter of transparency and fitness for purpose rather than a fixed level of data quality. The framework is intended as a basis for dialogue between data providers, infrastructure operators and AI practitioners.