Yu J1, Baker P3, Petridis R2, Munnerley J4, Thomas L5
1CSIRO, Clayton, Australia, 2CSIRO, Hobart, Australia, 3CSIRO, Black Mountain, Australia, 4CSIRO, Dutton Park, Australia, 5CSIRO, Hobart, Australia
Biography:
Jonathan Yu is a research scientist and team leader at CSIRO working at the intersection of environmental science, data systems, and digital innovation with and interest in cloud computing and AI technologies. He leads projects that integrate models, datasets, and decision-support tools to address complex challenges in bushfire risk, water resources, and ecosystem management. His work focuses on building scalable, transparent, and reproducible workflows, with an emphasis on FAIR data, provenance, and knowledge graph approaches. Jonathan collaborates with government and research partners to translate science into operational capabilities that support policy, planning, and real-time decision-making across Australia.
Abstract:
Computational modelling workflows supporting data-driven decision-making are often constrained by the specialised domain and technical knowledge required to understand and run the underlying models. Addressing this requires developing usable interfaces that encapsulate domain knowledge and model details.
We present a framework for deploying modelling workflows that integrates model run and data orchestration with automated provenance coordinated via AI-assisted user interfaces. The framework combines four components: (1) Apache Airflow for distributed workflow execution; (2) Provena provenance platform for recording model information, model run provenance, and relevant datasets, managed in a metadata registry to ensure transparent and auditable outputs across complex pipelines; (3) Model Context Protocol servers that provide AI interfaces to LLM tools such as Cursor and Claude; and (4) Containerised models deployed via Docker and cloud infrastructure provide a modular execution environment supporting rapid integration of new model versions.
A key innovation is an AI-mediated interface that enables users to configure, execute, and query workflows using natural language. This abstracts system complexity, allowing users to invoke experiments without needing detailed knowledge of model run orchestration or data management processes. The framework is evaluated through a pilot involving multiple reef models in the Reef Restoration and Adaptation Program. Results show ease of scaling model execution and lowered onboarding effort, while automated registration of workflow metadata and data artifacts in Provena demonstrates feasibility of end-to-end provenance capture. Together, these capabilities demonstrate how combining orchestration, provenance, and AI interfaces can lower barriers to computational modelling while potentially increasing rigour in reproducibility and auditability.
