Agentic AI for Autonomous Gravitational Lensing Simulation Workflows

DeepLenseSim (built on lenstronomy) requires substantial manual intervention for configuring parameters, managing outputs, and iterating on failures across multi-step simulation pipelines, creating bottlenecks in large-scale dataset generation and limiting parameter space exploration. I will build an Agentic AI framework following the HEPTAPOD philosophy to autonomously orchestrate DeepLenseSim workflows across three phases: (1) implementing schema-validated Pydantic tools wrapping the DeepLens class for all four model configurations, an LLM orchestration layer via OpenRouter function calling, human-in-the-loop confirmation flow, and a compatibility layer addressing pyHalo/lenstronomy/numpy version conflicts; (2) building a ParameterScanTool for automated multi-point sweeps (axion mass, halo mass, redshift), a ValidationTool for automated image quality checks, and exploration of RL-based parameter optimization; (3) comprehensive documentation, tutorial notebooks, a test suite with >80% tool function coverage, and integration with downstream DeepLense ML training pipelines via standard export formats (numpy, HDF5, FITS).