Speaker
Description
CMS is developing FlashSim, a machine learning–based framework that produces analysis-level (NANOAOD) events directly from generator-level inputs, reducing simulation costs by orders of magnitude. Efficient integration of preprocessing, inference, and output is essential, and ROOT RDataFrame provides the backbone of this workflow.
Certain operations required for FlashSim are not yet part of the native RDataFrame API. These include event batching to optimize GPU utilization, efficient writing of ML-generated events through the RDataFrame interface, and support for oversampling to reuse inputs across multiple ML inference iterations. We implemented these features as custom extensions, but a native ROOT implementation would provide substantially better performance and scalability.
We present FlashSim through a simplified demonstrator that illustrates these operations and motivates discussion with the ROOT community on possible solutions and future directions.