Digital Twins in Action: AI, HPC, and End-to-End Workflows at Scale

In this invited talk, Anuj Kapadia of Oak Ridge National Laboratory, section head for advanced computing and health sciences, shares practical lessons from building and scaling human digital twins for cancer applications. He stresses that a digital twin is a systems problem, not merely a model, combining data, physics-based simulation, AI, workflows, validation, and computing infrastructure across multiple biological scales. Starting from Duke's virtual imaging trial XCAT phantoms, his team added DNA, chromatin structure, and cellular response mechanisms, coupling whole-body, multicellular, and subcellular scales into a single mechanistic simulation blending Monte Carlo and AI. He illustrates with three efforts: a population-scale virtual imaging trial that generated over 10,000 abdominal CT scans (125 terabytes) in two weeks on Summit, a 150-fold speedup over single-node execution; radiopharmaceutical distribution and DNA-damage studies comparing treatments such as lutetium versus yttrium; and extensive parameter sweeps requiring uncertainty quantification. Kapadia distills seven lessons: data becomes the bottleneck before compute; optimize the workflow rather than every line of code, prioritizing scientific over computational throughput; build quality control into the infrastructure; digital twins are fundamentally workflow and orchestration problems; integration is harder than modeling; validation against ground truth is a real bottleneck and the true source of trust; and AI and physics need each other, with HPC as enabler. His overarching message is that the future belongs to integrated AI-simulation ecosystems, not better individual models.