Workshop: XLOOP 2026: The 8th Annual Workshop on Extreme-Scale Experiment-in-the-Loop Computing
Authors: Shunyu Yao, K. Todd Lowe, and Dimitrios Nikolopoulos (Virginia Tech)
Abstract: Experimental aerodynamics facilities stream surface-pressure measurements continuously, but the CFD models used to interpret those measurements require minutes to hours per evaluation. This multi-order-of-magnitude gap prevents deploying the model inside the experiment’s control loop. We present DART, a workflow engine that represents a CFD digital-twin pipeline as a directed acyclic graph of typed computational steps and makes surrogate substitution a first-class operation: a machine-learned model replaces a node’s implementation while preserving its declared input/output contract, so adjacent stages remain unchanged. DART then monitors the substituted loop by comparing prediction residuals to incoming measurements and raises an alarm when the surrogate leaves its fitted regime. On a steady aerodynamic pipeline, a substituted surrogate answers in 4.1 ms what the solver answers in 220s while reproducing 99.5% of held-out lift variation. On an unsteady cylinder benchmark, an alarm calibrated only on trusted-regime data detects out-of-envelope deployment within six shedding cycles.
Back to XLOOP 2026: The 8th Annual Workshop on Extreme-Scale Experiment-in-the-Loop Computing Archive Listing
Back to Full Workshop Archive Listing