Fidelity Regret: Measuring Simulator Misspecification in Verifiable Design

Impact 6 arXiv cs.AI

Summary

  • The paper introduces fidelity regret to measure simulator misspecification in verifiable design workflows, such as genetic-circuit design.
  • Evaluating a 21,362-design combination subspace shows that mean normalized regret is 0.369 across 72 cells, with only 6 cells showing zero regret.
  • Workflows optimized using a cheap simulator or digital twin can easily satisfy top-level verifiers while performing poorly on physical objectives.
  • A budget-aware mitigation strategy re-scores top cheap-simulator designs to achieve a lower mean excess regret of 0.016 compared to cheap-only selection.

Why it matters

Fidelity regret provides a transferable template for detecting the gap between digital twin simulations and physical objectives, helping engineers mitigate simulator misspecification across scientific domains.

Who should care