Fixing the Measurement Model Before the Selection Heuristics: Optimiser Repairs for Living-Cell Self-Driving Laboratories

A measurement-aware Bayesian optimisation model for simulated living-cell self-driving laboratories reduced final simple regret versus plain GP-EI by 0.042 under heavy tails and by 0.039 with all pathologies.

• Why it matters: Closed-loop Bayesian optimisation breaks under living-cell constraints like one-time measurements, culture drift, batch offsets, heavy-tailed cell variation, and probe failures, so selection-only repairs may be insufficient.
• What they did: In a simulated living-cell laboratory (8,820 campaigns of 30 rounds; five destroyed cells per round; 60 seeds), the authors compared two surrogate models with four selection repairs across three response surfaces and four pathology regimes.
• The result: The measurement-aware model improved 90% band coverage under drift from 0.83 to 0.95 and beat GP-EI, while selection-level repairs did not help once data were contaminated; integrated NEI paired with the measurement-aware model gave the lowest regret.