Robustness Testing is the systematic evaluation of whether a model maintains accurate predictions when inputs are perturbed, corrupted, or shifted — measuring the model's stability and reliability under realistic variations that it will encounter in production.

Robustness Test Categories

• Input Perturbation: Small changes to inputs (noise, rounding, sensor drift) should not change predictions significantly.
• Corruption: Missing values, outliers, and sensor failures should be handled gracefully.
• Distribution Shift: Performance on data from different tools, time periods, or process conditions.
• Adversarial: Worst-case perturbations that maximally degrade model performance.

Why It Matters

• Reliability: A model that fails with minor input perturbations is unreliable for production use.
• Sensor Noise: Real-world fab data always contains noise — robustness to noise is essential.
• Confidence: Robustness testing builds confidence that the model will perform well under real operating conditions.

Robustness Testing is testing for the real world — verifying that models maintain performance amid the noise, drift, and variations of production.