Simulation-driven machine learning approach for high-speed correction of slope-dependent error in coherence scanning...

Area of Science:

• Metrology
• Optical Engineering
• Machine Learning Applications

Background:

• Slope-dependent errors are a significant challenge in coherence scanning interferometry (CSI) for measuring complex engineering surfaces.
• Existing correction methods, like phase inversion of the 3D surface transfer function, are computationally intensive due to CSI instrument non-shift-invariance.

Purpose of the Study:

• To introduce a novel, efficient machine learning approach for correcting slope-dependent errors in CSI measurements.
• To significantly reduce the computational time required for error correction in CSI.

Main Methods:

• A deep neural network was trained using simulated surface measurements from a validated physics-based virtual CSI method.
• The network directly learns error characteristics to perform slope-dependent error correction.

Main Results:

• The trained deep neural network corrected a 1024x1024 surface height map within 0.1 seconds.
• The achieved accuracy is comparable to traditional phase inversion methods.
• The new method is two orders of magnitude faster than previous approaches under identical computational conditions.

Conclusions:

• Machine learning, specifically deep neural networks, offers a highly efficient solution for slope-dependent error correction in CSI.
• This approach dramatically improves measurement efficiency for functional engineering surfaces with complex geometries.