(583b) Towards Real-Time Inference of Thin Liquid Film Thickness Profiles from Interference Patterns Using Transformer Models.

Measuring the thickness of thin liquid films is vital for mechanistically understanding a range of colloidal phenomena, from the stability of foams and emulsions to the behavior of coatings like human tear film. White light interferometry is a powerful technique that can be used to non-invasively measure thin film thickness with high resolution in both space and time. Unfortunately, recovering film thickness from white light interferograms is complicated by the transcendental phase-periodic governing equations that non-uniquely relate intensities in interferograms to film thicknesses. Overcoming this ambiguity has traditionally required confining the analysis range to a few hundred nanometers or relying on time-consuming manual interpretation, hindering wider adoption of the technique.

In this work, we develop a scalable solution to this problem leveraging advances in transformer models. Specifically, we engineer vision transformer networks with a U-Net architecture capable of end-to-end mapping of interferograms to thickness profiles. To train these data-intensive networks, we generate very large synthetic interferogram-thickness profile pairs utilizing known physics governing thin film structure and evolution. Preliminary evidence suggests that these models can recover thickness from interferograms agnostic of their source with an accuracy comparable to existing manual annotation methods. The film thickness recovery time per frame is less than 2 seconds on a conventional CPU, which is a 300-fold speedup compared to the existing manual method. Deploying this model on GPUs with further improvements promises real-time inference of film thickness and has the potential to accelerate the translation of thin film interferometry-based methods to industrial and clinical applications.