Multimodal fusion network for micro-displacement measurement via Michelson interferometer

Single-wavelength Michelson interferometry offers high sensitivity but is limited by λ / 2 phase ambiguity and strong noise sensitivity, hindering practical deployment. Conventional iterative fitting methods for phase recovery are computationally expensive and prone to convergence failure under severe noise, limiting throughput and reliability. We present a multimodal fusion network (MFN) for quasi-absolute micro-displacement measurement that is pretrained on simulated interferograms and fine-tuned with only ∼ 500 real images, enabling rapid adaptation to experimental conditions with minimal calibration effort. MFN fuses complementary interferometric observables and uses a dual-head design (regression for sub- λ / 2 displacement and classification for integer-order) with an orthogonality regularizer to separate continuous and discrete components. Experimentally validated on a calibrated Michelson testbed, the method achieves nanometer-scale precision, strong robustness under severe mixed-noise conditions, an order-classification accuracy of 98%, and real-time inference of ∼ 10 ms per frame. This hardware-efficient approach resolves single-wavelength ambiguity without additional optical channels or scanning, offering a practical, fast, and robust solution for interferometric metrology.