Machine Learning from Advanced Nano-Optical Imaging

The scattering-type scanning near-field optical microscopy(s-SNOM) technique has recently spread to many research areas and enabled notable discoveries in the field of quantum materials, extraterrestrial particles, catalytic substances, and biological samples. Various forms of nano-optical measurement, such as monochromatic nano-imaging, broadband nano-spectroscopy, time-resolved pump-probe spectroscopy and hyperspectral imaging produce terabytes of image-like, spectrum-like and higher dimensional data in a daily basis. In our group, we apply and develop advanced Machine learning algorithm for faster data acquisition, more quantitative data interpretation and automated data collection for s-SNOM. With the assistance of AI- and ML-enhanced data acquisition and analysis, scanning probe optical nanosopy is poised to become more efficient, accurate and intelligent.

In collaboration with Mengkun Liu (Stony Brook University)

Recent review articles: Coming Soon

Recent results:

Machine Learning for Optical Scanning Probe Nanoscopy
Chen, Xu et al. Advanced Material, 2022, 2109171 (2022). Ref [332]

Read More

The optical scanning probe microscopy enables us to study the behavior of quantum material with unprecedented spatial resolution. The versatility of s-SNOM is endowed by combining near-field optics with synchrotron radiation, ultrafast optics and other experimental probes. However, the high versatility of s-SNOM would lead to the increasing complexity in data analysis and data acquisition process.  In this perspective, we discuss how s-SNOM can benefit from various machine learning algorithms. The supervised learning algorithms enable fast material properties extraction based on nano-spectroscopy or polaritonic images, while the unsupervised learning algorithms, autoencoder for instance, provides ways to abstract rules and laws in the raw experimental data. Moreover, the experiment atomization can be realized by reinforcement learning and the combinations the various of statistical models and machine learning algorithms. The proposed Reinforced & Advanced learning s-SNOM(REAL-SNOM) could partially or fully automate the data acquisition, data processing and find new physical rules in the optical scanning probe microscopy data collected on quantum materials.

Deep Learning Analysis of Polaritonic Wave Images
Xu et al. Acs Nano, 15, 18182 (2021). Ref. [317]

Read More

Scattering-type scanning near-field optical microscope(s-SNOM) can excite and probe propagating modes (plasmon polariton, phonon polariton, exciton polariton, etc.) in the quantum materials. By extracting the wavelengths and the quality factors of the propagating modes, one can infer the local optical and electronic properties. The traditional analysis of polaritonic images requires intensive manual processing and fitting of the interference pattern. In this work, we showed that the convolutional neural network (CNN) enables regression on images in a time scale that is at least 3 orders of magnitude faster than common fitting/processing procedures. This protocol can also be adapted to other scanning probe microscopies dataset and applied for parameter extractions on more complicated nano-scale features.

Hybrid Machine Learning for Scanning Near-Field Optical Spectroscopy
Chen et al. Acs Photonics 8, 2987 (2021). Ref. [312]

Read More

In the scanning probe microscopy techniques, the local optical properties of samples are probed by measuring the scattering signal from the tip. However, the tip-sample electromagnetic interaction is non-trivial, and it is always a challenge to get an analytical and accurate model without approximation. As a powerful tool to establish complex correlation between data, the artificial neural network (ANN) is proved to be an alternative approach for understanding the tip-sample interaction and provide a one-to-one correspondence between scattering signal and the local optical conductivity. In this work, we also show that the physics-infused hybrid neural network, enabled by training dataset combined with both modeled data and experimental data, can predict the tip the probe-sample interaction with better accuracy and generalization.