Bayesian Inference-enabled Precise Optical Wavelength Estimation using Transition Metal Dichalcogenide Thin Films

Davoud Hejazi, Shuangjun Liu, Sarah Ostadabbas, Swastik Kar

Abstract

Despite its ability to draw precise inferences from large and complex datasets, the use of data analytics in the field of condensed matter and materials sciences -- where vast quantities of complex metrology data are regularly generated -- has remained surprisingly limited. Specifically, such approaches could dramatically reduce the engineering complexities of devices that directly exploit the physical properties of materials. Here, we present a cyber-physical system for accurately estimating the wavelength of any monochromatic light in the range of 325-1100nm, by applying Bayesian inference on the optical transmittance data from a few low-cost, easy-to-fabricate thin film "filters" of layered transition metal dichalcogenides (TMDs) such as MoS2 and WS2. Wavelengths of tested monochromatic light could be estimated with only 1% estimation error over 99% of the stated spectral range, with lowest error values reaching as low as a few ten parts per million (ppm) in a system with only eleven filters. By step-wise elimination of filters with the least contribution toward accuracy, mean estimation accuracy of 99% could be obtained even in a two-filter system. Furthermore, we provide a statistical approach for selecting the best "filter" material for any intended spectral range based on the spectral variation of transmittance within the desired range of wavelengths. And finally, we demonstrate that calibrating the data-driven models for the filters from time to time overcomes the minor drifts in their transmittance values, which allows using the same filters indefinitely. This work not only enables the development of simple cyber-physical photodetectors with high accuracy color-estimation, but also provides a framework for developing similar cyber-physical systems with drastically reduced complexity.

Introduction

Traditional optical wavelength estimation typically requires bulky and expensive hardware, such as mechanical diffraction gratings or dense arrays of narrow-band "notch" filters. These systems are difficult to miniaturize for portable electronics and often ignore "off-target" data from filters that do not perfectly match the incident wavelength. Researchers at Northeastern University have addressed these limitations by developing a "cyber-physical" system that utilizes the broadband optical responses of 2D materials. By combining low-cost thin films with sophisticated Bayesian statistics, they have created a high-precision wavelength estimator that thrives on overlapping, "messy" data rather than requiring narrow spectral selectivity.

Methodology: The Cyber-Physical Framework

The system is divided into two distinct layers that work in tandem to triangulate the wavelength of incident light.

The Physical Layer

The physical component consists of 11 thin-film filters fabricated from Transition Metal Dichalcogenides (TMDs), specifically $MoS_2$ and $WS_2$. These materials were prepared using liquid-phase exfoliation—a cost-effective "top-down" method—and drop-casted onto glass substrates. By varying the concentration and mixture of these two materials, the researchers produced a set of filters with unique, overlapping transmittance spectra across the 325–1100 nm range.

Experimental setup showing (a) the physical TMD filters, (b) micrographs of the film surface, (c) the measurement schematic, and (d) the resulting broadband transmittance spectra.

The Cyber Layer

To interpret the data, the team employed Bayesian Inference. Each filter was characterized by measuring its transmittance 120 times at every wavelength to establish a baseline probability distribution (mean and standard deviation). When unknown monochromatic light passes through the filters, the system calculates the "posterior probability"—determining which wavelength most likely corresponds to the observed set of 11 transmittance readings. This approach allows the system to remain accurate even in the presence of physical noise and fluctuations.

Key Insights and Results

The synergy between material properties and data analytics yielded several significant findings:

  • High Precision and Accuracy: With the full 11-filter array, the system achieved an estimation error of less than 0.1% for 71% of the tested spectrum. In optimal regions, accuracy reached the tens of parts per million (ppm).
  • Efficiency through Sparsity: Using a greedy selection algorithm, the researchers discovered that just two filters (pure $MoS_2$ and pure $WS_2$) could maintain approximately 99% mean accuracy, as these materials provide the most independent signals.
  • Handling Material Drift: Nanomaterials often degrade over time, causing shifts in optical properties. The study demonstrated that periodic software recalibration—updating the model's statistical parameters—could compensate for six months of material drift without needing to replace the physical filters.
Statistical processing of transmittance data: (a-b) establishing the Gaussian distribution of noise at a specific wavelength, and (c) the resulting sharp Bayesian posterior probability used for precise estimation.

Conclusion and Future Outlook

This work proves that sophisticated data analytics can effectively replace complex hardware in optical sensing. While currently limited to monochromatic light sources, the framework provides a clear path toward low-cost, high-accuracy "lab-on-a-chip" devices. Future research will focus on integrating these TMD films directly onto CMOS photodetector chips and expanding the detectable spectral range into the ultraviolet and infrared regions by selecting alternative 2D materials.

Summary updated

Abstract

Despite its ability to draw precise inferences from large and complex datasets, the use of data analytics in the field of condensed matter and materials sciences -- where vast quantities of complex metrology data are regularly generated -- has remained surprisingly limited. Specifically, such approaches could dramatically reduce the engineering complexities of devices that directly exploit the physical properties of materials. Here, we present a cyber-physical system for accurately estimating the wavelength of any monochromatic light in the range of 325-1100nm, by applying Bayesian inference on the optical transmittance data from a few low-cost, easy-to-fabricate thin film "filters" of layered transition metal dichalcogenides (TMDs) such as MoS2 and WS2. Wavelengths of tested monochromatic light could be estimated with only 1% estimation error over 99% of the stated spectral range, with lowest error values reaching as low as a few ten parts per million (ppm) in a system with only eleven filters. By step-wise elimination of filters with the least contribution toward accuracy, mean estimation accuracy of 99% could be obtained even in a two-filter system. Furthermore, we provide a statistical approach for selecting the best "filter" material for any intended spectral range based on the spectral variation of transmittance within the desired range of wavelengths. And finally, we demonstrate that calibrating the data-driven models for the filters from time to time overcomes the minor drifts in their transmittance values, which allows using the same filters indefinitely. This work not only enables the development of simple cyber-physical photodetectors with high accuracy color-estimation, but also provides a framework for developing similar cyber-physical systems with drastically reduced complexity.
BibTeX
@inproceedings{hejazi2019bayesian,
  author = {Davoud Hejazi and Shuangjun Liu and Sarah Ostadabbas and Swastik Kar},
  title = {{Bayesian Inference-enabled Precise Optical Wavelength Estimation using Transition Metal Dichalcogenide Thin Films}},
  booktitle = {ACS},
  year = {2019},
  doi = {10.1021/acsanm.9b00489},
}
Figure