SlitNET: A Deep Learning Enabled Spectrometer Slit

Abstract

The efficiency and resolution of dispersive spectrometers play crucial roles in optical spectroscopy. Achieving optimal analytical performance in optical spectroscopy requires striking a delicate balance between employing a narrow spectrometer input slit to enhance spectral resolution while sacrificing throughput or utilizing a wider slit to increase throughput at the expense of resolution. Here, we introduce a spectrometer slit empowered by a deep learning model SlitNET. We trained a neural network to reconstruct synthetic Raman spectra with enhanced resolution from low-resolution inputs. Subsequently, we performed transfer learning from synthetic data to experimental Raman data of materials. By fine-tuning the model with experimental data, we recovered high-resolution Raman spectra. This enhancement enabled us to distinguish between materials that were previously indistinguishable when using a wide slit. SlitNET achieved a resolution enhancement equivalent to employing a 10 μm slit size but with a physical input slit of 100 μm. This, in turn, enables us to simultaneously achieve high throughput and resolution, thereby enhancing the analytic sensitivity and specificity in optical spectroscopy. The incorporation of deep learning into spectrometers highlights the convergence of photonic instrumentation and artificial intelligence, offering improved measurement accuracy across various optical spectroscopy applications.

Figure 1

SlitNET deep learning architecture for spectral resolution enhancement. (A) Schematic representation of the ResUNet architecture used in this study. The ResUNet combines residual connections and U-Net-like skip connections to enable effective information flow and feature extraction, facilitating the reconstruction of high-resolution spectra from low-resolution inputs. (B) Random delta function generation and intensities are used for synthetic Raman data generation. These were convolved with Lorenzians to create natural peak broadening. A pair of high-resolution target spectra and low-resolution input spectra were generated by convolving the spectra with different Gaussian functions to simulate a narrow and broad slit. A total 100,000 pairs of synthetic Raman spectra were generated. (C) The workflow developed in this study involves training the model on synthetic Raman data and applying transfer learning to a limited set of experimental spectra to further enhance the model’s performance, accounting for complex spectrometer aberrations.

Figure 2

SlitNET application to synthetic Raman data. (A) A comparison of the low-resolution input spectra (black), their corresponding target high-resolution spectra (blue), and the spectra predicted by the deep learning-enabled spectrometer slit (tangerine). The predicted spectrum more closely resembles the target high-resolution spectrum, indicating the model’s ability for resolution enhancement. (B) Training and validation learning curves during model training. The graph shows the mean absolute error (MAE) plotted against the number of training epochs. (C) The full width at half-maximum (fwhm) indicates that the model can improve the resolution of synthetic spectra (n = 5 isolated peaks). (D) Comparison of peak intensities showing less than 2% deviation compared to high-resolution target spectra (n = 5 isolated peaks).

Figure 3

SlitNET transfer learning application to experimental Raman data. (A) Application of synthetic SlitNET model to real Raman spectrum of polystyrene before transfer learning (TL). This illustrates that the model’s capability can enhance the resolution of the spectra but that it overpredicts the resolution for most peaks. (B) TL application significantly improves predictive capability, resulting in spectra closely resembling the high-resolution target spectrum (blue) in terms of both resolution and intensity. (C) Comparison of the full width at half-maximum (fwhm) of Raman peaks relative to the target (n = 5 isolated peaks). These data shows that Post-TL the SlitNET model enhances the resolution. (D) Relative intensity in the spectra before and after TL. (E) Residual spectra between Pre- and Post-TL models and the target. The residual spectra show that the Post-TL model more accurately replicate the target. (F) Correlation coefficients of Pre-TL model (0.92) and Post-TL model (0.98) with respect to the target.

Figure 4

Application of the SlitNET model for identification of two different materials. (A) High resolution Raman spectra of l-Arginine and urea using a narrow slit (10 μm). These spectra show two close peaks located near 970 and 1003 cm^–1. (B) Low resolution Raman spectra of l-Arginine and urea obtained using a wide slit (100 μm). (C) Wide slit input (100 μm) and target spectra (10 μm) of the superposition of the spectra from the two materials. The predicted spectrum from the input (depicted in tangerine) clearly resolves the overlapping peaks near 970 and 1003 cm^–1.