Neural Integral Operators for Inverse problems in Spectroscopy

Emanuele Zappala, Alice Giola, Andreas Krämer, Enrico GrecoarXiv 2025 · DOI · arXiv

Abstract

Deep learning has shown high performance on spectroscopic inverse problems when sufficient data is available. However, it is often the case that data in spectroscopy is scarce, and this usually causes severe overfitting problems with deep learning methods. Traditional machine learning methods are viable when datasets are smaller, but the accuracy and applicability of these methods is generally more limited. We introduce a deep learning method for classification of molecular spectra based on learning integral operators via integral equations of the first kind, which results in an algorithm that is less affected by overfitting issues on small datasets, compared to other deep learning models. The problem formulation of the deep learning approach is based on inverse problems, which have traditionally found important applications in spectroscopy. We perform experiments on real world data to showcase our algorithm. It is seen that the model outperforms traditional machine learning approaches such as decision tree and support vector machine, and for small datasets it outperforms other deep learning models. Therefore, our methodology leverages the power of deep learning, still maintaining the performance when the available data is very limited, which is one of the main issues that deep learning faces in spectroscopy, where datasets are often times of small size.

Introduction

Spectroscopy is a foundational technique across chemistry, forensics, and food science used to identify substances from their spectral signatures. Deducing the underlying substance, molecular fingerprint, or concentration from a raw spectrum is a classic "inverse problem." While Deep Learning (DL) has revolutionized data analysis, it is notoriously data-hungry. In spectroscopy, datasets are often very small (e.g., just a few hundred samples) due to the cost and time required to collect physical samples.

When applied to these small datasets, standard DL models severely overfit, memorizing the training data but failing on new samples. Traditional machine learning models (like Support Vector Machines or Decision Trees) handle small data better but lack the expressive power to capture highly complex, non-linear spectral relationships. This research bridges this gap by designing a DL architecture inherently regularized against small-data overfitting.

Methodology: Neural Integral Operators

The proposed solution frames the spectroscopic inverse problem as learning an integral operator. Specifically, the model solves an integral equation of the first kind: $\int G_\theta(u(t), \sigma, t)dt = f(\sigma)$, where $f$ contains the target label information, and $G_\theta$ is a "Urysohn kernel" parameterized by a feed-forward neural network.

Schematic of the Neural Integral Operator framework for spectral classification.

The architecture consists of an encoder neural network ($E_\theta$) that compresses the input spectrum into a continuous function representation $u$. The kernel network ($G_\theta$) then takes $u$ and integration variables to compute the integral. Numerical integration is performed using Monte Carlo sampling. This integration step acts as a powerful, natural defense against overfitting; by sampling many random points during every forward pass, it effectively provides continuous data augmentation.

Experimental Setup and Results

The model was evaluated on three real-world datasets of varying sizes: Fruit Purees (983 spectra), Meat Samples (120 spectra), and Textile Samples (221 spectra). The Neural Integral Operator consistently outperformed or matched baselines across all data regimes.

Comparison of classification metrics on the Fruit Purees dataset, showing the Integral Operator matching or exceeding state-of-the-art DL models.

The advantage became even more pronounced on the smaller datasets. On the Meat dataset (120 samples), standard Feed-Forward Neural Networks (FFNN) collapsed to 38.5% accuracy due to overfitting, while the Integral Operator achieved 95.7% accuracy. Similarly, on the high-dimensional Textile dataset, the Integral Operator reached 92.2% accuracy, significantly outperforming CNNs (82.4%) and FFNNs (37.3%).

A key innovation is the use of asymmetric compute for stability. By using more sampling points during inference (5,000) than during training (2,000), the model gains stability at test time without increasing the computational burden of gradient calculations during training.

Conclusion

Neural integral operators offer a robust alternative to standard deep learning for spectroscopic analysis, particularly when data is scarce. By leveraging the mathematical structure of inverse problems and the regularizing properties of numerical integration, this approach maintains high accuracy where traditional DL models fail. Future work will explore the application of this framework to other domains involving inverse problems, such as medical diagnostics and materials science.

Summary updated

Abstract

Deep learning has shown high performance on spectroscopic inverse problems when sufficient data is available. However, it is often the case that data in spectroscopy is scarce, and this usually causes severe overfitting problems with deep learning methods. Traditional machine learning methods are viable when datasets are smaller, but the accuracy and applicability of these methods is generally more limited. We introduce a deep learning method for classification of molecular spectra based on learning integral operators via integral equations of the first kind, which results in an algorithm that is less affected by overfitting issues on small datasets, compared to other deep learning models. The problem formulation of the deep learning approach is based on inverse problems, which have traditionally found important applications in spectroscopy. We perform experiments on real world data to showcase our algorithm. It is seen that the model outperforms traditional machine learning approaches such as decision tree and support vector machine, and for small datasets it outperforms other deep learning models. Therefore, our methodology leverages the power of deep learning, still maintaining the performance when the available data is very limited, which is one of the main issues that deep learning faces in spectroscopy, where datasets are often times of small size.
BibTeX
@misc{zappala2025neural,
  author = {Emanuele Zappala and Alice Giola and Andreas Krämer and Enrico Greco},
  title = {{Neural Integral Operators for Inverse problems in Spectroscopy}},
  year = {2025},
  eprint = {2505.03677},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.03677},
  howpublished = {arXiv},
}
Figure