Understanding the patterns that neural networks learn from chemical spectra

Laura Hannemose Rieger, Max Wilson, Tejs Vegge, Eibar Flores*

*Corresponding author for this work

Research output: Contribution to journalJournal articleResearchpeer-review

3 Downloads (Pure)

Abstract

Analysing spectra from experimental characterization of materials is time consuming, susceptible to distortions in data, requires specific domain knowledge, and may be susceptible to biases in general heuristics under human analysis. Recent work has shown the potential of using neural networks to solve this task, and assist spectral interpretation with automated and unbiased analysis on-the-fly. However, the black-box nature of most neural networks poses challenges when interpreting which patterns from the data are being used to make predictions. Understanding how neural networks learn is critical to assess their accuracy on unseen data, justify critical decision-making based on predictions, and potentially unravel meaningful scientific insights. We present a 1D neural network to classify infrared spectra from small organic molecules according to their functional groups. Our model is within range of state-of-the-art performance while being significantly less complex than previously used networks reported in the literature. A smaller network reduces the risk of overfitting and enables exploring what the model has learned about the patterns in the spectra that relate to molecular structure and composition. With a novel two-step approach for explaining the neural network's classification process, our findings not only demonstrate that the model learns the characteristic group frequencies of functional groups, but also suggest it uses non-intuitive patterns such as tails and overtones when classifying spectra.
Original languageEnglish
JournalDigital Discovery
Volume2
Issue number6
Pages (from-to)1957-1968
Number of pages12
ISSN2635-098x
DOIs
Publication statusPublished - 2023

Fingerprint

Dive into the research topics of 'Understanding the patterns that neural networks learn from chemical spectra'. Together they form a unique fingerprint.

Cite this