Physics informed machine learning for material identification: initial steps

Machine learning methods are rapidly improving in the field of gamma spectroscopy, demonstrating the remarkable ability to classify complex spectra. However, most existing literature trains classical machine learning architectures on internal datasets, abstracting away the known physical laws surrounding photon emission, interaction, and detection. This analysis considers AI from a different angle – how can we maintain sight of the physics? We explore the broad field of physics informed machine learning for material identification and detection and investigate different opportunities to embed physics within the traditional machine learning workflow. For example, we may be able to achieve improved classification accuracy by pretraining the model on a massive quantity of cheap, low fidelity, physically derived synthetic training data. Results of this analysis will present a new framework for generalizable and interpretable machine learning models for material identification.