Physics-informed ML models for processing of spectroscopic data

Physics-informed ML models for processing of spectroscopic data

Abstrakt

Massive adoption of machine learning (ML) techniques in spectroscopy brought entirely new possibilities in analytical performance for applications, and also for basic research. However, several problems emerged, e.g. ML models are often utilized as “black-boxes”, or considerably overtrained. Another issue is a blind transition of successful models (architecture, parameters) from distinct applications (e.g. image processing) to spectroscopic tasks, without taking into account the properties of data. We study the influence of (spectroscopic) data properties and incorporate them into ML models in form of weight initializations, specific parameter penalizations, and invariances. This leads to an increased analytical performance of models and better interpretability

Massive adoption of machine learning (ML) techniques in spectroscopy brought entirely new possibilities in analytical performance for applications, and also for basic research. However, several problems emerged, e.g. ML models are often utilized as “black-boxes”, or considerably overtrained. Another issue is a blind transition of successful models (architecture, parameters) from distinct applications (e.g. image processing) to spectroscopic tasks, without taking into account the properties of data. We study the influence of (spectroscopic) data properties and incorporate them into ML models in form of weight initializations, specific parameter penalizations, and invariances. This leads to an increased analytical performance of models and better interpretability

machine learning, spectroscopic data, interpretability, physics-informed, inductive bias

machine learning, spectroscopic data, interpretability, physics-informed, inductive bias

VRÁBEL, J.

2022

24.08.2021