Principles-Driven Machine Learning for UV Spectral Prediction

Hajime Shinohara, Akihiro Kishimoto, Hiroshi Kajino

Keywords: machine learning, spectral prediction, curriculum learning, physical constraints, UV spectroscopy

TL;DR: We propose three physics-inspired methods (PPA, CLIAS, SCL) that capture UV spectral characteristics, achieving 10-22% improvement over baselines and outperforming state-of-the-art UV-adVISor.

Abstract:

Despite recent advances in machine learning, accurate UV spectral prediction remains a challenging task. UV spectra exhibit inherently broad absorption bands characterized by peak positions, band shapes, and curvature profiles that machine learning models still struggle to capture effectively.
Based on the ideas behind UV spectroscopic principles and analytical techniques, we present three methods that aims at identifying these three characteristics: Peak Position Awareness (PPA), Curriculum Learning for Interpolated Abstracted Spectra (CLIAS), and Spectrum Curvature Limitation (SCL).
Our performance evaluation shows that our methods can successfully capture genuine spectral characteristics, achieving consistent improvements over diverse models, especially when the training order of CLIAS and SCL is carefully considered in combination with PPA. We also show important evidence that our best models outperform the state-of-the-art UV-adVISor model.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 10056