Enhancing UV Spectral Prediction through Auxiliary Task, Curriculum Learning, and Curvature Limitation

Hajime Shinohara, Akihiro Kishimoto, Hiroshi Kajino

Keywords: UV spectral prediction, auxiliary task learning, curriculum learning, spectroscopy, machine learning, peak position awareness, spectral generation, chemical property prediction

TL;DR: We propose three methods (PPA, CLIAS, SCL) that explicitly handle peak positions, band shapes, and curvature profiles to improve UV spectral prediction accuracy by 14% over baselines.

Abstract: Accurate UV spectral prediction is challenging for machine learning. UV spectra exhibit broad absorption bands characterized by the peak positions, band shapes, and curvature profiles. However, current models fail to capture these characteristics. We present Peak Position Awareness (PPA), Curriculum Learning for Interpolated Abstracted Spectra (CLIAS), and Spectrum Curvature Limitation (SCL) to handle the above characteristics, showing consistent improvements over diverse models.

Submission Track: Paper Track (Short Paper)

Submission Category: Automated Material Characterization

Institution Location: Tokyo

AI4Mat RLSF: Yes

Submission Number: 20