[3C01-05] Data-Driven Surface Science 

9:00 AM - 9:30 AM JST(12:00 AM - 12:30 AM UTC)

[3C01【依頼講演】] Recent progress of statisital peak fitting via EM algorithm: rust acceleration with AI coding

*Yasunobu Ando¹ (1. Institute of Science Tokyo (Japan))

EM-based spectral peak fitting is key for high-throughput spectroscopic analysis (XPS, Raman), but optimizing complex functions poses computational and numerical challenges. We leveraged Claude Code, an AI coding agent, to migrate the 'EMPeaks' library to a Rust backend with robust safeguards. Over 23 development days in June 2026, we executed five phases: (1) establishing the Rust backend and porting peak models (Lorentzian, PseudoVoigt, TSDC, VoigtMixtureModel) via PyO3; (2) integrating B-spline backgrounds; (3) implementing modular priors and Shirley/Tougaard mixtures; (4) resolving numerical issues (L-BFGS-B flat-region line search traps and Shirley float64 catastrophic cancellation); and (5) implementing M-step safeguards Delta Q criterion, All-Fail Restart, active component masking, and EFIM-preconditioned GEM fallback). This human-AI co-development successfully turned EMPeaks into a highly robust, fast, and parallelized computation engine.

Password required to view