SMARTERS paper shows simulated-spectrum models collapse on real experimental frequency ranges

bravo_abad · x · 2026-09-26

An empirical observation from the SMARTERS paper (Sethi et al.) exposes a hidden pitfall of training scientific ML on simulated data.

The paper's attention U-Net reconstructs atomic positions from images of molecular vibrations. Given simulated spectra covering 0–3200 cm⁻¹, it recovers much of iron phthalocyanine's outer structure. Restrict the inputs to the experimentally measurable range of 300–2000 cm⁻¹, and the reconstruction deteriorates sharply. The molecule is unchanged — the available information is not.

This doesn't mean the narrow range can never work with a purpose-trained model; it shows why successful synthetic-data tests can mislead when training and measurement cover different ranges. Practical lesson: when training on simulations, match the simulation to the instrument's actual measurement window.

Original post →

More from Research

Research channel →