Transfer-learning enhanced adaptive sampling for accelerating ultrafast spectroscopy

Abstract

Ultrafast transient absorption (TA) spectroscopy is a versatile tool for probing photoinduced dynamics in complex materials, but it often requires dense temporal sampling and extensive signal averaging, resulting in lengthy data acquisition times. Here, we present a data-driven sampling method, called Transfer-Learning Enhanced Adaptive Sampling (TEAS), which significantly accelerates TA measurements by reducing the number of required time points while preserving the full spectral and temporal information content of the original data. TEAS combines transfer learning with adaptive sampling to exploit cross-wavelength patterns, concentrating measurements on the most informative regions for greater efficiency and accuracy. The method does not rely on any specific mathematical or kinetic model, making it a flexible and general framework for accelerating measurements across a wide range of spectroscopic modalities. We demonstrate that this method can accurately reconstruct TA data using less than 1% of the total experimental measurements under varying signal-to-noise conditions, consistently outperforming traditional approaches that lack transfer learning or adaptive sampling. These results highlight the potential of TEAS as a broadly applicable, cost-effective solution for speeding up ultrafast spectroscopy and enabling real-time, data-driven experimentation.