Real-time concentration prediction in column chromatography purification using NIR optical sensing and evolutionary attention-LSTM modeling

Ginsenosides, the principal bioactive constituents of American ginseng ( Panax quinquefolius ), require precise monitoring during column chromatography to ensure product quality. Conventional assays are time-consuming and unsuited for real-time process control, particularly for low-abundance compounds. Here, we present a rapid, non-destructive near-infrared (NIR) optical sensing system coupled with an Evolutionary Attention–based Long Short-Term Memory (EA-LSTM) model to predict ginsenoside concentrations dynamically. The approach exploits the high-speed acquisition and reproducibility of NIR spectra while leveraging EA-LSTM's temporal feature extraction and adaptive attention to focus on chemically informative elution segments. Offline spectra from eight elution batches, referenced by UPLC, enabled model training and validation, with high-frequency sampling simulated to mimic industrial deployment. The system demonstrated high accuracy (R 2  = 0.9841, RPD = 7.94) and robustness under dynamic conditions (R 2  = 0.9694, RPD = 5.72, response time = 0.34 ms), highlighting its potential for real-time, non-destructive monitoring and intelligent control in natural product purification.