Microchannel gas‑liquid sulfonation is an efficient route for synthesizing high‑performance surfactants such as sodium α‑olefin sulfonate (AOS). However, real‑time yield monitoring is challenging due to complex spatio‑temporal dynamics and the high cost of offline HPLC analysis. In a study published in Frontiers of Chemical Science and Engineering, researchers propose a non‑invasive soft measurement method using ConvLSTM to predict yield levels directly from microreactor image sequences.
The experimental system consists of a cross‑shaped glass microreactor chip (main channels 1 mm × 0.26 mm). Gaseous SO₃ and liquid 1‑dodecene react under varying conditions. A high‑speed camera (2000 fps) captures 1‑second video clips (≈2000 frames) every 50 seconds. Product yields (range 53 % to 85 %) are measured by HPLC. The raw data set contains only 32 experimental runs – a severe data scarcity challenge.
To address this, the team developed a frame‑sampling spatio‑temporal augmentation strategy. Due to high gas velocity (≥12.8 m·s⁻¹) and intense interfacial dynamics, consecutive frames show distinctly different visual details. Four frames are uniformly sampled from each 1‑second video (positions 400, 800, 1200, and 1600). For each selected index, the corresponding frames from all 20 video clips of the same experiment are concatenated to form a new 20‑step image sequence. This expands the original 32 runs into 156 labeled samples without additional experiments. After cropping, resizing to 200×60 pixels, and normalization, the data set is used for classification.
Yield values are divided into three levels using quartile thresholds: High (79–85 %, top 20 %), Medium (69–79 %), and Low (53–69 %). The ConvLSTM network – which replaces fully connected gates with convolutional operations – extracts both spatial features and temporal dependencies. A TimeDistributed module applies convolution independently to each frame before temporal integration. Five‑fold cross‑validation was used.
The augmented ConvLSTM model achieved an average accuracy of 97.44 % across folds. The same model without augmentation reached only 77.78 % (a gain of 19.66 percentage points). A conventional CNN (processing frames independently) achieved only 87.50 % on the augmented data – 9.94 % lower than ConvLSTM. The optimal input sequence length was 12 frames (≈240 ms). Inference time per prediction is about 75 ms on a standard GPU, far faster than offline HPLC analysis (hours). The confusion matrix shows perfect classification for the High yield class (precision = recall = 1.00) and strong performance for Medium and Low classes.
This work demonstrates that combining frame‑sampling augmentation with ConvLSTM effectively overcomes data scarcity and captures critical spatio‑temporal dynamics, offering a robust soft measurement tool for real‑time monitoring of microchannel gas‑liquid sulfonation. The approach is transferable to other microreactor multiphase systems.
DOI
10.1007/s11705-026-2636-8