Lactose Crystallization: Integrating Machine Learning with Process Analytical Technologies

Lactose is recovered from whey through crystallization process, where a concentrated supersaturated solution is cooled to crystallize the lactose, leaving the impurities in the mother liquor. Designing this process requires considerations over various parameters, particularly the concentration of the feed solution and the cooling profile. To optimize the parameters, most developers depend on trial-and-error methods, a manageable task for the experienced but challenging for novices. This study presents a novel system that leverages machine learning (ML) and process analytical technologies (PAT) to streamline lactose crystallization process development, going beyond manual trial and error interpretations.

Highlights

Automated crystallization platform integrating ML and PAT for lactose recovery.

Utilized Adaptive Neuro-Fuzzy Inference System (ANFIS) for dynamic process adjustment.

Achieved quantitative enhancements in lactose crystal size and yield productivity.

AI-driven control approach outperforms conventional crystallization methods

The automated system initiated with Direct Chord Length (DCL) feedback control run, which provided the foundational data for the ML model, which was then employed in subsequent AN1 and AN2 iterative runs. These iterative runs have smoother concentration and temperature curves, and it generates larger crystal with enhanced productivity and yield. The results indicate that the ML-driven approach can significantly outperform conventional methods, enabling the precise control of nucleation and growth phases to produce larger lactose crystals.

Experimental preparations

Lactose monohydrate, FlowLac® 100 (Meggle GmbH & Co. KG, Germany), was used for all experiments. To accurately measure solution concentration in terms of lactose monohydrate per gram of solution, the React-IR needs to be calibrated to convert raw spectrum into solution concentration. The IR signal was calibrated at varying concentrations (0.63 – 0.1 g lactose/g solution) and temperature (5 – 95 ^oC). Out of all the signals collected, 85% of the data was (randomly) used for calibration and the remaining 15% for validation. The baseline of the IR spectrum was first corrected using Savitzky-Golay second derivative. A Partial Least Square (PLS) calibration model was developed in Matlab® software using the calibration data collected. The prediction error % is 1.17% and 1.18% for both training and validation datasets, respectively. Some of the crystals were isolated from the crystallizer using vacuum filtration. The particle size distribution of isolated lactose crystals was measured with Malvern Mastersizer 3000.

See and download the full research paper (pre-proof): Lactose Crystallization – Integrating Machine Learning with Process Analytical Technologies

or continue here

Cha Yong Jong, Akshay Mittal, Felix Lee Jun Jie, Lee May Loo, Goh Yongkai, Qiaolin Yuan, Eunice Yeap Wan Qi, Srinivas Reddy Dubbaka, Harsha Nagesh Rao, Wong Shin Yee,
Lactose Crystallization: Integrating Machine Learning with Process Analytical Technologies,
Food and Bioproducts Processing, 2025, ISSN 0960-3085,
https://doi.org/10.1016/j.fbp.2025.02.008.

Tags: excipients

Lactose Crystallization: Integrating Machine Learning with Process Analytical Technologies

Related Posts

Hot-melt extruded ibuprofen ternary solid dispersions using in-line UV–Vis: Impact of an ionizable polymer on thermodynamics and dissolution

Roller compaction: Size distribution during different stages of the product pathway using three types of lactose

Capsugel® Enprotect® Capsule – First of its kind enteric technology

3D Printing of Tailored Veterinary Dual-Release Tablets: A Semi- Solid Extrusion Approach for Metoclopramide

Cart

Shop Search

Top Pharma-Excipient Links

Welcome Back!

Retrieve your password