(467f) A Physics-Informed Model for Solution Composition Prediction from ATR-FTIR Spectrum Measurement

Understanding the dynamic change of the solute concentration in response to process operation parameters is critical to crystallization process design and control of the product quality. One approach for solute concentration analysis is through ATR-FTIR, a widely used PAT tool. However, it suffers from lengthy calibration process for tens of conditions, at different compositions and temperatures required by statistical models, such as principal component analysis or partial least square regression.

Herein, we reported a physics-informed modeling approach relying on fewer calibration points, thus significantly reducing the calibration efforts involved in PAT tools. This machine learning approach incorporates domain knowledge for feature design, Gaussian filtration for IR spectrum analysis to develop a predictive model for solution composition, for process relevant solvent compositions and temperatures. The results demonstrate an accuracy of 95% in the testing data sets on two model compounds. We anticipate our findings to promote the usage of PAT tools and to benefit process development by quickly deploying the calibration, especially for systems with limited information.