Artificial Neural Network–Based Prediction of Instability‐Prone Operating States in Routine HPLC Systems Using Pressure and Reinjection Data

ABSTRACT

Typically, there is inherent vulnerability to gradual, nondestructive operational instability in routine high‐performance liquid chromatography (HPLC) systems under continuous pharmaceutical production. Although traditional approaches to monitoring involve either past‐due alerts to failures or binary thresholds, all of them are unable to identify the onset of multivariate process instability. In this paper, an intelligent solution driven by an artificial neural network (ANN) is proposed for forecasting operational instability (in terms of stable, warning, and unstable conditions) based on routine HPLC parameters including column pressure, mobile phase pH, organic modifier composition, flow rate, analytical production load, and reinjection correction history. To overcome the limitations of pure modeling approaches, this study explicitly links operational data to underlying physical degradation mechanisms such as localized hydrodynamic friction, stationary phase mechanical collapse, and mass transfer kinetics. For modeling the complex nonlinear operational process dynamics and improving interpretability, Explainable artificial intelligence (XAI) by means of permutation feature importance was considered. The refined ANN model showed a very good ability to classify data points on the testing dataset in an independent test setting with an overall accuracy rate of 92.6%, an F1 score of 0.91, and a multiclass area under the receiver operating characteristic curve (ROC‐AUC) score of 0.97, surpassing the gradient boosting, random forest, and support vector machine (SVM) models, as well as a traditional multinomial logistic regression mathematical model. According to the results of the XAI analysis, the most important predictors of system degradation included column pressure differential (Δ P ), process protocol (mobile phase aging), and the number of historical correction injections performed. This result proves the correctness of the paradigm shift in the use of historical correction injections to predict future system failures. The presented approach demonstrates a successful transition from reactive problem solving to proactive intelligent system prediction.