Heuristic Scheduling Strategies for Single-Reactor Pharmaceutical Batch Production Under Uncertainty: A Comparative Statistical and Machine Learning Analysis.

Abstract Background: Pharmaceutical batch production faces significant scheduling challenges due to operational uncertainties including equipment failures, yield variability, and demand fluctuations. While scheduling heuristics are widely used in practice, their comparative performance under varying uncertainty conditions remains insufficiently characterized, particularly for single-reactor configurations common in specialty pharmaceutical manufacturing. Methods: A discrete-event simulation model was developed for a single 10,000L bioreactor producing three antibiotic products with fermentation times of 48, 72, and 120 hours. Three scheduling heuristics (FIFO, SPT, LPT) were evaluated across three uncertainty levels (Low, Medium, High) using 150 randomly generated demand scenarios, yielding 450 total observations. Statistical analyses included two-way factorial ANOVA, Kruskal-Wallis tests, multiple linear regression, and ANCOVA. Machine learning classification models (Random Forest, Gradient Boosting, SVM, Decision Tree, Logistic Regression) were trained to predict schedule robustness. Results: Both heuristic type (F = 225.71, p < .001, η² = 0.285) and uncertainty level (F = 346.69, p < .001, η² = 0.437) significantly affected makespan, with no interaction effect (p = .981). SPT achieved mean makespan of 1,992 hours compared to 2,387 hours for FIFO (16.5% improvement). However, SPT and LPT were not significantly different from each other (p = 1.000 after Bonferroni correction). Uncertainty increased makespan by 28.2% from low to high conditions. Machine learning models achieved 90-97% classification accuracy, with Polynomial SVM performing best (96.7% accuracy, AUC = 0.972). Feature importance analysis consistently identified heuristic choice and uncertainty level as the dominant predictors, together explaining approximately 63% of classification accuracy. Conclusions: Campaign-based scheduling strategies (SPT, LPT) significantly outperform round-robin approaches (FIFO) across all uncertainty conditions, with benefits remaining consistent regardless of uncertainty level. The equivalence of SPT and LPT suggests that either campaign strategy is effective, providing operational flexibility. Machine learning models can reliably predict schedule robustness, enabling proactive risk management in pharmaceutical manufacturing. Keywords: batch scheduling, pharmaceutical manufacturing, scheduling heuristics, uncertainty quantification, machine learning classification, discrete-event simulation