Machine Learning-Based Fault Detection and Recovery Prediction in Digital Microfluidic Biochips: An Integrated XGBoost-SMOTE Framework

Abstract—Digital Microfluidic Biochips (DMFBs) are rapidly emerging as critical platforms for biochemical processing,point-of-care diagnostics, and on-chip assay automation. Despite their programmability and miniaturization advantages,DMFBs remain susceptible to operational failures, including droplet routing errors, electrode actuation faults, dielectricdegradation, and unsafe recovery transitions. These faults threaten the integrity and reproducibility of microfluidic assays.This paper proposes a comprehensive, two-stage machine learning framework for real-time fault detection and recoveryoutcome prediction in DMFBs. In the fault detection stage, an XGBoost classifier trained on SMOTE-balanced operationaldroplet log data achieves 98.7% accuracy and an ROC-AUC of 0.99, demonstrating near-perfect discrimination betweenfaulty and normal operations. In the recovery prediction stage, a comparative analysis of Random Forest andXGBoost-SMOTE classifiers reveals that the proposed XGBoost-SMOTE model substantially improves minority-classrecall—identifying 27% of recovery failures that the baseline model misses entirely. Feature analysis consistently identifiesrecovery_method, activated_pins, and safe_factor_avg as the most significant predictors. Interpretability tools including ROCcurves, precision-recall plots, normalized confusion matrices, gain-based importance rankings, and correlation heatmaps areemployed to provide actionable insights. This integrated framework advances DMFB fault tolerance toward autonomous,data-driven, and self-healing microfluidic control.