Cyber-Twin for Quantum Dot Systems: Robust AI Against Experimental Outliers and Cyberattacks Injection

—The rapid evolution and application of quantumdot (QD) materials in chemistry, chemical engineering, andmedical science has created a crucial need for robust accuratemodeling frameworks, interpretable and secure design. Withoutpractical implementation of the physical product and applicationin solar cells and TV, it is important to modeling the QDthrough virtual representation of the molecules. Even-thoughthe quantum simulation and machine learning are acceleratedmaterial screening, there are 3 business challenges remain unresolved: (1) prediction accuracy is low for the DFT quantumsimulation due to noise and systematic errors, (2) experimentalinconsistencies due to outliers, and (3) most ML models lackinterpretability, undermining trust and scientific insight. Tothis end, we propose a cyber-twin framework that combinesgraph neural networks, specifically GINEConv, Bayesian Gaussian Process Regression, and advanced anomaly detection basedon Isolation Forests. The proposed model learns chemicallyrich representations from simulated quantum dot structures,while Bayesian inference enables uncertainty-aware predictionscalibrated against experimental variability. We simulate realisticscenarios including 5 % experimental outliers, 10% noise-basedattacks, 7% bias-targeted attacks, and 3% full corruption, anddesign multi-feature anomaly scores to detect these threats. Thesystem achieves over 90% detection accuracy for outliers, witha predictive R² of 0.95 and strong calibration (Pearson r =0.87) between uncertainty and absolute error. Additionally, SHAPanalysis identifies key atomic features—such as electronegativityand bond disorder—as dominant drivers of NIR absorption,offering explainability aligned with chemical knowledge.