Development of Chemistry-Informed Machine Learning Models for Cyber Threat Detection and Protection of Water and Wastewater Infrastructure

Abstract

 Water treatment plants increasingly depend on networked control, and an intruder who manipulates dosing while replaying normal analyser readings can defeat monitors that judge each signal alone. We developed a chemistry informed machine learning (CIML) framework that turns the chemistry of treatment into consistency checks on supervisory control and data acquisition (SCADA) data. Carbonate equilibrium, Kohlrausch conductivity, Nernstian redox potential, first order chlorine kinetics and reservoir mass balance were expressed as eleven residuals computed from historian tags, with coefficients calibrated by least squares on normal operation. A dual channel detector combined a joint autoencoder on raw signals and residuals with a residual energy score, and a gradient boosted triage stage trained on a digital twin attack library separated attacks from benign anomalies. We tested it on a chemically explicit twin of a coagulation, filtration and chlorination plant over 110 test days containing seven attack scenarios, natural storms and four types of instrument fault or process upset. Calibration recovered dosing stoichiometry within 1 % and electrode and decay parameters within 5 %. Storms left every balance residual at baseline, whereas each attack left a distinct fingerprint. CIML with triage reached an area under the precision recall curve of 0.67 and an F1 score of 0.79, against 0.48 and 0.41 for the strongest data driven reference, with 0.19 false alarm episodes per day. A masked hypochlorite underdose missed by a raw signal autoencoder was detected 2.25 h after onset. Triage generalised poorly to unseen attack types, and attacks that leave the chemistry consistent remain hard to detect.