Physics-Informed AI and Digital Twin Framework for Real-Time Predictive Maintenance and Energy Optimization of Hydrogen-Integrated Industrial Systems

Abstract

Hydrogen-integrated industrial systems combine electrolyzers, compressors, storage units, pipelines, heat exchangers, and power-electronic equipment with digital sensing, control, and energy-management infrastructures, creating operating environments in which equipment degradation, process stability, and energy consumption develop together rather than independently. Existing research has examined physics-informed artificial intelligence, digital twins, real-time industrial data, predictive maintenance, and energy optimization largely as separate technical streams, leaving limited quantitative evidence on how these capabilities jointly relate to integrated maintenance and energy performance. This study developed and empirically evaluated a physics-informed AI and digital twin framework grounded in Cyber-Physical Systems Theory, specifying Physics-Informed AI Capability (PIAIC), Digital Twin Integration Capability (DTIC), Real-Time Industrial Data Intelligence (RTIDI), Predictive Maintenance Intelligence Capability (PMIC), and Energy Optimization Intelligence Capability (EOIC) as predictors of Predictive Maintenance and Energy Optimization Performance (PMEOP). A quantitative, explanatory, cross-sectional, case-study-based survey design was employed. Of 320 structured five-point Likert questionnaires distributed to maintenance, process, automation, energy, AI, and managerial professionals across hydrogen, refining, power, steel, and industrial-gas settings, 289 valid responses were retained (90.3%). Reliability and validity were satisfactory (Cronbach’s α = .85–.92; KMO = .894; Bartlett’s test p < .001; 72.4% cumulative variance explained). Pearson correlations between PMEOP and all five predictors were positive and significant (r = .67–.78, p < .001). Multiple linear regression explained 73.2% of the variance in PMEOP (adjusted R² = .727; F(5, 283) = 154.63, p < .001), with EOIC (β = .31), PMIC (β = .27), DTIC (β = .23), PIAIC (β = .20), and RTIDI (β = .17) each retaining significant unique effects (p < .001) and no multicollinearity concerns (VIF = 1.67–1.92). All five hypotheses were supported. The findings indicate that maintenance and energy performance in hydrogen-integrated systems is associated with the coordinated maturity of physically grounded intelligence, synchronized virtual representation, real-time data, maintenance reasoning, and energy-oriented decision-making rather than with any single technology, offering an empirically validated basis for integrated cyber-physical asset management.