The term "digital twin" is used for everything from static CAD models to autonomous predictive systems, yet implementations are rarely classified by capability level. This creates two critical problems. First, procurement teams, regulators, technical managers, and researchers cannot assess what is actually being delivered when someone claims to have built a "digital twin." Second, unrealistic expectations develop around predictive capability and bidirectional data flow without understanding fundamental limitations. This paper addresses both problems. We present a six-level framework (Level 0-5) that classifies digital twins by capability: from static models through monitoring, diagnostics, prediction, prescription, to autonomy. We then consider the mathematical barriers that inhibit the development of Level 3 digital twins for complex systems. When thousands of time-varying parameters (fuel thermal conductivity, cladding oxide thickness, gap conductance, material degradation) change unmeasurably during operation while only dozens of outputs can be measured, adaptive estimation fails. The system is severely underdetermined. Nonlinearity creates non-convex error surfaces where gradient methods converge to wrong solutions. Lyapunov convergence conditions are violated. We use barrier analysis to establish realistic expectations for realizing Level 3 systems and to inform related validation requirements.
Rube Ben Williams (Mon,) studied this question.