Digital Twins Lack Verification, Validation, and Uncertainty Quantification Methods for Bidirectional Feedback Loops

Digital twins — computational models dynamically coupled to physical systems through bidirectional data flows — are expected to transform medicine, manufacturing, climate science, and infrastructure management. But no verification, validation, and uncertainty quantification (VVUQ) framework exists that can handle their defining feature: continuous bidirectional updating between the virtual model and its physical counterpart. Traditional VVUQ methods from modeling and simulation assume a fixed model is validated once against test data. Digital twins break this assumption because the model state, parameters, and even structure change over time as new data arrives and decisions are fed back to the physical system. A digital twin without trustworthy VVUQ is, as the National Academies committee stated, "not trustworthy" — but the methods to make it trustworthy don't exist yet.

Digital twins are being promoted across high-stakes domains: patient-specific cancer treatment planning, predictive maintenance for aircraft engines, real-time weather forecasting, and critical infrastructure monitoring. The U.S. government has launched multi-agency digital twin initiatives (DoD, DOE, NIH, NSF all co-sponsored this study), and a Fast-Track Action Committee was established to develop a National Digital Twins R&D Strategic Plan. Without credible VVUQ, none of these applications can be responsibly deployed in safety-critical settings, and it becomes impossible to distinguish legitimate digital twin capabilities from what the report calls "merely aspirational" claims. The gap between enthusiasm and rigor threatens both public safety and the credibility of the field.

Traditional model validation compares simulation outputs against experimental data at a fixed point in time — this does not account for models that continuously assimilate new data and evolve. Sensitivity analysis and Monte Carlo uncertainty propagation work for individual simulations but scale poorly to the coupled, multiphysics, multiscale systems that digital twins represent. Surrogate models (reduced-order models, neural network emulators) reduce computational cost but introduce their own unquantified approximation errors, and papers in the literature routinely fail to acknowledge the true cost of generating sufficient training data for these surrogates. Data assimilation methods from weather forecasting handle sequential updates but assume known model structure and well-characterized observation errors — neither holds for general digital twins. Privacy-preserving approaches conflict with the need for comprehensive data flows. The report found that "little attention has been given to sustainability and maintenance or life-cycle management of digital twins," meaning even basic questions about when a digital twin becomes untrustworthy remain unanswered.

A framework for continual VVUQ that treats validation as an ongoing process rather than a one-time gate. This would require: scalable uncertainty quantification algorithms for high-dimensional parameter spaces; methods for quantifying and propagating uncertainty through coupled multiscale models; approaches for detecting when a digital twin has drifted outside its validated operating regime; and standards for reporting VVUQ results that enable comparison across implementations. The DOE Predictive Science Academic Alliance Program was cited as a possible organizational model. Adjacent fields like Bayesian online learning and adaptive experiment design may offer transferable techniques.

A team could build a simple digital twin (e.g., a thermostatically controlled room, a battery charge/discharge cycle, or a structural health monitoring beam) with a deliberate model mismatch, then implement and compare different approaches for detecting when the model diverges from reality: statistical tests on prediction residuals, Bayesian model evidence tracking, or ensemble-based uncertainty bounds. The key deliverable would be a quantitative comparison of how quickly each method detects model degradation and at what computational cost. Skills needed: numerical methods, statistics, and basic control systems or physics modeling.