Source-linked AI summary
A Probabilistic Graphical Model Foundation for Enabling Predictive Digital Twins at Scale
Michael G. Kapteyn, Jacob V. R. Pretorius, Karen E. Willcox
TL;DR
The paper tackles the lack of a unifying mathematical foundation for robust digital twins at scale. It formulates assets and twins as coupled dynamical systems within a probabilistic graphical model, then demonstrates calibration and dynamic updating for a UAV structural twin. The demonstrations cover asset-specific calibration and sensor-informed responses to changing structural health.
Problem
Custom digital-twin implementations require considerable deployment resources and expertise, creating a need for a unifying foundation for robust implementations at scale.
Method
The paper represents a physical asset and its digital twin as coupled dynamical systems and formalizes their interactions with a probabilistic graphical model.
Results
The framework is applied to calibrate and evolve a structural digital twin of an unmanned aerial vehicle using experimental calibration data and sensor-informed mission updates.
Takeaways & Limitations
The graphical-model formulation provides a principled, unified basis for digital-twin calibration, updating, and dynamic decision making across physical assets.
Takeaways & Limitations
The framework still faces challenges in model parameterization and inadequacy, sensor design, and computational costs from repeated high-fidelity model evaluations.
Abstract
from arXiv · showhide
A unifying mathematical formulation is needed to move from one-off digital twins built through custom implementations to robust digital twin implementations at scale. This work proposes a probabilistic graphical model as a formal mathematical representation of a digital twin and its associated physical asset. We create an abstraction of the asset-twin system as a set of coupled dynamical systems, evolving over time through their respective state-spaces and interacting via observed data and control inputs. The formal definition of this coupled system as a probabilistic graphical model enables us to draw upon well-established theory and methods from Bayesian statistics, dynamical systems, and control theory. The declarative and general nature of the proposed digital twin model make it rigorous yet flexible, enabling its application at scale in a diverse range of application areas. We demonstrate how the model is instantiated to enable a structural digital twin of an unmanned aerial vehicle (UAV). The digital twin is calibrated using experimental data from a physical UAV asset. Its use in dynamic decision making is then illustrated in a synthetic example where the UAV undergoes an in-flight damage event and the digital twin is dynamically updated using sensor data. The graphical model foundation ensures that the digital twin calibration and updating process is principled, unified, and able to scale to an entire fleet of digital twins.
Introduction
The paper addresses the need for a unifying mathematical foundation to replace custom, resource-intensive digital-twin implementations with robust implementations at scale. It formulates the asset and twin as coupled dynamical systems and develops a probabilistic graphical model demonstrated through a UAV structural digital twin.
- Digital twins represent unique physical assets over time through computational models of their structure, behavior, and context.
- Custom digital-twin implementations require considerable deployment resources and expertise, motivating a unifying mathematical foundation for scalable implementations.
- The proposed model treats the physical asset and digital twin as coupled dynamical systems linked through observational data, model updates, and decision-making.
- A probabilistic graphical model mathematically defines the coupled system and supports data assimilation, state estimation, prediction, planning, and learning.
- The framework is demonstrated through calibration and evolution of a structural digital twin for an unmanned aerial vehicle.
Results
The paper develops a general digital-twin abstraction and dynamic decision network, then demonstrates principled calibration, operational updating, prediction, and control for a UAV structural twin.
- A Mathematical Abstraction of the Asset-Twin System: The abstraction represents an asset-twin system through six key quantities, including physical and digital states, control inputs, and application-specific quantities.The physical state captures asset variation, while the digital state contains parameters defining the twin’s computational models.
- A Mathematical Abstraction of the Asset-Twin System: The abstraction is intended as a unifying framework for defining digital twins across disciplines and application areas.UAV-specific examples instantiate quantities that remain abstract enough for broader asset-twin systems.
- A Probabilistic Graphical Model for Digital Twins: The proposed probabilistic graphical model encodes asset-twin structure, quantity evolution, and the end-to-end flow from sensing through inference and assimilation to action.It is formulated as a dynamic decision network with random variables over time and decision nodes for controls.
- A Probabilistic Graphical Model for Digital Twins: The factored representation identifies conditional probability distributions and supports sequential Bayesian inference for monitoring, prediction, and optimization.The model can represent computational interactions while leveraging models shared across fleets with asset-specific digital states.
- Demonstration: Data-driven calibration and evolution of a UAV structural digital twin: The UAV demonstration applies the model to calibration and service-life evolution, using experimental data to tailor models and sensor data to update structural health estimates.The calibration phase targets a unique as-manufactured asset, while the operational phase dynamically estimates evolving health parameters.
- Demonstration: Data-driven calibration and evolution of a UAV structural digital twin: The operational twin assimilates sensor data, estimates structural health and quantities of interest with quantified uncertainty, and suggests health-aware control inputs.The demonstration includes mission planning and a simulated degrading-structure scenario in which the twin maintains an accurate estimate of one health parameter with relatively low uncertainty.
Discussion
The paper presents a probabilistic graphical model foundation for principled digital-twin calibration, updating, uncertainty quantification, and decision-making. A UAV demonstration illustrates lifecycle extension, while model-definition, sensor-design, and computational challenges remain.
- Foundation: The graphical model defines asset-twin elements and interactions while incorporating end-to-end uncertainty quantification through Bayesian inference.It also supports dynamic model updating within data assimilation and feedback control.
- UAV demonstration: The UAV application uses calibration experiments to tailor digital-twin models to the manufactured asset.The calibrated twin is then used for operational health monitoring and adaptive mission planning.
- UAV demonstration: The framework supports principled data assimilation, model-based analysis and decision-making, and uncertainty quantification across asset-lifecycle phases.Its declarative and general formulation is presented as applicable beyond the UAV example to engineering and science applications.
- Extensions: Specialized formulations can be integrated into the framework for tasks such as detecting unknown or anomalous structural states.The paper gives damage detection and classification as an example extension.
- Limitations: The framework faces challenges in defining and parameterizing digital-twin models, managing model inadequacy, designing observable sensor systems, and meeting computational demands.High-fidelity inference and planning may require reduced-order, surrogate, or other model-compression techniques.
Methods
The UAV structural digital twin combines laboratory calibration with finite-element structural modeling, modal analysis, and operational prediction. Calibrated models assimilate loads and measurements to estimate deformation, strain, and dynamic behavior.
- Calibration: Calibration experiments use detached, inverted UAV wings mounted to a fixed support in a laboratory environment.The setup makes typical aerodynamic forces easier to apply as downward forces.
- Structural model: The computational core is a finite-element, physics-based model relating time-varying aircraft loads to structural displacement.Its mass, damping, and stiffness matrices are parameterized by the digital state.
- Calibration: A static load-displacement experiment applies a near-tip point load, solves for nodal displacement, and extracts tip displacement for calibration.The study uses the resulting relationship between aggregate wing stiffness and the Young’s modulus scale factor.
- Operational prediction: After calibration, the digital twin simulates wing deformation under arbitrary loads and predicts strain during steady level turns.The operational algorithm computes maneuver loads, solves for displacement, and post-processes strain at sensor locations.
- Dynamic characterization: Eigenanalysis characterizes structural modes using modal displacement vectors and natural frequencies, while Rayleigh damping combines mass- and stiffness-proportional damping.The calibration focuses on the first two bending modes, although the model can compute additional bending, torsional, and skin-buckling modes.
- Structural health representation: Structural health parameters represent percentage stiffness reductions in specified wing regions by reducing corresponding Young’s modulus parameters.These parameters capture damage or degradation effects in the demonstration and could represent causes such as cracks or delaminations.
Prior estimate for the digital state
The UAV digital twin is calibrated by updating prior beliefs about geometry, material properties, hardware masses, and damping through staged experiments and Bayesian inference. The resulting graphical model supports sequential monitoring, prediction, and optimal control under uncertainty.
- Prior estimates: Prior geometry is based on nominal technical-drawing values with ±2.5mm manufacturing tolerance, while material stiffness uses an expert Gaussian prior centered at scale factor 1 with ±5% variability.The geometry prior concerns l, croot, and ctip; the material prior may be refined using future manufacturing data.
- Prior estimates: The model represents servomotor and pitot-tube masses at measured locations, but calibrates their values from total mass because individual component masses cannot be measured after manufacturing.The point masses modify the mass matrix and therefore the model’s predicted natural frequencies.
- Calibration procedure: Calibration assumes zero prior Rayleigh damping and no structural damage during calibration, then updates geometry, Young’s modulus, masses, and damping using asset measurements.Geometry is measured directly; material properties use static load-displacement data; dynamic calibration fits modal frequencies.
- Calibration procedure: Eight load-displacement measurements update the Young’s modulus posterior iteratively with a particle filter after likelihood densities are estimated by sampling and kernel-density fitting.Each measurement is transformed from force and displacement into an equivalent Young’s modulus estimate before sequential assimilation.
- Dynamic updating: Sequential Bayesian inference assimilates new observations into the graphical model to update beliefs over digital states, while the same factorized representation supports monitoring, prediction, and optimization.In the demonstrative Bayesian network, discrete unobserved variables permit the classical sum-product algorithm.
- Dynamic updating: Planning treats the current-to-future graph segment as a POMDP and selects control inputs that maximize expected future reward, although exact solution is typically intractable.Prediction omits future data-assimilation factors and observed-data conditioning, allowing prediction and monitoring in one sum-product pass.
Data availability
Calibration experimental data are available in the public UAV-experimental-calibration repository.
- Calibration experimental data are available in the public repository UAV-experimental-calibration[30].
Code availability
The paper provides calibration and simulation code through public repositories, while proprietary structural-analysis source code is unavailable.
- Akselos Integra v4.5.91 is proprietary and licensed, so its source code cannot be provided; model output data are supplied instead.
Competing Interests
The authors disclose that a co-author co-founded Jessara Group, whose sensors were used in the UAV experiments, with procurement reviewed under MIT policies.
- Jessara Group sensors were used in the UAV experiments, and co-author Jacob Pretorius is a Jessara co-founder.Sensor purchase was reviewed and approved in compliance with applicable MIT policies and procedures.