PiGGO: Physics-Guided Learnable Graph Kalman Filters for Virtual Sensing of Nonlinear Dynamic Structures under Uncertainty

2026-04-29 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

The Physics-Guided Graph Neural ODE (PiGGO) framework is a novel physics-informed, graph-based Bayesian state estimation approach designed for virtual sensing of nonlinear dynamic structures under uncertainty. It addresses challenges in digital twin deployment, such as unknown nonlinear dynamics and sparse sensing, which limit purely physics-based or data-driven methods. PiGGO integrates a learned graph neural ordinary differential equation (GNODE) as a continuous-time state-transition model within an extended Kalman filter. This graph representation explicitly defines the system state-space, while physics-guided inductive biases encode known structural relationships and constrain the learning of nonlinear dynamics. Numerical case studies demonstrate PiGGO's enhanced robustness to model uncertainty and measurement noise, outperforming open-loop graph neural models and conventional filtering approaches in online prediction tasks.

Key takeaway

For Machine Learning Engineers developing digital twins for complex structures, PiGGO offers a robust solution for online state estimation. Its integration of learned graph neural dynamics with Bayesian filtering improves reliability under model uncertainty and sparse sensing. You should consider PiGGO for systems requiring uncertainty-aware virtual sensing, especially where traditional physics-based or purely data-driven methods fall short, to achieve better online prediction accuracy.

Key insights

PiGGO combines learned graph neural dynamics with Bayesian filtering for robust, uncertainty-aware state estimation in nonlinear systems.

Principles

• Integrate physics-guided inductive biases.
• Utilize graph representations for state-space.
• Combine learned dynamics with recursive Bayesian filtering.

Method

PiGGO employs a learned GNODE as a continuous-time state-transition model within an extended Kalman filter, using graph representation and physics-guided inductive biases.

In practice

• Apply to nonlinear systems with unknown model form.
• Enhance online virtual sensing capabilities.
• Improve robustness to measurement noise.

Topics

• PiGGO Framework
• Graph Neural ODE
• Extended Kalman Filter
• Virtual Sensing
• Nonlinear Dynamic Structures

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• WaveGraphNet: Physics-Consistent Guided-Wave Damage Localization through Coupled Inverse-Forward Graph Learning — Coupling inverse graph localization with a physics-consistent forward model improves damage localization extrapolation in unseen regions.
• Two-Layer Linear Auto-Regressive Models Estimate Latent States — Two-layer linear auto-regressive models approximate Kalman filtering for latent state estimation in partially observed linear dynamical systems.
• Subject to: Laureano Escudero — Laureano Escudero's career exemplifies dedication to mathematical optimization and its real-world applications.
• Bayesian 3D Steerable CNNs: Enabling Equivariance and Uncertainty Quantification Simultaneously — Bayesian Steerable-CNNs unify SE(3)-equivariance with uncertainty quantification via stochastic kernels and variational inference.
• One-shot learning for the complex dynamical behaviors of weakly nonlinear forced oscillators — MEv-SINDy enables one-shot learning of complex nonlinear dynamics from minimal data.
• Robust Parameter and State Estimation in Multiscale Neuronal Systems Using Physics-Informed Neural Networks — PINNs offer robust parameter and state estimation in multiscale neuronal systems, overcoming limitations of traditional methods.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.