Skip to content
Conference Open access

Physics-Informed Neural Network-Guided Reduced-Order Models for Real-Time Digital Twins

Oct 2026 · International Conference on Innovative Mechanisms for Industry Applications · pp. 681-688 · 0 citations · 17 references

Abstract

Real-time digital twin deployments demand surrogate models that simultaneously deliver high-fidelity state predictions and sub-millisecond response under continuously shifting operational conditions. Standard high-fidelity solvers, such as finite element analysis (FEA) and computational fluid dynamics (CFD), impose computational overheads that render real-time evaluation infeasible, while purely data-driven reduced-order models (ROMs) tend to lose accuracy when applied outside the conditions encountered during training. To address this gap, we introduce PINN-ROM-DT: a framework combining projection-based model order reduction with physics-informed neural network (PINN) regularization inside a unified, continuously updatable digital surrogate. The framework rests on three interlocking components: (i) an Adaptive Proper Orthogonal Decomposition (APOD) basis that enlarges its reduced subspace on the fly whenever incoming sensor data signals departure from the training manifold; (ii) a PINN residual regularizer that penalizes PDE-constraint violations in the reduced coordinate space, suppressing non-physical drift during online adaptation; and (iii) a Kalman-informed Latent State Observer (KLSO) that fuses sparse sensor measurements with reduced-order predictions to yield calibrated uncertainty estimates without offline retraining. We evaluate the framework on four representative industrial scenarios: thermomechanical loading in a gas turbine blade, flow dynamics in a subsea pipeline network, structural health monitoring of a wind turbine tower, and thermal regulation in a battery pack. PINN-ROM-DT attains a mean relative L2 error of 0.74% at 0.87 ms inference latency-a 61.4% reduction relative to POD-Galerkin ROMs-and sustains sub-1% error even when operational parameters deviate 35% beyond the training envelope.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.