GNN-Based Generative Models Reconstruct Missing Sensor Fields for Safety-Critical Digital Twins

Summary:Researchers at Oregon State University developed a sparse-sensor reconstruction framework for safety-critical digital twins, combining graph-based spatial encoding with recurrent temporal modeling. Six approaches were evaluated: conditional diffusion, VAE, and GAN models; deterministic GNN-GRU and GRU baselines; and ridge regression. In the PG-29 depressurized conduction-cooling experiment, GNN-GRU VAE achieved the best intra-experiment result, with 11.73°C MAE, 20.54°C RMSE, and 0.98 R². For cr

---

Core Findings

This study addresses partial observability in safety-critical digital twins. Thermocouples exposed to high temperatures and radiation may drift, fail, or become intermittently unavailable, leaving a digital twin without a complete temperature field. Using data from Oregon State University's High Temperature Test Facility, the researchers compared six approaches: conditional diffusion models (DM), variational autoencoders (VAEs), generative adversarial networks (GANs), deterministic GNN-GRU and GRU models, and ridge regression.

For intra-experiment reconstruction in the PG-29 depressurized conduction-cooling test, GNN-GRU VAE achieved the best results: 11.73°C MAE, 20.54°C RMSE, and 0.98 R². Deterministic GNN-GRU reached 25.26°C RMSE, while plain GRU reached 35.22°C. This difference indicates that explicitly modeling spatial relationships provides value beyond temporal sequence modeling alone. All three generative models reduced RMSE by more than half relative to ridge regression. The diffusion model retained an R² of 0.94 when the number of available sensors fell from 91 to 45.

The cross-experiment protocol was harder: models were trained on the PG-27 pressurized conduction-cooling experiment and evaluated on PG-29 using section-context conditioning. The section-conditioned diffusion model achieved 42.84°C RMSE and 0.91 R². The larger error than in intra-experiment reconstruction shows that operating-condition changes remain a significant source of uncertainty.

Technical Mechanism

Temperature-field reconstruction differs from conventional missing-value imputation. Available thermocouples form an irregular sparse sampling set, while the target is a continuous multivariate field. A GRU can capture cooling dynamics over time, but it does not inherently encode which measurement points are physically adjacent.

A graph encoder represents sensors or facility sections as nodes and propagates information over physical connections. A typical message update is:

[ \mathbf{h}_i^{(l+1)}=\sigma\left(W_l[\mathbf{h}i^{(l)},\operatorname{Agg}{j\in\mathcal{N}(i)}\mathbf{h}_j^{(l)}]\right) ]

where (\mathcal{N}(i)) is the neighborhood of node (i), and the aggregation operator combines information from neighboring sensors or sections. This topology-aware representation allows the model to infer missing locations from local relationships instead of treating every sensor channel as an independent time series.

The GRU models the nonstationary evolution of the cooling process:

[ \mathbf{z}t=(1-z_t)\odot\mathbf{h}{t-1}+z_t\odot\tilde{\mathbf{h}}_t ]

GNN and GRU therefore play complementary roles: the former encodes spatial structure, while the latter summarizes temporal dynamics.

The generative methods represent a conditional distribution rather than only a deterministic estimate:

[ p(\mathbf{x}{\text{miss}}\mid\mathbf{x}{\text{known}},c) ]

Here, (c) can include observed sensors, graph structure, and operating-context features. Diffusion models learn this distribution through iterative denoising; VAEs approximate it with an encoder-decoder architecture and latent variables; GANs use adversarial training to produce realistic samples. Multiple plausible reconstructions are particularly useful when sensors fail, because downstream systems need uncertainty information rather than only a point estimate.

Engineering Trade-offs

Superior interpolation accuracy does not automatically make a generative model safe to operate. A generated field may look smooth while violating heat-transfer physics, crossing an adiabatic boundary, or creating unrealistic gradients. Systems should therefore evaluate maximum error, gradient continuity, conservation constraints, and confidence-interval coverage in addition to average RMSE.

The cross-experiment RMSE of 42.84°C, compared with 20.54°C for intra-experiment reconstruction, demonstrates the difficulty of transferring across operating conditions. Section conditioning helps, but it does not replace explicit boundary conditions, sensor-health information, or governing equations. A practical deployment could use the learned model as a fast state estimator and validate each reconstructed field against a heat-transfer solver or finite-element model before allowing it to influence control and safety decisions.

Inference cost also differs by architecture. Diffusion models generally require multiple denoising steps, increasing latency relative to a single-pass deterministic predictor. VAEs and GANs may offer lower latency, although training stability and their ability to cover multimodal outcomes must be evaluated. System-level selection should consider P50 and P99 latency, memory consumption, reconstruction cost, and worst-case behavior under realistic sensor-failure patterns rather than RMSE alone.

Interpreting the Evaluation

The distinction between intra- and inter-experiment protocols is a strength. The first measures fit within a known experiment, while the second probes transfer under changed conditions. Together, they support the value of graph-based spatial modeling and contextual generative reconstruction. However, the abstract does not provide complete data-splitting details, training duration, variance across random seeds, coverage of sensor-failure combinations, or confidence intervals.

For safety-critical systems, an R² of 0.91 or 0.98 is not sufficient for deployment by itself. Operational decisions depend on whether errors remain below material-specific thresholds, whether sensor bias is correctly identified, and whether the model remains calibrated under unseen failure combinations. Future evaluations should stratify results by failure type, sensor location, severity, and reconstruction uncertainty while also reporting tail errors and physical-violation rates.

Deployment Recommendations

A practical implementation can encode the sensor topology as a graph, represent cooling dynamics with a GRU or lightweight temporal Transformer, and use a VAE, diffusion model, or deterministic regressor as the reconstruction head. Diffusion models are attractive when multiple samples and uncertainty estimates are required. If the control loop can accept only one estimate under a strict latency budget, a VAE or GNN-GRU may be preferable.

A three-layer safety design is advisable:

1. Data layer: record sensor drift, intermittent disconnections, and calibration status so the model can distinguish missing values from unreliable observations.
2. Model layer: produce both point estimates and uncertainty intervals, while sampling multiple candidate fields for high-uncertainty cases.
3. Physics layer: reject reconstructed states that violate heat-transfer constraints, section boundary conditions, or cross-sensor consistency checks.

The paper's broader contribution is not a single generative architecture, but an integrated approach to reconstructing incomplete sensor fields through topology, temporal dynamics, generative uncertainty, and experimental context. Extending validation to drift, unknown sensor layouts, and online distribution shifts would strengthen its role as a dependable state-estimation layer for safety-critical digital twins.

---

Source: ArXiv Machine Learning (cs.LG) (2026-11-09)

— END —

Tags: #Digital Twins Follow #Sparse Sensor Reconstruction Follow #GNN Spatiotemporal Modeling Follow #Diffusion Models Follow #Safety-Critical Systems Follow