Variational Graph Neural Networks for Uncertainty Quantification in Inverse Problems

arXiv:2603.29515 · cs.LG · Submitted 2026-03-31 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Variational Graph Neural Networks for Uncertainty Quantification in Inverse Problems".

Jane: Here's the gist of the paper. The gist: The authors present a variational graph neural network (VGNN) architecture that integrates variational layers into its architecture to model the probability distribution of weights, allowing for efficient estimation of cognitive and statistical uncertainty in inverse problems.

Tom: The increasingly wide use of deep machine learning techniques in computational mechanics has significantly accelerated simulations of problems that were considered unapproachable just a few years ago.

Jane: In this work, we present a variational graph neural network (VGNN) architecture that integrates variational layers into its architecture to model the probability distribution of weights.

Tom: The authors propose a hybrid architecture combining the capabilities provided by the geometric biases of graph neural networks (GNNs) with variational inference.

Paper summary: Jane: So wrapping up the "Variational Graph Neural Networks for Uncertainty Quantification in Inverse Problems" paper by David Gonz´alez and his team, the main point is that they successfully built a hybrid architecture to quantify both cognitive and statistical uncertainty efficiently.

Tom: They used graph neural networks combined with variational layers in the decoder to model this distribution of weights <ref:2603.29515#pg2>. It lets you get a non-deterministic solution for inverse problems, which is what matters most for reliability Abstract.

Lu: What this means practically is that when you're solving complex physics problems where the answer isn't certain, you get more than just a single output; you get a probability distribution of possible outcomes <ref:2603.29515#pg2>.

Meng: For deployment, this suggests that we can use AI in these critical applications because it provides the necessary confidence bounds to make decisions about system safety or material properties Abstract.

Jane: It moves the field toward models that are not just fast but also honest about what they know and what they don't know when solving inverse problems Abstract.

Tom: The authors proved this using concrete solid mechanics examples, like finding an elastic modulus distribution in a 2D problem and locating loads on a three dee beam Numerical Validation <ref:2603.29515#pg1>. Those results show the model recovers parameters with high precision while providing confidence intervals consistent with the physics of the problem.

Lu: So, it's not just about using AI for speed; it's about using AI to provide a level of rigor that deterministic methods currently lack in these tricky inverse scenarios <ref:2603.29515#pg1>.

Jane: That’s the essence of what they achieved here—integrating geometric structure with probabilistic modeling to give us reliable answers when the input data is messy or the solution space is broad.

Conclusion: Jane: So, we’re wrapping up this paper on "Variational Graph Neural Networks for Uncertainty Quantification in Inverse Problems." The authors are David Gonz´alez and his team.

Tom: Right, so what's the actual core idea again? It seems like they’re taking graph neural networks, which are great for modeling shapes or structures, and adding variational layers to them.

Lu: Exactly. They’re integrating variational inference into the architecture to model the probability distribution of weights instead of just getting one fixed set of numbers for those weights.

Meng: So they’re trying to figure out not just the best answer, but how sure we are about that answer. That sounds like something you need when you're actually shipping a model in production.

Jane: Right, so instead of just saying "this is the modulus," they’re saying "here is the range of possible values for that modulus." It’s about quantifying both cognitive uncertainty and statistical uncertainty.

Tom: That’s interesting because I always worry about models giving you a confident answer that's totally wrong because the input data was noisy.

Lu: They specifically use this structure to separate the epistemic uncertainty, which is about the model not knowing enough—the lack of data—from the aleatory uncertainty, which is just random noise in your measurements.

Meng: For deployment at scale, separating those two types of error is crucial. If we can tell if an error comes from a bad training set or just a noisy sensor reading, we know how to handle it differently.

Jane: So, the authors validated this on two solid mechanics problems—finding an elastic modulus in 2D and locating loads on a three dee beam—using only the displacement field as input <ref:2603.29515#pg1>.

Tom: And they showed that the results weren't just accurate; they gave confidence intervals that actually made sense with how physics works. That’s the real payoff for me as an engineer.

Lu: They also demonstrated that they can infer parameters on meshes different from what they used to train the model, which shows a nice level of generalization.

Meng: But what about the cost? How much more expensive is this variational decoder compared to just running a standard graph neural network inference? That’s my biggest question.

Jane: We’ll have to dig into that next, because if it's too slow or computationally heavy, it might be great for research but hard to put on a real-time system.

Tom: Yeah, let's see what they say about the complexity of the variational layers. But for now, this is showing how we can get more honest answers from these complex AI models in engineering.

David González, Alba Muixí, Beatriz Moya, ʓ El⅁as Cueto

Keysight-UZ Chair of the Spanish National Strategy on AI · Laboratori de C`alcul Num`eric (LaC`aN), Universitat Polit`ecnica de Catalunya - BarcelonaTech (UPC) · Centre Internacional de M`etodes Num`erics en Enginyeria (CIMNE) · PIMM Lab. Arts et M´etiers Institute of Technology

cs.LG

Submitted: 2026-03-31

Updated: 2026-10-20

Code: https://github.com/NASA/pigans-material-ID

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 82/100

The gist: The gist: The authors present a variational graph neural network (VGNN) architecture that integrates variational layers into its architecture to model the probability distribution of weights,

Key concepts

Variational Graph Neural Network (VGNN): A hybrid deep learning architecture combining Graph Neural Networks (GNNs) for spatial data processing with variational inference. It uses variational layers, specifically in the decoder, to model the probability distribution of network weights rather than just point estimates.
Cognitive Uncertainty: This is also known as epistemic uncertainty and relates to a lack of knowledge or data. In this context, it represents how uncertain the model is because it hasn't seen enough relevant training data or lacks sufficient knowledge about the underlying physical system being modeled.
Statistical Uncertainty: This uncertainty arises from the inherent noise present in the input data. The VGNN captures this by predicting a variance ($\sigma^2$) alongside its mean prediction, quantifying how much the output might vary due to random fluctuations in the input measurements.

Terminology

Inverse Problem — 2× in this paper · explained in 2 episodes
Decoder — 1× in this paper · explained in 2 episodes
Digital Twin — 1× in this paper · explained in 6 episodes
Initial Data — 1× in this paper · explained in 2 episodes
Uncertainty Quantification — 1× in this paper · explained in 9 episodes

Summary

The gist: The authors present a variational graph neural network (VGNN) architecture that integrates variational layers into its architecture to model the probability distribution of weights, allowing for efficient estimation of cognitive and statistical uncertainty in inverse problems.

Introduction and Motivation

The increasingly wide use of deep machine learning techniques in computational mechanics has significantly accelerated simulations of problems that were considered unapproachable just a few years ago. In this work, we present a variational graph neural network (VGNN) architecture that integrates variational layers into its architecture to model the probability distribution of weights. The authors propose a hybrid architecture combining the capabilities provided by the geometric biases of graph neural networks (GNNs) with variational inference. This approach allows for quantifying two types of uncertainty: cognitive (or epistemic) uncertainty, which is the uncertainty associated with a lack of data or knowledge of the model—this type of uncertainty can be reduced with more data—; and statistical (or random) uncertainty, which is the uncertainty associated with the noise inherent in the input data.

Neural Network Architecture

The VGNN architecture follows a structure consisting of an encoder, a processor, and a decoder. The Encoder consists of two independent multi-layer perceptron (MLP) networks for nodes and edges. The Processor is the interaction core that performs message passing steps to propagate information across the domain using the connectivity defined in the graph. For each message step k = 1... m, an edge's information is updated by calculating a “message”—interaction value from the information stored in the connected nodes. The Node aggregation and updating step involves aggregating messages from the neighborhood N(i) and updating the node's state.

Variational Decoder and Uncertainty Modeling

Unlike Graph Networks where all weights calculated by the network are constant, this work proposes a hybrid approach where uncertainty is modeled only in the decoder. The network has two outputs: one predicts a mean vector µ and another predicts its variance σ 2. This dual-output structure allows for explicit distinction between two fundamental types of uncertainty: epistemic (cognitive) uncertainty is modeled by the probability distribution of weights in the variational decoder. Statistical (aleatory) uncertainty is captured through the network’s predicted variance σ 2 and the learned global sampling noise σnoise.

Training Methodology

The training methodology utilizes stochastic training of the neural network using Variational Inference The training involves maximizing the Evidence Lower Bound (ELBO), which is equivalent to minimizing a loss function L = LNLL + β · LKL. The DKL term acts as a complexity-dependent regularizer.

Numerical Validation

The proposed VGNN architecture was validated in two cases of solid mechanics: the identification of the value of the elastic modulus with nonlinear distribution in a 2D elastic problem and the location and quantification of the loads applied to a 3D hyperelastic beam. In both cases using only the displacement field of each test as input data. For example in 4.1 estimation of Young’s modulus E(x, y) at nodes under plane stress. The model allows inference to be performed on meshes other than those used during training.

Improvements for AI systems

Variational Graph Neural Networks can estimate material properties like Young's modulus with high precision while simultaneously providing a position-dependent field of uncertainty, allowing users to see the areas of greater or lesser confidence in the results.
The system can locate and quantify applied loads on 3D hyperelastic beams, offering confidence intervals for that experiment, which is crucial for experimental verification.
The architecture distinguishes between two types of uncertainty: epistemic uncertainty from the probability distribution of weights and statistical uncertainty from predicted variance sigma squared, allowing the model to admit its inability to produce a deterministic answer when the input data is unreliable or the inverse problem is ill-conditioned.
The training methodology optimizes the Evidence Lower Bound (ELBO) by balancing likelihood with complexity regularization, ensuring that while accurately fitting displacement fields, it avoids overfitting and promotes sparse weights.
The model's ability to perform inference on meshes other than those used during training demonstrates versatility, allowing the technique to be applied effectively in changing scenarios and even those not seen in training for novel geometries.