💻 computer science

A manifold-aware Neural ODE surrogate model for stochastic induction heating with anisotropic electrical conductivity

This paper proposes a manifold-aware Neural ODE surrogate model that respects the symmetric positive-definite structure of anisotropic electrical conductivity to efficiently propagate uncertainty in stochastic induction heating processes for fibre-reinforced thermoplastic composites.

Original authors: Wouter J. Schuttert, Mohammed Iqbal Abdul Rasheed, Bojana Rosić

Published 2026-08-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Wouter J. Schuttert, Mohammed Iqbal Abdul Rasheed, Bojana Rosić

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to bake the perfect loaf of bread, but instead of flour and water, your ingredients are layers of super-strong plastic and carbon fiber. To stick these layers together into a sturdy structure, you need to melt the interface just right. Too cold, and they won't bond; too hot, and you ruin the material. The tool you use is a magical "induction heater," which is like a high-tech microwave that uses invisible magnetic waves to make the carbon fibers themselves get hot, kind of like how a metal spoon gets warm in a pot of soup.

But here's the tricky part: carbon fiber isn't a perfect, uniform block of metal. It's more like a bundle of thousands of tiny, individual straws. Sometimes, during manufacturing, these straws get squished together perfectly; other times, they are a bit crooked or have gaps. This messiness changes how electricity flows through the material. In the world of physics, we call this "anisotropic conductivity," which is just a fancy way of saying "electricity flows better in some directions than others." Because the material is a bit random and messy, the heat it generates is also unpredictable. If you try to predict the temperature using a standard computer model, it's like trying to guess the exact path of a leaf blowing in the wind by assuming the wind never changes. You need a smarter way to handle all that randomness.

This is where the story of the paper comes in. The researchers wanted to understand exactly how these tiny manufacturing imperfections in carbon fiber affect the heating process. They knew that if they just guessed, they might miss the mark. So, they built a super-smart "digital twin" of the heating process. Instead of running thousands of slow, heavy simulations to see what happens (which would take forever), they taught a special kind of artificial intelligence to learn the rules of the game.

The secret sauce of their AI is that it respects the shape of the data. Imagine the electrical conductivity as a spinning, squishy balloon. A normal computer sees this balloon as a list of numbers, which is like trying to describe a balloon's shape by just listing its weight. It loses the geometry. The researchers built a new type of neural network that understands the balloon is a 3D object that can stretch and rotate, but never pop or turn inside out. They call this a "manifold-aware" network. They also taught it to move through time like a smooth movie rather than a choppy slideshow, using a technique called "Neural ODEs."

When they tested this new AI, they found something important: the randomness in the material (the crooked fibers) causes the temperature to become very unpredictable as time goes on. The uncertainty doesn't stay small; it grows, like a snowball rolling down a hill. The AI was able to predict this growing chaos accurately. They tried different ways to make the AI "think" about time—some were like taking quick, small steps (Euler), while others were like taking a more careful, calculated path (Runge-Kutta and Adams-Moulton). They found that while the quick-step method was slightly faster, the more careful methods were more consistent and reliable, giving the same answer every time you asked them.

In the end, the researchers showed that by respecting the unique, wobbly nature of the material and using a smart, time-aware AI, they could predict how these high-tech composites heat up with much less computing power than before. This means engineers could one day design better welding processes for airplanes and cars, ensuring that the lightweight materials holding them together are fused perfectly, without needing to run millions of expensive simulations to find out what works. The paper suggests that this approach is a powerful new tool for handling the messy reality of manufacturing, turning a chaotic problem into a manageable one.

Technical Summary: A Manifold-Aware Neural ODE Surrogate for Stochastic Induction Heating

Problem Statement
Induction welding is a critical process for joining lightweight carbon fibre-reinforced thermoplastic (CFRTP) composites, where bond quality is governed by the temperature history at the interface. However, direct measurement of this temperature is often infeasible, necessitating predictive physics-based models. Existing deterministic models fail to account for significant variations in electrical conductivity caused by manufacturing imperfections, such as fibre placement deviations and fibre-fibre contact quality. These variations are inherently anisotropic and must be represented by a symmetric positive-definite (SPD) conductivity tensor. Standard surrogate modeling approaches treat inputs as Euclidean vectors, ignoring the Riemannian manifold structure of SPD tensors, and often fail to capture the full spatial and temporal evolution of the temperature field under uncertainty. Furthermore, direct uncertainty propagation via Monte Carlo (MC) simulations of the coupled electromagnetic-thermal system is computationally prohibitive.

Methodology
The authors propose a surrogate modeling framework that respects the geometric structure of the conductivity tensor and captures temporal dynamics efficiently.

Stochastic Material Model:
The electrical conductivity tensor $\sigma$ is modeled as a random SPD-valued variable. To ensure physical admissibility (symmetry and positive definiteness) and separate physical mechanisms, the tensor is decomposed into scaling (eigenvalues) and orientation (eigenvectors) uncertainties:
$\sigma(\omega) = V(\omega)\Lambda(\omega)V(\omega)^T$
- Scaling: Eigenvalues are modeled as lognormal random variables to represent magnitude uncertainty.
- Orientation: Eigenvectors are perturbed via a rotation matrix derived from a skew-symmetric matrix, with the rotation angle modeled by a von Mises distribution to represent directional uncertainty.
  This decomposition allows for independent control of magnitude and directional anisotropy while preserving the SPD property.
High-Fidelity Solver:
The coupled electromagnetic and heat-transfer equations are solved using the Finite Element Method (FEM) in COMSOL Multiphysics. The electromagnetic problem uses a magnetic vector potential formulation under the magneto-quasi-static (MQS) approximation. The thermal problem is a transient heat conduction equation driven by Joule heating from induced eddy currents. A Monte Carlo analysis with 20,000 realizations is performed to generate a dataset of temperature trajectories under three uncertainty scenarios: scaling-only, orientation-only, and combined.
Surrogate Architecture (CMNN-NODE):
To replace the expensive FEM solver, a data-driven surrogate is trained to learn the discrete-time evolution of the temperature field.
- Manifold-Aware Input (CMNN): The conductivity tensor is processed through a Constitutive Manifold Neural Network (CMNN) layer, specifically a "Strength-Angular" (StrAng) mapping. This layer performs eigendecomposition, applies logarithmic maps to the eigenvalues and rotation matrices (mapping them to their respective tangent spaces), and flattens the result. This ensures the network receives a Euclidean representation that preserves the intrinsic SPD geometry.
- Temporal Dynamics (Neural ODE): The surrogate is formulated as a Neural Ordinary Differential Equation (NODE). Instead of learning a direct state-to-state map, the network learns the vector field $f_\xi(\sigma, u)$ governing the rate of change of the temperature. The update rule is defined as:
  $\frac{du_h(t)}{dt} = f_\xi(\sigma, u_h(t))$
- Integration Schemes: The study evaluates three integration schemes for the Neural ODE: Explicit Euler, explicit Runge-Kutta 4 (RK4), and implicit Adams-Moulton (AM). The models are trained to minimize the mean squared error between predicted and observed one-step increments.

Key Contributions

SPD-Aware Surrogate: Development of a surrogate model that explicitly respects the Riemannian manifold structure of the conductivity tensor via the StrAng layer, avoiding the geometric distortions of standard vectorization.
Uncertainty Quantification: A comprehensive stochastic analysis of induction heating, separating the effects of conductivity magnitude and orientation uncertainties on the temperature field.
Neural ODE Integration: A comparative study of different Neural ODE integrators (Euler, RK4, AM) for learning transient multiphysics dynamics, assessing their trade-offs between accuracy, stability, and run-to-run consistency.

Results

Uncertainty Propagation: Monte Carlo simulations reveal that stochastic conductivity drives significant temperature variability. Scaling uncertainty is the dominant source of variance, while orientation uncertainty has a secondary but non-negligible effect. Uncertainty in the temperature field increases continuously over time due to the cumulative nature of the transient heat transfer process.
Surrogate Performance:
- All three surrogate models (CMNN-Euler, CMNN-RK4, CMNN-AM) achieve comparable accuracy on unseen test cases, with mean $L_2$ errors around $0.33\text{--}0.44^\circ\text{C}$ .
- The Euler baseline achieves the lowest mean errors on several metrics but exhibits significantly higher variability across repeated training runs (lower robustness).
- The Neural ODE models (RK4 and AM) demonstrate superior run-to-run consistency, indicating greater robustness in the learned dynamics.
- Among the Neural ODE integrators, RK4 yields lower pointwise errors, while the implicit Adams-Moulton scheme better preserves the distributional characteristics (Kullback-Leibler divergence) of the temperature field.
Stability Analysis: Stability analysis reveals that the learned dynamics are locally expansive (amplifying perturbations) rather than contractive, a known failure mode for learned time-steppers trained without explicit spectral constraints. However, the Neural ODE formulations provide more consistent stability properties across training instances compared to the Euler baseline.

Significance and Claims
The paper claims to provide a computationally efficient framework for uncertainty propagation in induction heating, enabling the analysis of full spatial temperature fields under manufacturing-induced variability at a fraction of the cost of full Monte Carlo FEM simulations. The primary significance lies in the integration of manifold-aware inputs (respecting the SPD nature of conductivity) with Neural ODEs to capture temporal dynamics. The authors modestly note that the surrogate is trained and evaluated within specific parameter ranges and does not claim generalization beyond the training distribution or to spatially varying random fields. The work demonstrates that while standard Euler-based surrogates can be accurate, Neural ODE formulations offer improved robustness and consistency, making them suitable for reliable uncertainty quantification in multiphysics processes involving SPD-valued parameters.

Technical Summary: A Manifold-Aware Neural ODE Surrogate for Stochastic Induction Heating

More like this