💻 computer science

Uncertainty Quantification in Calibration and Simulation of Thermo-Chemical Curing of Epoxy Resins

This study investigates the propagation of uncertainties during the multi-step calibration and numerical simulation of thermo-chemical curing in epoxy resins, demonstrating that the first-order second-moment method provides an efficient and reasonably accurate approximation despite the highly nonlinear nature of the stochastic model.

Original authors: Jendrik-Alexander Tröger, Christina Steinweller, Stefan Hartmann

Published 2026-01-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jendrik-Alexander Tröger, Christina Steinweller, Stefan Hartmann

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

The Big Picture: Baking a Perfect Cake

Imagine you are trying to bake a very complex cake (the epoxy resin) that needs to harden perfectly to hold a heavy structure together (like an airplane wing). The problem is that the cake doesn't just sit there; it reacts to heat, changes its own texture as it bakes, shrinks as it cools, and even generates its own heat while cooking.

If you get the recipe wrong, the cake might burn in the middle, crack on the outside, or never harden properly. To prevent this, engineers use computer simulations to predict exactly how the cake will behave. But here's the catch: no recipe is perfect. The ingredients (material parameters) you measure in the lab always have tiny errors, like a scale that is off by a gram or a thermometer that fluctuates.

This paper asks: "If our measurements of the ingredients are slightly wrong, how much does that mess up our computer prediction of the cake?"

The Problem: A Chain Reaction of Guesses

The researchers explain that figuring out the "recipe" for this epoxy isn't a one-step process. It's like solving a puzzle where the pieces depend on each other.

First, you measure how the material changes temperature.
Then, you use that result to figure out how fast it hardens.
Then, you use that result to figure out how much it shrinks.

If your first measurement has a tiny error, that error gets passed down to the next step, then the next, and so on. By the time you get to the final simulation, that tiny error might have grown into a huge mistake. This is called uncertainty propagation.

The Solution: Two Ways to Check the Math

The authors wanted to see how these errors grow. They compared two methods:

The "Monte Carlo" Method (The Brute Force Approach):
Imagine you have a robot baker. To see how sensitive the cake is to bad measurements, you tell the robot to bake the cake 2,000 times. In every single bake, you slightly tweak the ingredients (sometimes a bit more sugar, sometimes a bit less heat) based on random errors. Then, you look at all 2,000 cakes to see how much they vary.
- Pros: It's very accurate.
- Cons: It takes a long time and a lot of computer power.
The "FOSM" Method (The Shortcut):
This stands for First-Order Second-Moment. Think of this as a smart math shortcut. Instead of baking the cake 2,000 times, the mathematician looks at the recipe once and asks, "If I change the sugar by a tiny bit, how much does the cake change?" They use a simple linear approximation (a straight line) to guess the result.
- Pros: It is incredibly fast. It takes a fraction of the time.
- Cons: It assumes the relationship is simple and straight. If the cake behaves in a crazy, non-linear way, this shortcut might miss the details.

What They Found

The researchers tested these methods on real epoxy resin data. Here is what they discovered:

The Shortcut Works (Mostly): Even though the epoxy reaction is very complex and "non-linear" (like a cake that suddenly doubles in size if you add one extra egg), the fast FOSM shortcut gave results that were very close to the slow, expensive Monte Carlo method.
The "Non-Gaussian" Trap: The shortcut works best when errors are random and average out (like a bell curve). However, for one specific ingredient (the "diffusion factor"), the errors didn't follow a normal pattern. In this specific case, the shortcut wasn't perfect, but it was still reasonable.
The "Sparse Data" Warning: They found that if you don't have enough data points to start with (like trying to guess a recipe with only 5 taste tests), the errors can get huge. They showed that using the "shortcut" method to track these errors is crucial; otherwise, you might think your prediction is perfect when it's actually very shaky.
Boundary Conditions Matter: They also tested what happens if the oven temperature (the boundary condition) is uncertain. They found that if the oven temperature fluctuates, the uncertainty in the final result explodes. It turns out that knowing exactly how hot the oven is is even more important than knowing the exact chemical properties of the resin.

The Bottom Line

This paper proves that you don't always need to run thousands of expensive computer simulations to understand how uncertain your results are. A smart, fast mathematical shortcut (FOSM) is usually good enough to tell you if your prediction is reliable.

However, you have to be careful:

If your initial data is very scarce, the errors will grow fast.
If the "oven" (boundary conditions) is unstable, your whole prediction could be wrong, regardless of how good your recipe is.

In short: Use the fast shortcut to save time, but always double-check your oven temperature and make sure you have enough data to start with.

Technical Summary: Uncertainty Quantification in Calibration and Simulation of Thermo-Chemical Curing of Epoxy Resins

Problem Statement
The curing of epoxy resins is a complex thermo-chemo-mechanical process essential for manufacturing high-performance composite materials, particularly in the aviation industry. Accurate modeling of this process requires constitutive relations for glass transition temperature, curing kinetics, thermal expansion, chemical shrinkage, specific heat capacity, and thermal conductivity. These models rely on numerous material parameters that must be identified through a multi-step calibration procedure using experimental data (e.g., DSC, TMDSC, density measurements).

A critical challenge identified in this work is the propagation of uncertainties. Previous studies, such as those by Leistner [4] and extensions in [5], established the constitutive relations but neglected uncertainty propagation. In a multi-step calibration scheme, where later steps depend on parameters identified in earlier steps, uncertainties inevitably accumulate. Furthermore, the presence of uncertain material parameters renders the deterministic model response uncertain, potentially leading to inaccurate predictions of temperature distributions, residual stresses, and delamination risks. The paper addresses the need to quantify these uncertainties efficiently during both the inverse calibration problem and the forward numerical simulation of curing processes.

Methodology
The study employs a multi-step model calibration approach using the ordinary nonlinear least-squares method. Material parameters are identified sequentially based on specific experimental datasets, as dependencies exist between constitutive relations (e.g., glass transition temperature influences curing kinetics and specific heat capacity).

To quantify uncertainties, the authors utilize the First-Order Second-Moment (FOSM) method. This perturbation-based approach approximates the mean and variance of the model response using Taylor expansions. The FOSM is applied in two distinct phases:

Inverse Problem (Calibration): The method propagates uncertainties from previously identified parameters to subsequent calibration steps. The covariance matrix of the current step is updated to include contributions from both measurement noise and the propagated uncertainty of input parameters.
Forward Problem (Simulation): The method quantifies the uncertainty in the model response (temperature and degree of cure) resulting from the identified parameter uncertainties.

The FOSM results are rigorously evaluated against the Monte Carlo (MC) method, which serves as a robust reference. While MC is computationally expensive due to slow convergence and the need for thousands of samples, it provides a benchmark for assessing the accuracy of the FOSM's first-order approximation, particularly in the presence of highly nonlinear constitutive relations and non-Gaussian parameter distributions.

Key Contributions

Systematic Uncertainty Propagation: The paper implements a comprehensive framework for propagating uncertainties through a multi-step calibration scheme for epoxy resin curing, a process previously treated deterministically in the referenced literature.
Efficiency of FOSM in Nonlinear Systems: The study demonstrates that FOSM, despite being a first-order approximation, yields reasonable results for highly nonlinear thermo-chemical models, offering a significant computational advantage over Monte Carlo simulations.
Coverage Testing: The authors conduct extensive "in-silico" coverage tests using synthetic data to evaluate the robustness of the Jacobian-based covariance matrix approximation under various conditions, including sparse data, dense data, and different noise types (Gaussian, uniform, and heteroscedastic).
Boundary Condition Analysis: The study extends uncertainty quantification to include uncertain boundary conditions (Dirichlet and mixed), comparing their impact on model response against material parameter uncertainties.

Results

Calibration Uncertainties: The FOSM and Monte Carlo methods show good agreement for most material parameters. However, significant discrepancies arise for parameters with clearly non-Gaussian distributions (e.g., the diffusion factor $b_d$ in curing kinetics), where FOSM struggles to capture the true distribution shape.
Impact of Data Density: Coverage tests reveal that calibration steps based on sparse data (e.g., only 5 data points for glass transition temperature) can lead to substantial underestimation of uncertainty if not properly accounted for. Increasing the amount of data improves the reliability of the uncertainty estimates.
Propagation Effects: Uncertainty propagation significantly increases the variance of parameters in later calibration steps (e.g., specific heat capacity parameters) compared to estimates derived without propagation. Neglecting this propagation leads to confidence intervals that fail to cover the true parameter values at the nominal confidence level.
Simulation Results: In numerical simulations of curing in an aluminum container, FOSM and Monte Carlo results for nodal temperatures and degree of cure agree closely. The temperature uncertainty is relatively small, while the degree of cure exhibits larger uncertainties, particularly during rapid curing transitions.
Boundary Conditions: The study finds that uncertainties in Dirichlet boundary conditions (prescribed temperature profiles) have a far more significant impact on model response uncertainty (orders of magnitude higher) than uncertainties in mixed boundary conditions (convection and radiation coefficients).

Significance and Claims
The authors claim that the FOSM method is a viable, efficient, and sampling-free approach for uncertainty quantification in the calibration and simulation of epoxy resin curing, even when dealing with highly nonlinear constitutive relations. While acknowledging that FOSM is limited by its first-order approximation and may not fully resolve complex non-Gaussian distributions, the study demonstrates that it provides reasonable results with significantly lower computational cost than Monte Carlo methods.

The work emphasizes that uncertainty propagation is not merely a theoretical refinement but a practical necessity; ignoring it in multi-step calibration schemes can lead to unreliable parameter estimates and overly optimistic confidence intervals. Furthermore, the results highlight that controlling uncertainties in boundary conditions (specifically temperature profiles) is critical for reliable numerical simulations, as these uncertainties can dominate the error budget more than material parameter uncertainties. The study concludes that FOSM is a robust tool for engineering applications where computational efficiency is paramount, provided its limitations regarding distribution shapes and large input variances are understood.

The Big Picture: Baking a Perfect Cake

The Problem: A Chain Reaction of Guesses

The Solution: Two Ways to Check the Math

What They Found

The Bottom Line

More like this