Probabilistic Predictions of Process-Induced Deformation in Carbon/Epoxy Composites Using a Deep Operator Network

This study presents a hybrid physics-informed and data-driven framework utilizing Deep Operator Networks, transfer learning, and Ensemble Kalman Inversion to accurately predict and quantify uncertainty in process-induced deformation of carbon/epoxy composites, ultimately enabling the optimization of non-isothermal cure cycles for deformation mitigation.

The Gist

Imagine you are baking a very complex, high-tech cake. But instead of flour and sugar, your ingredients are layers of carbon fiber and a special liquid glue (epoxy resin). You want this cake to be perfectly flat and strong when it's done.

However, there's a problem: as the cake bakes, the different ingredients shrink at different rates. The fiber wants to stay the same size, but the glue shrinks as it hardens. This mismatch causes the cake to warp, twist, or curl up like a potato chip. In the engineering world, this is called Process-Induced Deformation (PID). If the part warps too much, it's useless for building airplanes or cars.

This paper is about a team of scientists who built a "super-smart crystal ball" to predict exactly how much this cake will warp before they even turn on the oven. They want to find the perfect baking schedule (temperature over time) to keep the cake flat.

Here is how they did it, broken down into simple steps:

1. The "Physics Simulator" (The Digital Twin)

First, the team built a super-detailed computer simulation. Think of this as a digital twin of their factory. They programmed it with the laws of physics: how heat moves, how the glue shrinks, and how the fibers stretch.

• The Result: They ran thousands of virtual baking cycles on the computer. This gave them a massive library of data showing exactly what happens to the material under every possible temperature schedule.

2. The "Deep Operator Network" (The Super-Translator)

Running the physics simulation every time they wanted to test a new idea is slow and expensive, like trying to bake a real cake just to see if a new recipe works.
So, they trained an AI called a Deep Operator Network (DeepONet).

• The Analogy: Imagine a translator who doesn't just translate words, but translates entire stories. If you give the AI a "temperature story" (a graph of how hot the oven gets over time), it instantly tells you the "deformation story" (how much the part will warp).
• The Trick: They added a special feature called FiLM (Feature-wise Linear Modulation). Think of this as a "volume knob" for the AI. The AI knows that if the glue starts out slightly harder (a higher "degree of cure") before baking, the final result will be different. FiLM lets the AI adjust its prediction based on that starting condition, just like a chef adjusting a recipe based on the humidity in the kitchen.

3. The "Transfer Learning" (The Smart Student)

Here was the tricky part: They had tons of computer data, but very little real experimental data. In the real lab, they could only measure how much the part warped at the very end of the process, not every second along the way.

• The Solution: They used a technique called Transfer Learning.
• The Analogy: Imagine a student who has read every physics textbook in the world (the computer simulation). They are an expert in theory. Now, they go to a real lab and take a test. They only get the final grade (the final warp measurement). Instead of re-learning everything from scratch, the student just tweaks their final conclusion to match the real grade, while keeping all their deep physics knowledge intact. This allowed them to predict the entire warping history of the real experiment, even though they only measured the end result.

4. The "Uncertainty Crystal Ball" (EKI)

In science, it's not enough to make a guess; you need to know how confident you are in that guess. What if the oven has a glitch? What if the material is slightly different?

• The Tool: They used a method called Ensemble Kalman Inversion (EKI).
• The Analogy: Instead of asking one expert for an answer, they asked 2,000 slightly different experts (an "ensemble"). They all looked at the data and gave their own predictions. By looking at how much their answers agreed or disagreed, the team could draw a "confidence band." If the experts all agree, the band is tight (high confidence). If they argue, the band is wide (low confidence). This tells the engineers, "We are 95% sure the part will warp this much."

5. The "Perfect Recipe" (Optimization)

Finally, they put it all together to solve the main problem: How do we bake this so it stays flat?

• They used their AI and uncertainty tools to search for the perfect "temperature schedule." They found a specific way to heat and cool the material that minimized the warping while ensuring the glue hardened completely.
• The Result: They found a "Goldilocks" temperature curve that reduced the warping by about 8–10% compared to the standard method.

The Big Picture

This paper is a story of combining three powerful tools:

1. Physics Simulations (The deep knowledge).
2. AI (DeepONets) (The fast translator).
3. Uncertainty Math (EKI) (The confidence checker).

By mixing them, the scientists created a tool that can predict how complex materials will behave, fix the mistakes before they happen, and design better manufacturing processes for everything from airplane wings to sports equipment. It's like having a time machine that lets you see the future of your manufacturing process and fix it before you even start.

Technical Summary

1. Problem Statement

Composite manufacturing, particularly for thermoset fiber-reinforced polymers (e.g., carbon/epoxy), is plagued by Process-Induced Deformation (PID). PID arises from the mismatch in thermal expansion coefficients between fibers and the matrix, combined with matrix shrinkage during the curing (cross-linking) process. These heterogeneities generate residual stresses that, upon release during demolding, cause the final part to deviate from design tolerances.

Key Challenges:

• Complexity: The cure process involves coupled thermo-chemical-mechanical phenomena (viscosity evolution, degree of cure, modulus changes) across multiple length scales.
• Data Scarcity: High-fidelity physics-based simulations are computationally expensive, while experimental data is often limited to final deformation measurements (time-history data is rarely available).
• Uncertainty: Manufacturing variability (e.g., initial degree of cure, material properties) introduces uncertainty that standard deterministic models fail to capture.
• Optimization: Designing non-isothermal cure cycles to minimize PID while ensuring full cure is a high-dimensional inverse problem.

2. Methodology

The authors propose a hybrid framework integrating physics-based simulation, Deep Operator Networks (DeepONets), Transfer Learning, and Ensemble Kalman Inversion (EKI).

A. Physics-Based Modeling & Data Generation

• Material System: AS4 carbon fiber/amine bi-functional epoxy prepreg (Hexply® 3501-6).
• Simulation Framework: A sequential two-step approach using ABAQUS and COMPRO:
  1. Thermo-chemical analysis: Models temperature evolution, Degree of Cure (DoC), and viscosity.
  2. Stress-deformation analysis: Calculates residual stresses and PID.
• Experimental Validation: Asymmetric $[0/90]$ cross-ply laminates were manufactured using a baseline isothermal cycle and two optimized non-isothermal cycles. PID was measured via mid-span deflection.
• Dataset: A diverse dataset of non-isothermal cure cycles (time-temperature profiles) was generated via simulation, varying an intermediate control point $A=(t_1, T_1)$ to create a wide range of thermal histories.

B. Deep Operator Network (DeepONet) Architecture

Instead of learning mappings between finite-dimensional vectors, the model learns operators mapping functions to functions (e.g., a temperature profile function $\to$ a deformation history function).

• Standard DeepONet: Uses a Branch Network (encodes the input function, i.e., temperature profile) and a Trunk Network (parameterizes dependence on query points, i.e., time).
• FiLM Enhancement (Feature-wise Linear Modulation): To account for the Initial Degree of Cure ($DoC_0$), which significantly affects the process, the branch network is augmented with FiLM. This allows the network to condition its internal features on $DoC_0$, enabling the prediction of DoC, viscosity, and deformation histories for varying initial material states.
• Architecture Details: Branch and trunk networks use 3 hidden layers with 20 neurons (tanh activation). Viscosity data was log-transformed to handle exponential growth.

C. Transfer Learning Strategy

Since experimental data provides only the final deformation (sparse data) rather than full time histories:

1. Pre-training: The DeepONet is first trained on the large, high-fidelity simulation dataset.
2. Fine-tuning: The network is applied to experimental cases. Only the final layer of the branch network is updated using a loss function that penalizes the mismatch between the predicted final deformation and the measured experimental value. The rest of the network (the learned operator) remains fixed. This prevents overfitting to sparse data.

D. Uncertainty Quantification (UQ) via EKI

To quantify epistemic uncertainty and optimize cure cycles:

• Ensemble Kalman Inversion (EKI): A gradient-free ensemble method is used to train the DeepONet. An ensemble of models (particles) is evolved to minimize the discrepancy between predictions and observations.
• Chebyshev-enhanced KANs (cKANs): To improve computational efficiency and reduce parameter dimension, standard MLPs within the DeepONet are replaced with cKANs, which adapt activation patterns to input data.
• Result: This approach generates an ensemble of predictions, providing mean estimates and uncertainty bounds (standard deviation) for DoC, viscosity, and deformation.

3. Key Contributions

1. FiLM-DeepONet Surrogate: Development of a data-driven surrogate that successfully predicts full time-histories of DoC, viscosity, and deformation for arbitrary non-isothermal cure cycles, conditioned on the initial degree of cure.
2. Transfer Learning for Sparse Data: A novel strategy to adapt simulation-trained operator networks to experimental data where only terminal states are known, effectively recovering full deformation histories.
3. Probabilistic Framework: Integration of EKI with DeepONets to provide rigorous uncertainty quantification, moving beyond deterministic predictions to probabilistic ones.
4. Optimization Capability: Demonstration of using the trained surrogate to solve constrained optimization problems (minimizing PID while ensuring $DoC \ge 0.99$).

4. Key Results

• Model Accuracy: The simulation model was validated against experiments, showing simulated PID within experimental standard deviation for the baseline and <10% error for optimized cycles.
• Surrogate Performance: The FiLM-DeepONet achieved high accuracy in predicting DoC, viscosity, and deformation histories for both simulation and experimental cases. It successfully distinguished responses based on different initial $DoC_0$ values.
• Uncertainty Quantification: The EKI-based ensemble provided well-calibrated uncertainty bounds. For experimental data, transfer learning significantly improved the match to the final deformation, though uncertainty increased in the early trajectory (reflecting the lack of intermediate data).
• Optimization: The framework successfully identified an optimized cure schedule (intermediate point $A \approx (1.61 \text{ min}, 133.01^\circ\text{C})$) that minimized final deformation while satisfying the full-cure constraint. This result aligned closely with previous optimization studies.
• PID Reduction: Optimized non-isothermal cycles achieved an 8–10% reduction in process-induced deformation compared to the baseline.

5. Significance

This work represents a significant advancement in Scientific Machine Learning (SciML) for composite manufacturing:

• Bridging Simulation and Experiment: It solves the critical "data gap" problem by leveraging rich simulation data to train models that can be efficiently adapted to sparse experimental data via transfer learning.
• Physics-Informed AI: By using DeepONets, the model respects the functional nature of the physical process (time-series evolution) rather than treating it as a static regression problem.
• Robust Decision Making: The integration of EKI allows engineers to not only predict outcomes but also quantify confidence in those predictions, which is essential for risk-averse manufacturing optimization.
• Scalability: The use of cKANs and EKI suggests a pathway to handling high-dimensional inverse problems that were previously computationally prohibitive.

In summary, the paper presents a robust, uncertainty-aware framework that enables the rapid design of optimized cure cycles for composite materials, reducing defects and improving manufacturing reliability.