⚡ electrical engineering

Adaptive Digital Twin of Sheet Metal Forming via Proper Orthogonal Decomposition-Based Koopman Operator with Model Predictive Control

This paper presents an adaptive Digital Twin framework for robotic sheet metal forming that integrates Proper Orthogonal Decomposition, the Koopman operator, and online Recursive Least Squares to enable real-time, data-driven Model Predictive Control of complex, nonlinear deformation processes.

Original authors: Yi-Ping Chen, Derick Suarez, Ying-Kuan Tsai, Vispi Karkaria, Guanzhong Hu, Zihan Chen, Ping Guo, Jian Cao, Wei Chen

Published 2026-07-14

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Yi-Ping Chen, Derick Suarez, Ying-Kuan Tsai, Vispi Karkaria, Guanzhong Hu, Zihan Chen, Ping Guo, Jian Cao, Wei Chen

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 shape a piece of metal into a perfect curve, like the sleek hood of a vintage car. In the old days, a master artisan would stand there, hammering and rolling the metal, feeling the shape with their hands and eyes, adjusting their force with every single push. It was a dance between human skill and stubborn metal. But what if you wanted a robot to do this dance? The problem is that metal is tricky. It bends, it springs back, and it gets harder the more you work it. If you tell a robot exactly how to move once, the metal might not listen the second time, or the third.

This is where the researchers from Northwestern University stepped in with a "Digital Twin." Think of this not as a simple 3D model, but as a super-smart, magical mirror that lives inside a computer. This mirror doesn't just look like the metal; it thinks like the metal. It predicts exactly how the metal will react before the robot even touches it.

The Problem: Metal is a Mood Ring

The team was working with a tool called an English wheel, a machine that rolls metal between two wheels to stretch and shape it. The challenge? The metal's behavior changes constantly. If you push it too hard, it hardens (a process called strain-hardening). If you push it too soft, it doesn't bend enough.

Previous attempts to control this with robots often failed because they tried to use simple rules (like "push here, then push there") or complex simulations that took too long to calculate. The paper argues that simply using standard machine learning isn't enough either; those models are often "black boxes" that guess without understanding the physics, and they get confused when the metal changes its mind mid-process.

The Solution: A Linear Map for a Non-Linear World

The researchers built a new kind of Digital Twin using a clever mix of three ideas:

The "Squishy" Compression (POD): Imagine you have a giant, messy pile of data showing how the metal bends at every single point. It's too much to handle. The team used a technique called Proper Orthogonal Decomposition (POD) to "squish" this massive pile down into just a few key numbers. It's like taking a high-definition photo of a storm and compressing it into a simple weather report that still tells you exactly where the rain is falling. They found that just 4 of these "key numbers" could capture 99.5% of the metal's behavior.
The "Magic Translator" (Koopman Operator): Metal bending is chaotic and non-linear (it doesn't follow a straight line). But the researchers used a mathematical trick called the Koopman Operator to translate this chaos into a simple, straight-line language. Imagine trying to describe a rollercoaster ride. It's wild and twisting. But if you could translate that ride into a simple list of "up, down, left, right" instructions that a computer could solve instantly, that's what the Koopman Operator does. It lifts the messy reality into a "lifted space" where the math becomes easy and predictable.
The "Self-Correcting" Brain (RLS & MPC): Even the best map can be wrong if the terrain changes. The metal gets harder as it's worked, which the initial map didn't know. To fix this, the team added a Recursive Least Squares (RLS) algorithm. Think of this as the Digital Twin having a "learning loop." Every time the robot makes a move and the metal reacts, the Twin checks: "Did I predict this right?" If the metal bent less than expected, the Twin instantly updates its internal "control knobs" (specifically the matrix B) to be more accurate for the next move.

Finally, they used Model Predictive Control (MPC). This is like a chess player who looks six moves ahead. Instead of just reacting to the current shape, the MPC plans the next 6 cycles of rolling to ensure the metal ends up exactly where it needs to be, while obeying rules like "don't bend too sharply" or "start and end at the same spot."

The Results: Does the Magic Work?

The team tested this on a real robotic system using 316L stainless steel blanks that were 30.48 mm × 7.62 mm × 0.06 mm thick. They didn't just simulate it; they built it and ran it.

Without the self-correcting update: The robot tried to bend the metal, but because the metal got harder than expected, the robot kept under-bending. After 6 cycles, the final shape was off by a whopping 12.64 mm at the edge. The robot was stuck in a loop of trying the same thing and failing.
With the self-correcting update: The Digital Twin noticed the metal wasn't bending enough. It adjusted its internal "gain" (specifically the coefficient for the first Chebyshev polynomial, T1) and told the robot to push harder. The result? The final shape was off by only 2.79 mm. In a case with a smaller target curve, the error dropped to just 1.09 mm, and the robot finished the job in just 5 cycles instead of 6.

The system solved the complex math problem (involving 1,590 design variables) in about 0.8067 seconds on a powerful computer, meaning it could make decisions fast enough to keep up with the robot.

What This Means (and What It Doesn't)

The paper shows that this adaptive Digital Twin is a powerful way to control complex metal forming. It proves that you can take a messy, non-linear process and turn it into a manageable, linear problem that a computer can solve in real-time.

However, the authors are careful to note that this isn't a "solved problem" for every possible metal shape. Their current model is an "aggregate approximation," meaning it works well for the specific conditions they tested, but it might struggle with completely new materials or shapes it hasn't seen before. They suggest that future work needs to gather even more data and perhaps combine this with human knowledge to make the system truly autonomous for any shape.

But for now, they have successfully built a robot that can "feel" the metal through its digital twin, learn from its mistakes in real-time, and adjust its strategy on the fly to create smooth, perfect curves. It's a step toward a future where factories don't just follow instructions, but adapt and learn like a master artisan.

Technical Summary: Adaptive Digital Twin of Sheet Metal Forming via POD-Based Koopman Operator with MPC

Problem Statement

Digital Twin (DT) technologies offer significant potential for real-time monitoring and control of manufacturing processes. However, applying DT to deformation-based metal forming, specifically the English wheel process, remains challenging due to the strongly coupled spatio-temporal behavior of sheet deformation and the nonlinear relationship between toolpaths and material response. Traditional methods face distinct limitations:

Physics-based simulations (e.g., Finite Element Analysis) are too computationally expensive for real-time optimization.
Purely data-driven models often lack physical interpretability, struggle with generalizability beyond training domains, and require expensive data acquisition.
Existing control strategies (e.g., PID) are typically limited to single-input single-output systems and cannot effectively handle the high-dimensional, spatially correlated fields inherent in sheet forming.
Reinforcement Learning is often data-intensive, making it impractical for scenarios where data collection is costly.

The core challenge is developing a model that accurately captures nonlinear, spatio-temporal deformation dynamics while remaining computationally tractable for real-time decision-making via Model Predictive Control (MPC).

Methodology

The authors propose an adaptive DT framework that integrates Proper Orthogonal Decomposition (POD), a Deep Koopman Operator, and Model Predictive Control (MPC) with Recursive Least Squares (RLS) for online model updating. The system was validated on a robotic English wheel setup using 316L stainless steel.

1. Dimensionality Reduction and Feature Extraction:
To manage high-dimensional data, the framework employs two distinct decomposition strategies:

State Reduction (POD/PCA): The sheet deformation field is reduced using Proper Orthogonal Decomposition (Principal Component Analysis). The first four principal modes were retained, capturing over 99.5% of the total variance, reducing the state dimension from 8 spatial points to 4 coefficients.
Input Reduction (Chebyshev Polynomials): Toolpaths are represented using Chebyshev polynomial decomposition. This ensures that the reconstructed trajectories are inherently smooth and differentiable, satisfying manufacturing constraints. Five basis functions were used to represent the toolpaths.

2. Surrogate Modeling via Deep Koopman Operator:
A Deep Koopman Operator with a residual-connected encoder-decoder architecture is trained to learn the system dynamics in a lifted linear space.

Architecture: The encoder maps reduced-order states (PCA coefficients) to a higher-dimensional lifted space ( $z$ ). A linear evolution module (matrices $A$ and $B$ ) governs the dynamics ( $z_{k+1} = Az_k + B\tilde{u}_k$ ). The decoder projects the lifted state back to the reduced-order space.
Training Strategy: The model is trained using a composite loss function including reconstruction loss, one-step and multi-step prediction losses, stability loss (penalizing eigenvalues $>1$ ), and forward linearity loss. Residual connections and dropout (applied only to non-linear layers) are used to enhance stability and prevent overfitting.
Key Design Choice: The control inputs (Chebyshev coefficients) are not mapped through the non-linear encoder; they remain in the linear subspace to preserve the linear structure required for efficient MPC.

3. Control Strategy (MPC):
The learned linearized model is integrated into a constrained MPC framework to generate optimal toolpaths.

Formulation: The problem is cast as a convex Quadratic Programming (QP) problem, minimizing the deviation from a target geometry and control effort over a prediction horizon of 6 cycles.
Constraints: The formulation includes constraints on control amplitude and geometric smoothness derived from Chebyshev polynomial properties (e.g., ensuring the toolpath starts at $z=0$ ).
Receding Horizon: Only the first toolpath of the optimized sequence is applied. After each cycle, the system state is updated with new measurements, and the optimization is re-solved.

4. Online Model Adaptation (RLS):
To address material variability (e.g., strain hardening) and model mismatch, a Recursive Least Squares (RLS) algorithm is employed to update the control matrix $B$ in real-time.

Mechanism: The state dynamics matrix $A$ is kept fixed to preserve global stability, while $B$ is updated based on prediction errors.
Trigger: Updates occur only when the deviation between measured and predicted geometry exceeds 3 mm or when the sheet deformation stagnates (change < 3 mm between cycles). This prevents overfitting to limited feedback data.

Key Contributions

Adaptive DT Framework: Integration of POD and a Deep Koopman Operator to capture spatio-temporal dynamics of sheet forming in a reduced-order, linearly representable form.
Real-Time Decision-Making: A control framework combining Koopman-based linear modeling with constrained MPC, enabling the generation of optimal, smooth toolpaths in real-time (average solve time ~0.8s).
Online Adaptation: A mechanism using RLS to adaptively refine the control matrix $B$ during operation, maintaining accuracy against material variability and uncertainty without retraining the entire deep network.
Experimental Validation: Demonstration of the framework on a fully autonomous robotic English wheel system, validating its ability to achieve target geometries under varying conditions.

Results

The framework was validated using 80 experimental parts (75 for training, 5 for validation/testing).

Model Accuracy: The learned Koopman operator achieved strong agreement with measured data. The majority of prediction errors remained within ±5 mm near the sheet tip, with reconstruction errors for deformation and toolpaths generally within ±0.5 mm and ±0.25 mm, respectively.
Control Performance (Large Deformation Case):
- Without Model Updating: The system suffered from "underbending" and stagnation from Cycle 3 onward due to the model overestimating toolpath effectiveness. The final deviation from the target was 12.64 mm.
- With Model Updating: The RLS algorithm corrected the control gains, generating steeper toolpaths to drive the sheet into the plastic regime. The final deviation was reduced to 2.79 mm.
Control Performance (Moderate Deformation Case):
- Without Model Updating: Final deviation was 7.34 mm.
- With Model Updating: The target was reached within 5 cycles with a final deviation of 1.09 mm.
Interpretability: Analysis of the updated matrix $B$ revealed specific adjustments in control gains (e.g., reduced influence of the first-order Chebyshev coefficient), demonstrating the model's ability to adapt to physical changes in the process.

Significance and Claims

The paper claims that this work establishes a generalizable approach for creating interpretable, adaptive, and computationally efficient Digital Twins for nonlinear manufacturing systems where inputs and outputs are vector fields.

Bridging Physics and Data: The framework successfully bridges reduced-order physics representations (via POD) with data-driven adaptability (via Koopman and RLS).
Autonomous Control: It demonstrates the feasibility of autonomous, feedback-driven forming, moving beyond artisan-dependent processes to systems capable of planning and adapting strategies.
Scalability: While demonstrated on the English wheel, the authors suggest the pipeline is extendable to other processes involving spatio-temporal dynamics, such as incremental forming, forging, and additive manufacturing.
Modest Limitations: The authors acknowledge that the current model serves as an aggregate approximation of material behaviors (elastic to plastic) and may be limited by data sparsity in the joint state-input space. They note that while the online update may not achieve global accuracy in all material states, it remains highly effective for guiding control laws toward desired outcomes. Future work is proposed to address full 3D forming and integrate multi-modal data (simulations, human demonstrations) to further enhance model refinement.