💻 computer science

Enhancing the Computational Efficiency of Aircraft Assembly Simulations Using Deep Learning Warm-Start

This paper proposes a deep learning warm-start method that utilizes neural networks to generate accurate initial guesses for quadratic programming problems, significantly accelerating the computational efficiency of aircraft assembly simulations within digital twin frameworks.

Original authors: Thi Thu Hoai Trinh, Roman Krasnikov, Sergey Lupuleac, Stanislav Baklanov

Published 2026-10-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Thi Thu Hoai Trinh, Roman Krasnikov, Sergey Lupuleac, Stanislav Baklanov

Original paper licensed under CC BY 4.0 (https://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

Building a modern airplane is an exercise in extreme precision. The metal panels, ribs, and frames that form the wings and fuselage are not perfect; they carry tiny, unavoidable variations from the manufacturing process. When engineers assemble these massive structures, they must force these slightly mismatched parts together, often using thousands of rivets. This process creates internal stresses and forces that can weaken the aircraft if not carefully managed. To ensure safety and efficiency, engineers rely on digital simulations that act as a virtual twin of the real assembly line. These simulations calculate exactly how the metal will bend and where the parts will touch, allowing designers to optimize the process before a single rivet is driven. However, running these simulations is computationally expensive. The mathematics required to predict how thousands of contact points behave are so complex that solving them can take a long time, especially for large aircraft structures, slowing down the entire design cycle.

Researchers at Peter the Great St. Petersburg Polytechnic University have developed a way to speed up these simulations significantly by using artificial intelligence to give the computer a better head start. In their study, they focused on the specific mathematical challenge of determining which parts of the aircraft structure are actually touching and which are not. Instead of asking the computer to figure this out from scratch every time, they trained a neural network—a type of computer program inspired by the human brain—to predict the pattern of contact based on the initial gaps between parts. The system does not replace the heavy mathematical calculations; rather, it provides a highly educated guess about where contact will occur. This guess serves as a "warm start," allowing the main solver to skip many unnecessary steps and reach the final answer much faster.

The core of the problem lies in the sheer number of potential contact points. In a typical assembly model, there can be hundreds of pairs of surfaces that might touch. The computer must determine for each pair whether the gap between them closes completely or remains open. This is a binary decision for thousands of points simultaneously. The researchers treated this as a classification task, similar to sorting a large pile of items into two distinct groups. They generated thousands of random scenarios representing different manufacturing gaps and fastener arrangements, solved the contact problems for each one, and used the results to teach the neural network. The network learned to look at the initial conditions—the size of the gaps and the forces involved—and predict the final pattern of contact with high accuracy.

The results of this approach were measured using two different models of aircraft assembly, one with 121 potential contact pairs and a larger, more complex one with 621 pairs. The neural network proved remarkably effective at predicting the correct contact pattern. For the smaller model, the system correctly identified the contact status in over 91 percent of cases, while for the larger model, it achieved an accuracy of about 80 percent. Crucially, the researchers found that perfect accuracy was not necessary to achieve a major speedup. Even when the prediction was not flawless, providing the solver with a starting point that was closer to the truth than a standard guess allowed the computer to converge on the final solution much more quickly.

When the researchers tested the system on 200 independent problems, the time savings were substantial. For the smaller model, the use of the neural network reduced the total time required to solve the problem by between 24 and 41 percent, depending on the specific mathematical formulation used. For the larger, more complex model, the reduction was still significant, ranging from 24 to 28 percent. The time it took for the neural network to make its prediction was negligible, taking only a few milliseconds, which is a tiny fraction of the total time saved. The distribution of solving times also shifted, meaning that the computer was less likely to get stuck in long, difficult calculations. The method worked by identifying the "active set" of constraints—the specific points where contact occurs—and feeding this information into the solver. This allowed the solver to bypass the initial, slow phase of figuring out where to look and focus immediately on refining the solution.

The study confirms that deep learning can serve as a powerful accelerator for engineering simulations without altering the underlying physics or the mathematical algorithms used to solve them. The researchers demonstrated that a neural network does not need to solve the entire complex problem itself; it only needs to predict the structure of the solution with reasonable accuracy. This insight allows the traditional, highly accurate solvers to do what they do best, but starting from a much better position. The approach is flexible and does not depend on a specific type of solver, suggesting it could be applied to a wide range of engineering problems beyond aircraft assembly. By automating the generation of training data and integrating the neural network directly into their existing software, the team created a practical tool that balances the speed of prediction with the accuracy required for safety-critical design. The work highlights a future where artificial intelligence acts as a silent partner to traditional engineering, handling the initial guesswork so that human experts can focus on the design and optimization of the aircraft itself.

Technical Summary: Enhancing the Computational Efficiency of Aircraft Assembly Simulations Using Deep Learning Warm-Start

Problem Statement
The simulation of aircraft assembly processes, particularly the stress-strain analysis of structures involving contact interactions, relies heavily on solving convex quadratic programming (QP) problems. These problems arise from modeling the elimination of initial technological gaps between mating parts and the subsequent deformation under fastener installation. While specialized solvers exist (e.g., active-set algorithms, interior-point methods, operator splitting), their runtime grows polynomially with the problem dimension. In large-scale aircraft assembly models, where the number of potential contact pairs can reach several thousand, this computational cost becomes a bottleneck for digital twin applications. The core challenge is to accelerate these solvers without altering the underlying mathematical formulation or the optimization algorithm itself.

Methodology
The authors propose a "Deep Learning Warm-Start" approach that leverages machine learning to generate high-quality initial guesses for the QP solver. The methodology is structured as follows:

Problem Formulation as Classification: Instead of attempting to predict the continuous solution vector (displacements or gaps), the problem is reframed as a multi-label binary classification task. The goal is to predict the active set of constraints ( $I_{act}$ ), which identifies which potential contact pairs will be in contact (gap eliminated) versus those that remain inactive (residual gap) in the final solution.
Data Generation: A fully automated pipeline within the ASRP (Assembly Simulation of Riveting Process) software package generates training data. This involves:
- Creating random fields of initial gaps using a Gaussian random process model (based on Zaitseva et al.).
- Randomly generating fastener configurations.
- Solving the resulting contact problems to obtain the ground-truth active sets.
- Datasets of 10,000 and 30,000 examples were generated for models with 121 and 621 potential contact pairs, respectively.
Neural Network Architecture: A Multilayer Perceptron (MLP) is employed.
- Input: A vector of size $2m$ containing contact forces and initial gaps for all $m$ potential pairs.
- Output: A vector of size $m$ representing the probability of each constraint being active.
- Structure: Four fully connected hidden layers with LayerNorm, LeakyReLU activation, and Dropout. The hidden layer width is experimentally tuned (512 for $m=121$ , 1024 for $m=621$ ).
- Training: The network is trained using a weighted binary cross-entropy loss function to address the significant class imbalance (where non-contact pairs often outnumber contact pairs). Optimization is performed via the Adam algorithm with early stopping based on the Macro-F1 metric.
Warm-Start Integration: During inference, the trained network predicts a binary vector indicating the active set. This prediction is converted into an initial guess for the solver by setting the initial active set $I_{act}^{(0)}$ based on the predicted labels. The residual and Lagrange multipliers are initialized to zero. The solver (a first-order gradient projection method with catalyst acceleration) then proceeds to refine the solution from this starting point without modification to its core algorithm.

Key Contributions

Novel Application to Contact Problems: Unlike existing "learning-to-optimize" literature focused on Model Predictive Control (MPC) or general QP, this work specifically addresses the unique characteristics of aircraft assembly contact problems: large numbers of potential contact pairs, strong class imbalance, and dense Hessian matrices.
Active Set Prediction Strategy: The paper demonstrates that predicting the structure of the active set (a classification problem) is sufficient to accelerate the solver, avoiding the complexity of directly regressing high-dimensional continuous solution vectors.
Integration with ASRP: The method is fully integrated into the ASRP software, providing a practical tool for aerospace informatics that accelerates existing solvers without requiring algorithmic changes.
Handling Class Imbalance: The study explicitly addresses the imbalance in contact data (where active constraints are the minority) through weighted loss functions and specific evaluation metrics (Macro-F1).

Experimental Results
The method was evaluated on two aircraft assembly models ( $m=121$ and $m=621$ ) using 200 independent test problems.

Prediction Quality: The model achieved a Macro-F1 score of 0.91 for the smaller model and 0.80 for the larger model. The authors note that perfect recovery of the active set is not necessary for effective acceleration.
Computational Speedup:
- For the $m=121$ model, solution time was reduced by 37% (dual formulation) and 41% (relative formulation).
- For the $m=621$ model, reductions were 28% (dual) and 24% (relative).
Inference Overhead: The time required for the neural network to generate the initial guess was negligible (2.83 ms for $m=121$ and 5.81 ms for $m=621$ ), representing less than 3% and 0.12% of the total solution time, respectively.
Distribution Shift: The use of the warm-start systematically shifted the solution time distribution to lower values and reduced the "right tail" (outliers with long runtimes), improving the predictability of solver performance.

Significance and Claims
The paper claims that the proposed deep learning warm-start method offers a practical and efficient way to accelerate large-scale contact simulations in aircraft assembly. The primary significance lies in demonstrating that approximate recovery of the active set structure is sufficient to achieve substantial speedups (24–41%) without compromising the accuracy of the final solution, as the numerical solver refines the remaining components.

The authors emphasize that this approach is algorithm-agnostic regarding the final solver, making it applicable to a broad class of computational mechanics problems. They conclude that while increasing model dimension naturally decreases prediction accuracy, the method remains effective, and the trade-off between network complexity and classification accuracy can be managed to balance construction speed and solver performance. Future work is suggested to explore graph neural networks and transformers, but the current study establishes the viability of MLPs for this specific domain.

Technical Summary: Enhancing the Computational Efficiency of Aircraft Assembly Simulations Using Deep Learning Warm-Start

More like this