(579f) Optimizing Crashworthiness of Hexagonal Composite Rings Using Inverse Physics-Informed Neural Networks

Composite materials, formed by blending two or more distinct components while retaining their individual characteristics, offer enhanced properties such as increased strength, reduced weight, and improved crashworthiness [1]. Given their widespread use and advantageous properties, understanding and optimizing specific composite materials, such as woven roving glass/epoxy composites, becomes crucial. These materials find extensive applications across industries, including aerospace, automotive, marine, construction, and civil engineering, benefiting from their exceptional mechanical properties, such as high strength-to-weight ratios and resistance to environmental factors [2]. Woven roving glass/epoxy composites offer notable advantages, including excellent performance-to-cost ratios, corrosion resistance, and efficient manufacturing potential [3]. Moreover, their architectural versatility makes them ideal for applications such as stadium and airport ceilings, where lightweight yet durable materials are essential [4]. Therefore, investigating and optimizing their crashworthiness parameters is essential for advancing safety standards in industrial sectors. Researchers use mechanical criteria to evaluate crashworthiness, often through axial crushing stress tests and random stress tests, but the lack of comprehensive data makes finding optimal structures time-consuming [5]. Machine learning algorithms, computer-aided smart manufacturing, offer efficient analysis of experimental data to build accurate predictive models, aiding in exploring alternative composite materials [6]. Traditional methods (Response Surface Methodology, Gradient-based Optimization, Genetic Algorithm, Simulated Annealing) struggle to encapsulate the nuanced relationships between the crashworthiness parameters of the hexagonal composite ring specimens under lateral compressive, energy absorption, and failure modes [7]. This is precisely where Physics Informed Neural Networks (PINNs) emerge as a vital tool [8]. PINNs offer a distinct advantage by seamlessly integrating mechanistic principles with data-driven insights, allowing us to delve deep into the underlying physics of the depolymerization process.

This study presents a novel approach utilizing Inverse Physics-Informed Neural Networks (IPINNs) comprising a multi-modal Random Forest Classifier-Artificial Neural Network (RFC-ANN) ensemble framework to optimize the crashworthiness of hexagonal composite ring structures. The aim is to enhance the safety and crash resilience of woven roving glass/epoxy composite materials for diverse Industry 4.0 applications. Leveraging advanced machine learning algorithms, specifically RFC-ANN(IPINNs), we delve into the intricate relationship between crashworthiness parameters, such as lateral compressive strength, energy absorption, and failure modes, exhibited by hexagonal composite ring specimens. Drawing inspiration from previous advancements in the optimization of complex systems [3], our study focuses on elucidating the complex interplay between structural configurations and crash performance metrics. Through extensive experimentation and model refinement using Reinforcement Learning (RL) algorithms, our approach enables the accurate prediction of crashworthiness parameters for various hexagonal ring angles. Experiments were conducted in-house to determine the load-displacement curves and deformation history of composite hexagonal ring structures under uniaxial compression directions. Depending on the hexagonal angle, the hexagonally packed systems reacted differently to the crushing load. Therefore, the crashworthiness parameters were calculated based on the load-displacement curves, deformation history, energy absorption capability and listed for a one-dimensional composite hexagonal array system. The detailed results and figures are available in a previous publication [3]. In this work, PINNs are trained using DeepXDE (library for scientific machine learning and physics-informed learning) at PyTorch backend to accurately predict the angles between the experimental data points and outside the experimental data point range, indicating the developed model's robustness. It generates the necessary composite configurations while leveraging the established complex relationship among the six crashworthiness parameters (e.g., crushing load, crush force efficiency, specific energy absorption). The model successfully forecasted the optimal configuration of the composite for different random sets of crashworthiness characteristics. In addition, the developed model was serialized using the pickle module, allowing for the conversion of Python objects into byte sequences, while the ANN model's features are saved as a JSON file containing structure, weight, and bias values. These files enable us to estimate the crashworthiness characteristics for any specific composite sample with known hexagonal angle and loading condition (i.e., model trained can predict characteristics for a 63° angle). Reinforcement Learning (RL) algorithms, such as Proximal Policy Optimization (PPO) and Trust Region Policy Optimization (TRPO), are employed to refine the Physics Informed Neural Network (PINN) model iteratively using experimental feedback [9, 1]. These RL algorithms are well-suited for continuous control tasks, making them suitable for optimizing complex processes. We utilize popular RL frameworks (OpenAI's Baselines), which provide implementations of state-of-the-art RL algorithms along with tools for training, evaluation, and monitoring. These platforms offer robust support for experimenting with RL algorithms and integrating them into the optimization pipeline seamlessly. Additionally, custom RL implementations using libraries like TensorFlow or PyTorch enable fine-tuning the algorithms to suit the specific requirements of the PINN model refinement process. The developed IPINN model not only offers accurate predictions of crashworthiness parameters but also provides valuable insights into the underlying physics governing the crash behavior of composite materials. By integrating mechanistic principles with data-driven insights, our model transcends traditional boundaries, offering a holistic understanding of the complex interactions between structural design and crash performance. Furthermore, the IPINN framework enables inverse design capabilities, allowing for the targeted optimization of composite materials based on desired crashworthiness characteristics. This inverse design approach holds promise for revolutionizing the design and engineering of advanced composite materials, paving the way for safer and more resilient structures in various Industry 4.0 applications.

References

1. Kazi M-K, Eljack F, Mahdi E. Design of composite rectangular tubes for optimum crashworthiness performance via experimental and ANN techniques. Composite Structures. 2022;279:114858. doi:https://doi.org/10.1016/j.compstruct.2021.114858.
2. Mahdi E, Faraj M, Hamouda AMS, Sapuan SM, Attia MA, Eltai EO. Effect of hexagonal on the in-plane crushing behaviour of plain weave composite hexagonal quadruple ring system. International Journal of Crashworthiness. 2020;25(2):192-202. doi:10.1080/13588265.2018.1555025.
3. Kazi M-K, Mahdi E. Crashworthiness optimization of composite hexagonal ring system using random forest classification and artificial neural network. Composites Part C: Open Access. 2024;13:100440. doi:https://doi.org/10.1016/j.jcomc.2024.100440.
4. Sreejith M, Rajeev RS. 25 - Fiber reinforced composites for aerospace and sports applications. In: Joseph K, Oksman K, George G, Wilson R, Appukuttan S, editors. Fiber Reinforced Composites. Woodhead Publishing; 2021. p. 821-59.
5. Kazi M-K, Eljack F, Mahdi E. Optimal Filler Content for Cotton Fiber/PP Composite based on Mechanical Properties using Artificial Neural Network. Composite Structures. 2020:112654. doi:https://doi.org/10.1016/j.compstruct.2020.112654.
6. Kazi M-K, Eljack F, Mahdi E. Data-Driven Modeling to Predict the Load vs. Displacement Curves of Targeted Composite Materials for Industry 4.0 and Smart Manufacturing. Composite Structures. 2020:113207. doi:https://doi.org/10.1016/j.compstruct.2020.113207.
7. Nicolle A, Deng S, Ihme M, Kuzhagaliyeva N, Ibrahim EA, Farooq A. Mixtures Recomposition by Neural Nets: A Multidisciplinary Overview. Journal of Chemical Information and Modeling. 2024;64(3):597-620. doi:10.1021/acs.jcim.3c01633.
8. Seo J. Solving real-world optimization tasks using physics-informed neural computing. Scientific Reports. 2024;14(1):202. doi:10.1038/s41598-023-49977-3.
9. Lin P, Wu M, Xiao Z, Tiong RLK, Zhang L. Physics-informed deep reinforcement learning for enhancement on tunnel boring machine's advance speed and stability. Automation in Construction. 2024;158:105234. doi:https://doi.org/10.1016/j.autcon.2023.105234.