Deep neural operator for learning transient response of interpenetrating phase composites subject to dynamic loading

Minglei Lu, Ali Mohammadi, Zhaoxu Meng, Xuhui Meng, Gang Li, Zhen LiComputational 2023 · arXiv

TL;DR

This paper introduces a Deep Neural Operator (DNO) model that can rapidly and accurately predict the transient mechanical response of complex 3D composite structures under dynamic loading. By leveraging sequence-to-sequence training and incremental learning, their DNO achieves a significant speedup (up to 200x) compared to traditional methods, enabling faster design and optimization of these advanced materials.

What Problem Does This Paper Address?

Additive manufacturing (AM) allows for the creation of intricate 3D structures like Interpenetrating Phase Composites (IPCs). These IPCs offer superior mechanical properties compared to conventional composites, but their performance, especially under dynamic loading, is highly dependent on their complex 3D microstructures. Traditionally, predicting these responses requires computationally expensive Finite Element Analysis (FEA) or time-consuming experiments, which can take hours or days for each design iteration. This bottleneck significantly slows down the design and optimization process for new composite materials with desired properties. The paper aims to bridge this gap by developing a fast and accurate surrogate model for predicting these transient mechanical responses.

Methods

The core of this work is a Deep Neural Operator (DNO) model, specifically a variant of DeepONet, adapted for sequence-to-sequence learning. Unlike traditional DeepONets that process data point-by-point, this approach handles entire time sequences, improving training efficiency for time-dependent problems.

Key methodological innovations include:

  • Sequence-to-Sequence Training: The DNO is trained to map input strain sequences to output force and stress sequences, rather than individual time points. This is facilitated by using a "trunk net" for time information and a "branch net" for the corresponding strain data.
  • Incremental Learning: To further accelerate training and improve accuracy, especially with limited data, an incremental learning strategy is employed. This allows the model to learn from one element of the composite structure and then use that learned knowledge as a starting point for training subsequent elements, avoiding retraining from scratch.
  • JAX Implementation: The model is implemented in JAX, a high-performance numerical computation library, which contributes to the significant speedups observed.

Datasets

The study uses data generated from Finite Element Analysis (FEA) simulations of a 3D IPC beam.

  • FEA Simulation Data: 5,000 FEA simulations were performed on a 3D IPC beam (composed of stainless steel and aluminum alloy) subjected to 5,000 random time-dependent strain loads generated by a Gaussian process. These simulations provided transient reaction forces and stress distributions as training and testing data.
  • No New Benchmarks Introduced: The paper does not introduce new benchmark datasets but rather generates its own data from FEA simulations for the specific IPC beam problem.

Key Insights and Ideas

  • DNOs as Efficient Surrogates: The research demonstrates that DNOs can effectively learn the complex, time-dependent mechanical behavior of IPCs, acting as highly accurate and fast surrogates for computationally intensive FEA.
  • Sequence-to-Sequence Learning for Dynamics: Applying sequence-to-sequence learning to DNOs is crucial for capturing the temporal dependencies inherent in dynamic loading scenarios, leading to improved prediction capabilities.
  • Incremental Learning for Composites: The incremental learning approach is particularly well-suited for composite structures where individual elements share underlying physical laws. This allows for efficient transfer of learned knowledge between elements, significantly reducing training time.
  • Order Matters in Incremental Learning: The order in which elements are trained in the incremental learning process can impact performance. Training elements with similar material properties sequentially leads to smoother loss curves and potentially faster convergence.
  • Robustness to Noise: The trained DNO model exhibits good robustness to Gaussian white noise in the training data, maintaining reasonable accuracy even with moderate levels of contamination.

Key Results

  • Speedup: The proposed DNO model with sequence-to-sequence training and incremental learning achieves a training speedup of up to 200 times compared to traditional DeepONet models implemented in TensorFlow. Specifically, it's about 10 times faster for force prediction and significantly faster for stress distribution prediction.
  • Accuracy:
    • Force Prediction: Average prediction accuracy for transient force response is 98.04%, with a best accuracy of 99.58% on test data.
  • Stress Distribution: Average predicted accuracy for cross-section stress distribution is 98.18%.
  • Time-Extended Prediction: The model can predict mechanical responses for time durations longer than the training data (e.g., 2 seconds prediction from 1 second training) with reasonable accuracy (e.g., 89.56% for random Gaussian process input).
  • Robustness: Even with a noise level of β = 0.2, the DNO model maintains an average prediction accuracy of 86.48%.

Limitations and Future Work

  • Time-Extended Prediction Limitations: While the model shows extensibility, its ability to predict significantly longer time sequences than trained on has limitations, especially when faced with inputs exhibiting higher frequency amplitude changes.
  • Noise Impact: Although robust, prediction accuracy decreases with increasing noise levels. For very noisy data, more advanced techniques might be needed.
  • Future Work: The authors suggest exploring other recent methods like Bayesian neural networks for more reliable prediction results with noisy data. Further research into optimizing the incremental learning route for even greater efficiency is also implied.

Summary updated

BibTeX
@misc{lu2023deep,
  author = {Minglei Lu and Ali Mohammadi and Zhaoxu Meng and Xuhui Meng and Gang Li and Zhen Li},
  title = {{Deep neural operator for learning transient response of interpenetrating phase composites subject to dynamic loading}},
  year = {2023},
  eprint = {2303.18055},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2303.18055},
}
Figure