Speaker
Description
In material science, 4D Scanning Transmission Electron Microscopy (4D STEM) is a dataset of images formed by electrons passing through a thin specimen with the electron beam focused on a fine spot [1], allowing material scientists to learn some structural properties. Oxley et al. showed that deep learning is powerful for distinguishing structures embedded within the data [2]. However, Oxley et al’s model relies on approximations requiring test examples to be within the training dataset distribution, and lacks the ability to directly learn spatial parameters. Qin et al. proposed a neural network structure called cycle-consistent-spatial-transforming auto-encoders (CC-ST-AE) to extract spatial parameters like crystallographic strain, shear, and rotation to understand material properties [3], [4]. However, to have a more precise result, we need to send more electron beams, which is likely to cause more damage to the specimen. To minimize the specimen’s damage, we could control the total fluence by reducing it in the convergent beam [5]. Unfortunately, this introduces noise into the collected data, making it harder to analyze. It is possible to reduce damage by collecting less precise data, if we enhance the model’s robustness.
Qin et al. trained CC-ST-AE on 4D STEM data at different noise intensity levels and compared results using Mean Absolute Error (MAE) of the strain along different directions, between CC-ST-AE and py4DSTEM, a widely used tool analyzing 4D STEM. However, they only used the ground truth with training noise as a base. Instead, We tested performance by comparing the test set with both noise-free and same level noisy bases. Firstly, to determine the maximum noise level that the neural network can handle, we have trained and tested the model with various intensities of the background noise [6], [7]. To analyze the performance of the model, we have compared the training loss, testing loss, and the angle difference between the prediction and input, which shows the accuracy of the rotation predicted. We noticed the increase of both training loss and testing loss as the intensity of the background noise increases. The testing loss with 25% testing noise increased by 187% from 1.3731 for noisy-free training to 2.5778 for 50% noise level training. Also, with the same training noise, 25%, the testing loss increased 284% from 0.8339 to 3.20835 by increasing the training noise from 10% to 50%. This data shows that as the noise level of the 4D STEM dataset increases, the CC-ST-AE model is not likely to perform well enough. Although some increase in test loss with higher training noise is expected, this data shows that we still need to make every effort to improve robustness. We will work on data augmentation with training with a combination of various noise levels to try and improve the model’s performance under high levels of noise. We will perform hyper-parameter tuning for the learning rate, seed, and batch size to further increase its robustness. Additionally, to make this suitable for edge deployment, we will explore techniques such as regularization, pruning, and quantization.
References
[1] I. 4D STEM — Gatan, “4d stem,” 2024, https://www.gatan.com/techniques/4d-stem [Accessed: 9-1-2024].
[2] M. P. Oxley, J. Yin, N. Borodinov, S. Somnath, M. Ziatdinov, A. R. Lupini, S. Jesse, R. K. Vasudevan, and S. V. Kalinin, “Deep learning of interface structures from simulated 4d stem data: cation intermixing vs. roughening,” Machine Learning: Science and Technology, vol. 1, no. 4, p. 04LT01, 2020.
[3] S. Qin, J. Agar, and N. Tran, “Extremely noisy 4d-TEM strain mapping using cycle consistent spatial transforming autoencoders,” in AI for Accelerated Materials Design - NeurIPS 2023 Workshop, 2023. [Online]. Available: https://openreview.net/forum?id=7yt3N0o0W9
[4] S. Wang, T. B. Eldred, J. G. Smith, and W. Gao, “Autodisk: Automated diffraction processing and strain mapping in 4d-stem,” Ultramicroscopy, vol. 236, p. 113513, 2022. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0304399122000481
[5] K. C. Bustillo, S. E. Zeltmann, M. Chen, J. Donohue, J. Ciston, C. Ophus, and A. M. Minor, “4d-stem of beam-sensitive materials,” Accounts of chemical research, vol. 54, no. 11, pp. 2543–2551, 2021.
[6] F. Croce, M. Andriushchenko, V. Sehwag, E. Debenedetti, N. Flammarion, M. Chiang, P. Mittal, and M. Hein, “Robustbench: a standardized adversarial robustness benchmark,” arXiv preprint
arXiv:2010.09670, 2020.
[7] J. Gilmer, N. Ford, N. Carlini, and E. Cubuk, “Adversarial examples are a natural consequence of test error in noise,” in Proceedings of the 36th International Conference on Machine Learning, ser. Proceedings
of Machine Learning Research, K. Chaudhuri and R. Salakhutdinov, Eds., vol. 97. PMLR, 09–15 Jun 2019, pp. 2280–2289. [Online]. Available: https://proceedings.mlr.press/v97/gilmer19a.html
| Focus areas | HEP |
|---|
