Courtesy of D. Wang

---

---

Reconstruction of force networks from experimental data is rather complicated. It involves solving an inverse problem whose solution is not analytically tractable. In this project, we investigate different deep-learning architectures for carrying out force reconstruction in photo-elastic granular materials. We leverage the power of convolutional neural networks to classify complex spatial patterns exhibited by photo-elastic particles and use them to predict the forces acting between the particles [1]. The main drawback of using neural networks is that training them usually requires a large labeled data set, which is hard to obtain experimentally. We showed that this problem can be successfully circumvented by pre-training the networks on an extensive synthetic data set and then fine-tuning them on much smaller experimental data sets.

To make our method available, we developed well-documented and easy-to-use software for constructing force networks from digital data [2]. By taking advantage of machine learning techniques' computational efficiency and adaptability, the software outperforms traditional algorithms in the quality of the reconstruction and speed.