建構以影像為基底之光彈顆粒受力之深度學習模型及其於剪切流之初步應用

Abstract

光彈技術是一種非接觸式的應力測量方法，其因雙折射效應所產生之光彈性條紋的光強度場可用來量測二維顆粒流實驗中的應力場。但現今將光強度場對應至應力之校準方法不僅計算需求大，且準確度的不確定性高。本研究採用快速發展之深度學習方法建構出新型的光彈顆粒受力量測方法，我們以實驗光彈影像與電腦產生之模擬光彈影像訓練出兩種可對顆粒尺度進行受力量測的卷積神經網路模型，其可直接從光強度場預測顆粒尺度下的總受力與各分力的量值及受力角度，最終將此快速且高效之模型應用至剪切流的瞬時影像之分析。由預測施加在邊界的正向壓力之結果顯示深度學習模型的架構還需進行修正，但已嶄露深度學習模型輔助顆粒尺度下的應力量測的潛力。

To experimentally investigate the stress field of a 2D granular system, the photoelastic technique is a promising intrusive measurement method that links the stress magnitude with the light intensity field of photoelastic fringes. Unfortunately, the intensity-to-stress calibration is both case-dependent and computation-demanding with accuracy uncertainty. This work adopts the fast-growing deep learning strategy to establish a novel method for measuring forces acting on photoelastic disks. With the help of the experimental and computer-simulated photoelastic images, we have trained two convolutional neural network models capable of predicting total force, individual force components, and force angles based on the light intensity field at the particle scale. By utilizing these fast and efficient models, analyses were conducted on the instantaneous images of a plane photoelastic flow in simple plane shear. The result of predicting the pressures subjected to the boundary indicated the architecture of the models requires further refinement. However, it reveals the promising potential of analyzing the rheology of granular flow via artificial intelligence.