以物理引導之生成式AI預測RC橋柱破壞模式與遲滯行為

Abstract

對地震工程而言，鋼筋混凝土(Reinforcement Concrete, RC)橋柱的遲滯行為、損壞機制及結構性能指標如損害指數（Damage Index, DI）可以提供工程師了解橋柱在反覆載重下的反應行為，因此如何在設計階段時獲取橋柱的遲滯行為及損壞機制是當今重要的研究課題。本研究提出了一個基於閘門循環單元（Gated Recurrent Unit, GRU）的遲滯迴圈預測網路(HysGRU)，旨在彌平有限元模擬和真實實驗之間的差異。經由實驗，在模擬所提供的物理限制下，HysGRU 達到了31.51 kN 的均方根誤差，相較於純粹數據驅動的方法和模擬所獲得之誤差，分別降低了24％和62％。此外，HysGRU 於預測遲滯迴圈時，也會一併提取每個層間位移角(Drift Ratio)下所對應之潛在特徵，於此研究中，此潛在特徵作為所提出的損傷模式生成模型HysGAN之條件，藉由輸入不同時間點之潛在特徵以控制HysGAN生成每個層間位移角下相對應的損傷影像。為了評估 HysGAN的模型表現，此研究亦額外提出了一個損傷指標預測網路，透過預測生成之損傷影像對應之損傷指標， 計算預測與真實值之間的R^2分數，藉此評估模型生成表現。HysGAN 在驗證和測試數據集上皆可達到0.92 的R^2分數，相對於基準模型，分別增加了 0.05 和 0.55，顯示出經由潛在特徵的控制可以有效提升模型的通用性及生成影像之品質。此研究中所提出之橋柱損害預測框架可以提供工程師根據設計預定之柱設計參數，預測設計橋柱之遲滯行為，並可以透過選定特定層間位移角，觀察橋柱破壞形式，使工程師可以更有效評估設計橋柱在反覆載重下之結構行為。

Seismic attributes of reinforced concrete (RC) bridge columns, including hysteretic response, damage mechanism, and structural performance measures such as the Damage Index (DI), are pivotal to the domain of seismic engineering. In this study, a sequence-to-sequence hysteresis prediction network based on Gated Recurrent Units (GRU), HysGRU, is proposed to bridge the discrepancy between finite element simulation and experiment. Guided by the physics constraints extrapolated from simulation, HysGRU achieves a root mean square error (RMSE) of 31.51 kN, representing a reduction of 24% and 62% compared to the purely data-driven methodology and simulation, respectively. Moreover, during the prediction of HysGRU, the physics features of hysteretic behavior are extracted as a condition of the proposed damage pattern generative model, HysGAN, which is based on a conditional generative adversarial network (CGAN). To evaluate the performance of HysGAN, a DI prediction network is proposed to assess R^2 scores by the DI prediction of the synthetic damage patterns. Comprehensive experimental evaluations have demonstrated that the integration of latent feature conditioning substantially enhances the quality, stability, and generalization performance. HysGAN attains R^2 scores of 0.92 for both the validation and test datasets, representing significant improvements over the baseline model, with increases of 0.05 and 0.55, respectively. The proposed prognostic framework enables engineers to delineate the hysteresis loop and damage patterns based on predefined column design parameters and chosen drift ratios, facilitating an efficient assessment of seismic design and structural capacity.