فصلنامه
نوع مقاله : مقاله پژوهشی
نویسندگان
1 گروه مهندسی کامپیوتر، دانشکده فنی و مهندسی، دانشگاه، اراک اراک، ایران
2 گروه طراحی کاربردی، دانشکده مهندسی مکانیک، دانشگاه صنعتی اراک، اراک، ایران
3 گروه مهندسی مکانیک، دانشکده فنی و مهندسی، دانشگاه خوارزمی، تهران، ایران
کلیدواژه‌ها
عنوان مقاله English
نویسندگان English
This study presents a unified three-dimensional convolutional neural network (3D CNN) for the direct prediction of the instantaneous lift coefficient in turbulent flow-induced vibration of a two-degree-of-freedom elastically mounted square cylinder. Physics-consistent data were generated using a validated two-way coupled URANS-FSI model with the k-ω SST turbulence closure. After excluding the initial transient response, 1017 synchronized wake-field snapshots and lift-coefficient values were collected over 101.7 s. Consecutive flow images were processed as a single spatio-temporal volume, allowing the Conv3D kernels to learn vortex morphology, shear-layer evolution, and short-term wake dynamics without recurrent units or reduced-order preprocessing. A chronological train-validation-test division was adopted to preserve temporal causality. Sensitivity analysis identified a sequence length of T = 18 and a batch size of 14 as the optimal configuration. The final network achieved R2train = 0.932 and R2test = 0.919, with approximate MAE = 0.19 and MSE = 0.06. Time-history comparisons showed accurate recovery of oscillation amplitude, extrema, zero crossings, and phase over the unseen test interval. These results demonstrate that direct three-dimensional convolution can provide an effective image-to-force surrogate for turbulent fluid-structure interaction.
کلیدواژه‌ها English
