本研究以離散元素法模擬高爐料倉與爐喉區域旋轉佈料行為，探討1/2縮尺模型取代全模型進行高爐佈料分析之可行性，並進一步整合romAI、Twin Activate與HyperStudy，建立爐料分布代理模型之快速預測及佈料矩陣最佳化篩選之架構。研究首先比較全模型與1/2縮尺模型之料流軌跡、料面輪廓、礦焦比、孔隙率與粒徑分布等特徵，並分析添加HBI與添加Pellet兩種含鐵爐料條件對料層結構之影響。其次，以EDEM模擬資料建立礦焦比與孔隙率之romAI代理模型，先採用11組模擬資料訓練集進行模型訓練，並以第12組模擬資料作為測試集，再增加5組模擬資料將訓練集擴充至16組，以評估訓練集增加對模型預測能力之影響。最後，透過MELS-DOE、GRSM及局部鄰近搜尋之三階段最佳化流程，分別針對礦焦比與孔隙率之目標徑向分布搜尋最佳佈料矩陣，並將所得矩陣以EDEM重新模擬以進行驗證。 模擬結果顯示，1/2縮尺模型可將旋轉滑槽佈料的模擬時間縮短至全模型之約1/3∼1/6，並能合理重現主要爐料流動、料面輪廓及礦焦比與孔隙率徑向分布趨勢。添加Pellet因小粒徑顆粒數量較多，不僅增加計算負荷，亦較容易填入大顆粒間孔隙，使整體孔隙率低於添加HBI。在代理模型方面，訓練集由11組增加至16組後，礦焦比之平均絕對百分比誤差由13.6%降至12.3%，孔隙率則由6.72%降至5.93%，顯示增加訓練集資料有助於改善代理模型預測能力。 在最佳化方面，經三階段最佳化後，礦焦比之目標函數(平均絕對百分比誤差， Objective_J)由25.33%降至14.33%，最佳化的礦焦比佈料矩陣為C=[3，4，2，4，2]而含鐵爐料佈料矩陣為Fe=[1，1，5，6，2]；孔隙率由3.39%降至1.69%，最佳化的孔隙率佈料矩陣為C=[2，1，4，5，3]與Fe=[3，3，1，3，5]。將上述最佳化矩陣重新代入EDEM進行模擬後，其Objective_J則分別為19.06%與1.86%。因此，本研究所建立之代理模型暨最佳化流程可有效縮小預測分布與目標分布之差異，並快速識別整體誤差較低之佈料矩陣。綜合而言，本研究所建立之縮尺DEM模型與代理模型最佳化流程，可有效降低高爐佈料分析之計算成本，並作為礦焦比與孔隙率快速預測及佈料矩陣篩選之工具。 關鍵字:高爐、旋轉佈料、離散元素法、縮尺模型、料層分布、代理模型、最佳化分析。;This study uses the Discrete Element Method (DEM) to simulate burden flow in blast furnace storage bins and rotating-chute charging in the furnace throat region. The feasibility of using a half-scale model as a substitute for the full-scale model in blast furnace burden distribution analysis is investigated. Furthermore， romAI， Twin Activate， and HyperStudy are integrated to establish a framework for rapid prediction using burden-distribution surrogate models and optimization-based screening of charging matrices. First， the full-scale and half-scale models are compared in terms of burden trajectories， burden surface profiles， ore-to-coke ratio， porosity， and particle-size distribution. The effects of two ferrous burden conditions， namely HBI-added and pellet-added cases， on the burden structure are also investigated. Subsequently， EDEM simulation data are used to construct romAI surrogate models for the ore-to-coke ratio and porosity. Eleven simulation cases are initially used as the training set， while the twelfth simulation case is used as the test set. Five additional simulation cases are then added to expand the training set to 16 cases， thereby evaluating the effect of increasing the training data on model prediction performance. Finally， a three-stage optimization procedure consisting of MELS-DOE， GRSM， and local neighboring search is applied to identify optimal charging matrices for prescribed radial distributions of the ore-to-coke ratio and porosity. The resulting matrices are subsequently re-simulated using EDEM for validation. The simulation results show that the half-scale model reduces the computational time required for rotating-chute charging simulations to approximately one-third to one-sixth of that required by the full-scale model， while reasonably reproducing the major burden flow behavior， burden surface profile， and radial distribution trends of the ore-to-coke ratio and porosity. Compared with the HBI-added case， the pellet-added case contains a larger number of fine particles， which not only increases the computational load but also allows the fine particles to more readily fill the voids between larger particles， resulting in a lower overall porosity. Regarding the surrogate models， when the training set is increased from 11 to 16 cases， the mean absolute percentage error of the ore-to-coke ratio model decreases from 13.6% to 12.3%， while that of the porosity model decreases from 6.72% to 5.93%， indicating that increasing the amount of training data improves the predictive performance of the surrogate models. Regarding optimization， after the three-stage optimization procedure， the objective function， defined as the mean absolute percentage error and denoted by Objective_J， for the ore-to-coke ratio decreases from 25.33% to 14.33%. The optimized coke charging matrix is C=[3，4，2，4，2]， while the optimized ferrous burden charging matrix is Fe=[1，1，5，6，2]. For porosity， Objective_J decreases from 3.39% to 1.69%， with the optimized charging matrices C=[2，1，4，5，3] and Fe=[3，3，1，3，5]. When these optimized matrices are re-simulated using EDEM， the corresponding Objective_J values are 19.06% and 1.86%， respectively. Therefore， the surrogate-model-based optimization procedure developed in this study can effectively reduce the discrepancy between the predicted and target distributions and rapidly identify charging matrices with lower overall errors. Overall， the developed half-scale DEM model and surrogate-model-based optimization procedure can effectively reduce the computational cost of blast furnace burden distribution analysis and serve as a tool for rapid prediction of ore-to-coke ratio and porosity distributions as well as charging-matrix screening. Keywords: Blast furnace， Rotating chute， Discrete Element Method， Scaled model， Burden distribution， Surrogate model， Optimization analysis.
