Quality Prediction and Process Optimization for Silicon Wafer Diamond Wire Sawing Based on Simcenter AI Studio

Abstract

In diamond wire sawing (DWS) of photovoltaic silicon wafers, total thickness variation (TTV) and micro-cracks are difficult to predict online, and process optimization still depends largely on trial and error. This study takes the silicon wafer DWS dataset (solar_raw for experiments, solar_sim for simulation) officially released by the "Lightweight Design and AI Application" track of the national final of the 19th "Gaojiao Cup" National Undergraduate Advanced Graphics Technology and Product Information Modeling Innovation Contest, and carries out cleaning, merging, training, evaluation, and reverse optimization on the Simcenter AI Studio 2025.0.1 platform. Cleaning leaves 141 experimental samples and 152 simulation samples, from which an inner join yields 62 merged samples. On the 62-row merged table, Linear Regression (OLS) in AI Studio reaches a test-set RMSE of 2.775 μm (relative error 9.08%) under a 70/30 split, and the Model Simulator predicts TTV=19.956 μm for a given set of process inputs. A separate Python re-computation (scikit-learn/xgboost, 5-fold cross-validation, 7 regression models) on the 141-row experimental table again finds OLS the most accurate (RMSE=1.136 μm, R2=0.907), with feed_rate the dominant factor for TTV (r=+0.706); for micro-crack classification, Logistic Regression (LogReg) performs best (F1=0.527, AUC=0.816). A multi-fidelity ablation shows that adding finite-element simulation features (wire stress/deflection) does not improve prediction accuracy for this dataset under the small-sample condition. Reverse optimization on the platform (Find Optimal Input Settings, with three modes Specific Value / Minimize / Maximize) and on the Python side (Ridge surrogate + RF-crack-probability ε-constraint) both recommend the same low-TTV, low-crack process window; the platform's Top1 falls within the Python top-3, and the Top1 window cuts the predicted TTV by about 43.5% relative to the experimental-table mean. The results suggest that a no-code AutoML platform is a practical option for silicon wafer sawing process optimization.