Critical samples focusing based intelligent optimization of shear thickening polishing process

Shear thickening polishing (STP) is an ultra-precision finishing method. The complex rheological behavior of shear-thickening fluids and the interaction of polishing parameters result in a highly nonlinear relationship between process parameters and surface quality after polishing. Neural network models have been widely applied to nonlinear system modeling due to their powerful nonlinear fitting capabilities. However, conventional modeling approaches typically treat all training samples equally, which limits the model's performance in critical sample regions. In this study, a training strategy based on Critical Samples Focusing (CSF) was proposed for surface roughness ( S _a ) prediction modeling, and its effectiveness was validated through STP experiments of quartz glass. The proposed strategy assigns higher training weights to samples of greater importance, as determined by the objectives of process optimization, thereby making the model more suitable for optimization-oriented tasks. Experimental results demonstrated that, compared with conventional approaches, the proposed strategy reduced the Primary-MSE (MSE for Critical Samples Region) from 32.13 to 1.80, and the Primary-MAPE (MAPE for Critical Samples Region) from 1.88 to 0.16. The optimized polishing parameters with 16wt% abrasive concentration, pH 12, and a polishing speed of 86 rpm—yielded a superior surface finish with an S _a of 0.85 nm, significantly outperforming the minimum S _a of 2.50 nm achieved by conventional experimental methods. The proposed modeling method is considered to be broadly applicable, offering a novel data-driven approach and experimental foundation for ultra-precision machining.