Frequency-spatiotemporal synergistic network for high-precision semiconductor wafer dicing quality prediction

Semiconductor wafer dicing quality prediction requires micrometer-level precision, yet faces three critical challenges: scarce labeled data, complex spatiotemporal dependencies, and fine-grained feature preservation. Traditional physics-based modeling approaches suffer from extensive parameter calibration and fail to capture nonlinear dynamics, limiting their applicability in practical manufacturing scenarios. To address these challenges, this paper propose FreStyle, the first deep learning framework specifically designed for semiconductor wafer dicing quality prediction. To capture fine-grained features, the framework constructs a frequency-spatiotemporal synergistic dual-branch architecture that overcomes the over-smoothing problem of conventional methods. Specifically, the spatiotemporal branch employs depthwise separable 3D convolutions and adaLN-Zero Transformer to model long-range spatiotemporal dependencies, while the frequency branch introduces complex-valued convolutions, constructs analytic signals via Hilbert transform and applies Fast Fourier Transform, with an Adaptive Radial Energy Partition (AREP) module designed to separate low- and high-frequency components and perform differentiated attention processing. To effectively fuse heterogeneous features, a Cross-Domain Consistency-Driven Bidirectional Interaction (CDCI) block is constructed to achieve deep complementary enhancement through an interaction mechanism where spatiotemporal dynamics guide frequency-domain aggregation and frequency-domain priors calibrate spatiotemporal responses. Moreover, we construct the first public semiconductor wafer dicing dataset (SCD) meeting industrial precision standards for realistic evaluation. Experimental results on the SCD dataset show that FreStyle achieves an MSE of 0.0028 and an SSIM of 0.8961, corresponding to a 26.3% reduction in MSE and a 1.3% improvement in SSIM compared to the best baseline method. Extensive experiments on WeatherBench, TrafficBJ, and NSE datasets further validate the generalization capability of the proposed framework.