As semiconductor manufacturing processes advance to smaller scales, devices become more vulnerable to defects that reduce performance. Wafers transported in front opening unified pods (FOUPs) are exposed to moisture in the airflow from the fan filter unit (FFU) when mounted to the equipment front end module (EFEM). This moisture reacts with airborne molecular contaminants (AMCs), forming harmful particles that cause wafer defects. Therefore, understanding internal FOUP flow structures is crucial for preventing moisture-induced contamination. Three wafer loading configurations, i.e., odd-layer loaded (OLL), top-loaded (TL), and bottom-loaded (BL), were investigated using magnetic resonance velocimetry (MRV) and computational fluid dynamics (CFD). Both qualitative and quantitative analyses were conducted to determine how wafer protection varies with different loading configurations. MRV measurements showed the load port purge (LPP) flow impinging on the FOUP ceiling and forming an outward flow exiting toward the EFEM at the top. Additionally, CFD results showed that the volumetric fraction of LPP flow in the upper region of the FOUP progressively increased over time. Therefore, wafers loaded in the upper region were found to receive enhanced protection against moisture-laden EFEM flow. This work proposes a strategy to protect wafers from defects by optimizing wafer loading configurations.
The electrically active defects at the InGaP/Al2O3 interface, representative of an AlInGaP red micro‑light‑emitting diodes (micro‑LED) quantum well sidewall, are profiled using planar metal/insulator/semiconductor (MIS) stacks. The planar structure enables a simple, material selective characterization and allows replicating the micro‑LED fabrication surface treatments. Multifrequency capacitance‑voltage (CV) and conductance‑voltage (GV) analysis reveals shallow InGaP traps and Al2O3 border traps. The effect of the surface treatment (no treatment, HCl wet etching, and BCl₃/Ar plasma dry etching) on the detected traps is reported.
5G/6G communication evolution and Moore’s law scaling impose stringent device–circuit co-optimization, balancing intrinsic device characteristics and circuit performance. We propose a multi-level framework integrating Design of Experiments (DoE), artificial intelligence (AI), and response surface methodology (RSM). At device level, for each device parameter, separate Gradient Boosting Regression (GBR) and Light Gradient Boosting Machine (LightGBM) models are trained on a dataset built from a compact experimental design; their decision boundaries are approximated by an RSM surrogate, yielding interpretable logic that predicts process parameters for target specifications. At cross-platform level (TCAD-to-SPICE), a human–machine approach combining prompt engineering with the DeepSeek large language model assists in generating data-transfer scripts, accelerating SPICE parameter extraction and circuit simulation. For N-channel RF MOSFETs, we co-optimize cutoff frequency (fT), breakdown voltage (BVDS), off-state current (Ioff), and saturation drain current (Id(sat)) using TCAD, extract BSIM4 parameters, and validate with class-AB power amplifiers. Compared to industry benchmarks, the framework improves Id(sat) by 32.8% to 797 μA/μm without compromising fT, BVDS, or Ioff. Device–circuit co-optimization increases RF power gain by 46% without area overhead and improves P1dB at 3.4 GHz by 28%, with negligible third-order intermodulation distortion (IMD3) penalty. These improvements directly benefit 5G/6G coverage. Multi-model comparisons uncover physical principles; this methodology generalizes to multi-level co-optimization.
Integrated-circuit test equipment operates under multiple programs and electrical conditions, making equipment-induced variation difficult to separate from legitimate device and operating-condition variation. This paper presents a revised closed-loop perception–diagnosis–compensation framework for automatic test equipment (ATE) stability management. The framework combines a Gaussian mixture model (GMM) for operating-condition perception, cluster-referenced residual-domain particle-swarm-optimized Kalman filtering (PSO-KF), Spearman-weighted batch indicators, confidence-guided batch re-estimation (CGBR), quadrant diagnosis, and low-confidence retest routing. The complete evaluation uses 44,509 industrial IC test samples with 149 test items and a leakage-free chronological split of 26,705 training, 6,676 validation, and 11,128 testing samples. A two-component full-covariance GMM is selected by BIC. Under batch point stress, validation-calibrated linear drift, and validation-calibrated random walk, confidence-guided diagnosis returns 33/33, 20/22, and 18/22 displaced batches to the clean Q1 diagnostic region, corresponding to 100%, 90.9%, and 81.8% diagnostic recovery. Measurement-level RMSE reductions remain moderate at 0.62%, 2.27%, and 3.14%, respectively. The results therefore support improved diagnostic consistency rather than direct production-yield recovery. Low-confidence cases remain subject to retest, and no false-pass, defect-coverage, or outgoing-quality claim is made.
This work presents a comparative framework for target-driven plasma etch-rate optimization using a neural-network process surrogate trained on physics-inspired synthetic data incorporating sample-level noise and batch drift across seven process variables. The surrogate achieved a test coefficient of determination of R2 = 0.990. Quantitative comparison with first-order linear and quadratic-plus-interaction models yielded test R2 values of 0.676 and 0.920, respectively, supporting the use of a nonlinear surrogate for the modeled response. Four optimization strategies, differential evolution (DE), particle swarm optimization (PSO), Gaussian-process Bayesian optimization (BO/GP), and limited-memory Broyden-Fletcher-Goldfarb-Shanno with bounds (L-BFGS-B), were evaluated for nine target etch rates from 50 to 450 Å/min. L-BFGS-B matched all targets to numerical precision at the reported resolution with a mean optimization time of 0.075 s, while PSO provided the strongest derivative-free performance, with a mean absolute optimization error of 0.264 Å/min and mean time of 0.315 s. Each optimized recipe was then independently evaluated using the deterministic process model to distinguish surrogate-level target matching from modeled-process performance. Despite near-zero surrogate-level error, L-BFGS-B exhibited a mean process-model validation error of 13.646 Å/min, demonstrating that increasingly precise surrogate optimization does not necessarily yield proportionally improved process-level accuracy. The results identify surrogate fidelity as an important limiting factor once the inverse optimization problem is solved with high accuracy and support a modular framework for rapid plasma-etch recipe optimization and eventual closed-loop process control.
