Organizations: Seoul National University, Seoul, Republic of Korea · KAIST, Daejeon, Republic of Korea · Samsung Electronics, Hwaseong, Republic of Korea
Virtual metrology (VM) aims to estimate costly or delayed metrology variables from heterogeneous sensor signals and in-line measurements, and it acts as a critical component of semiconductor manufacturing. While recent data-driven models have improved VM accuracy, most are trained in batch settings and assume stationary data distributions. In practice, however, production environments evolve continuously due to equipment degradation, process adjustments, and maintenance events, making conventional batch-trained VM models difficult to maintain under temporal distribution shifts. In this work, we propose GRAPE (Global Regression with Adaptive Prediction Enhancement), a practical online learning framework for VM in evolving wafer-processing environments. GRAPE integrates a stable global predictor based on interpretable representations with two lightweight streaming adapters that correct feature-level and environment-level shifts through online updates. Experiments on two real-world semiconductor manufacturing datasets, including chronologically held-out blind-test streams, show that GRAPE improves robustness and generalization under real temporal drift compared with both conventional batch-learning methods and existing online learning baselines.
Figures & tables
Figure 1: Distinct properties of industrial VM data. (a) Heterogeneity: different sensing modalities exhibit distinct signal patterns and sampling rates. (b) Irregularity: sensor observations are recorded at non-uniform time intervals across sensors.
Figure 2: Representative temporal distribution shifts in industrial VM data. (a) The trajectories of representative sensors, colored by recording time, show substantial covariate shift. (b) Distribution of prediction targets by recording month shifts toward larger values over time, illustrating temporal label shift.
Figure 3: Overall framework of Grape . For each wafer i , the global model fθ∗g produces an initial prediction based on target-aware representations. Adapter I performs feature-level adaptation by adjusting the contributions of shift-sensitive features, while Adapter II corrects prediction offsets at the environment level. The final prediction is obtained by combining the outputs of these components, and the adapters are updated online to continuously adapt the model to temporal distribution shifts.
Feature Group
Description
Global/window statistics
Mean, min, max, std of sensor signals.
Temporal dynamics
Rising time, peak time, settling time.
Shape characteristics
Max gradient, monotonicity, curvature.
Frequency-domain features
Dominant frequency, top FFT magnitudes.
Cross-sensor correlations
Mean correlation across sensor groups.
Window transitions
Changes in sensor statistics across windows.
Table I: Target-aware feature groups derived from wafer sensing signals.
Method
D1
D2
Avg. Rank
RMSE
MAE
R2
RMSE
MAE
R2
Batch Methods
Linear
0.1712
0.1293
0.1256
1.0920
0.8980
-0.4827
5.67
MLP
0.1715
0.1271
0.1231
1.0483
0.8532
-0.3664
5.33
DLinear
0.1860
0.1441
-0.0311
0.9313
0.7390
-0.0783
6.50
Transformer
0.1823
0.1389
0.0095
0.8784
0.6840
0.0406
5.17
Table II: Test performance on two semiconductor VM datasets. Grape achieves the best average rank across various metrics and datasets, demonstrating improved robustness compared with both batch and online learning approaches.
Method
D1
D2
Avg Rank
RMSE
MAE
R2
RMSE
MAE
R2
Batch Methods
Linear
1.2968
0.9080
-0.5966
1.1381
0.9469
-0.6078
6.50
MLP
1.2924
0.9076
-0.5858
1.1511
0.9599
-0.6447
6.50
DLinear
1.3245
0.9394
-0.6612
0.9828
0.7893
-0.2313
7.00
Transformer
1.3015
0.9223
-0.6041
0.8689
0.6854
0.0377
6.00
Table III: Performance under a blind-test setting on two semiconductor VM datasets. Grape achieves the best performance, indicating improved generalization. Batch methods degrade substantially, while online learning methods remain markedly stronger.
Figure 4: Case-study examples from data streams under temporal shift: (a) D1 and (b) D2.
Method
D1
D2
Avg Rank
RMSE
MAE
R2
RMSE
MAE
R2
Global-MLP
0.1649
0.1252
0.1887
0.8076
0.6485
0.1891
2.00
Global-kNN
0.1651
0.1254
0.1874
0.8321
0.6692
0.1392
3.50
Adapter-I-RLS
0.1695
0.1290
0.1430
0.8098
0.6350
0.1847
3.33
Adapter-II-None
0.1718
0.1312
0.1196
1.0993
0.9012
-0.5025
5.00
Grape (ours)
0.1647
0.1270
0.1910
0.7658
0.5990
0.2709
1.17
Table IV: Ablation study on the test split, which shows the contribution of each component in Grape . The full Grape model generally achieves the best results.
Method
D1
D2
Avg Rank
RMSE
MAE
R2
RMSE
MAE
R2
Global-MLP
0.7340
0.5118
0.4885
0.8149
0.6378
0.1758
3.00
Global-kNN
0.7297
0.5091
0.4945
0.8225
0.6438
0.1602
3.00
Adapter-I-RLS
0.7660
0.5223
0.4430
0.8104
0.6294
0.1848
3.00
Adapter-II-None
1.2331
0.8625
-0.4435
1.1309
0.9171
-0.5875
5.00
Grape (ours)
0.2099
0.1540
0.9582
0.7962
0.6165
0.2132
1.00
Table V: Ablation study on the blind-test split. The full Grape model shows significant performance improvements over the four variants, especially on D1.
Chemical mechanical planarization (CMP) is a critical semiconductor manufacturing process where material removal rate (MRR) is a key metric for process monitoring and advanced process control (APC). Virtual metrology enables MRR estimation from process time series signals, supporting run-to-run(R2R) control and process optimization while reducing reliance on physical metrology. Existing approaches typically rely on simplified linear assumptions or require handcrafted feature engineering. In this paper, we propose Global Regression Pretraining(GREP), a self-supervised, attention-based framework for CMP MRR prediction. GREP learns temporal representations from unlabeled sensor signals through global proxy-target regression, followed by supervised fine-tuning for MRR estimation. The framework is evaluated against a feature-engineered supervised baseline under three conditions: limited labeled data, sensor noise, and temporal process drift. Experiments on the PHM 2016 CMP dataset demonstrate competitive performance under limited labeled-data conditions, with improvements in selected cases and performance largely comparable to the baseline across different labeled-data fractions. Both models exhibit relatively small changes in prediction error with increasing noise, while GREP achieves a 17.0% RMSE reduction in the final temporal interval.
Jonathan Adam Rico, Nagarajan Raghavan, Senthilnath Jayavelu
Engineering Product Development, SUTD, Singapore · Institute of Advanced Intelligence and Computing, A*STAR, Singapore · Dept. of Electrical and Computer Engineering, NUS, Singapore
Online adaptation can help edge time-series forecasting under distribution drift, but its measured benefit is sensitive to evaluation choices. We study six public multivariate streams, including building-sensor and smart-meter data, under a leakage-free streaming protocol. We identify two additional sources of comparison bias. First, the warmup budget of the static baseline has a two-sided effect: insufficient warmup undertrains the baseline, whereas excessive warmup can degrade its pre-drift generalization. Across six dataset-backbone settings, the estimated adaptation benefit changes by 3.0 to 18.8 percentage points (pp) over the 1,000-20,000-step warmup range. Second, comparing SGD with momentum (SGD+m) and Adam at a shared default learning rate conflates optimizer quality with rate sensitivity. We select both the warmup budget and each optimizer's online rate using a held-out pre-drift validation slice without accessing test data. Under this validation-only procedure, Adam outperforms SGD+m in 310 of 360 evaluated cells, while 4 Adam cells remain below the static baseline. We further characterize accuracy against adaptation-state memory and A100-measured per-update latency for full, head-only, and calibration-based adaptation. In the evaluated PatchTST frontier settings, several parameter-efficient variants are nondominated on the adaptation-state-memory axis. Smart-meter analyses also show that reported gains depend on meter-selection rules. These findings support a validation-only commissioning procedure, while target-device latency and energy remain to be measured. Code, data, and all reported numbers: https://github.com/keiotakmin/tsf-edge-adaptation.
Takumi Fujimoto, Hiroaki Nishi
School of Science for Open and Environmental Systems, Graduate School of Science and Technology, Keio University, Yokohama, Japan
Accurate workload forecasting is critical for elastic resource provisioning in web-scale cloud services, where distribution shifts driven by viral content, product launches, and user behavior degrade offline-trained models rapidly. Naive online learning recovers accuracy but incurs prohibitive per-step compute cost. We propose AdaptLSTM, an adaptive online framework that detects drift via validation-calibrated thresholds and applies selective, targeted updates. On the Alibaba Machine Trace, AdaptLSTM recovers 54% of Naive Online's improvement at 20% cost (2.7× efficiency, p=0.002 over 10 seeds). On the more volatile Container Trace, it achieves 96% at 20% cost (4.8× efficiency, +75% MAE reduction over Static). Unlike classical drift detectors (ADWIN, DDM, Page-Hinkley) which fail to trigger on regression-scale error streams, AdaptLSTM fires 42 times over 301 steps and outperforms matched-budget baselines. Wall-clock profiling shows 1.33× throughput gain and 45% update-time reduction. The framework is model-agnostic: identical Pareto patterns hold for LSTM, GRU, and Transformer backbones.
Xinhua Miao, Bowei Yang, Zhengong Cai
Zhejiang University Ningbo, China · Zhejiang University Hangzhou, China