International Journal of Innovative Research in Computer and Communication Engineering
ISSN Approved Journal | Impact factor: 8.771 | ESTD: 2013 | Follows UGC CARE Journal Norms and Guidelines
| Monthly, Peer-Reviewed, Refereed, Scholarly, Multidisciplinary and Open Access Journal | High Impact Factor 8.771 (Calculated by Google Scholar and Semantic Scholar | AI-Powered Research Tool | Indexing in all Major Database & Metadata, Citation Generator | Digital Object Identifier (DOI) |
| TITLE | Automated Optical Inspection (AOI) and AI-Based Defect Classification in Silicon Solar Cell Manufacturing: Inline Implementation and Yield Correlation |
|---|---|
| ABSTRACT | This bulletin entry presents a 12-month retrospective on the Automated Optical Inspection (AOI) and AI-based defect classification program at the ES Foundry one-gigawatt PERC manufacturing facility in Greenwood, South Carolina. The retrospective spans the period from the initial AOI v1.0 release in March 2025 through the current production v4.5 release in March 2026, covering three distinct phases of capability evolution: Phase 1 (single-modal AOI, v1.0 → v3.2, Mar 2025 - Aug 2025), Phase 2 (multimodal AOI+EL fusion, v4.0, Sep 2025 - Nov 2025), and Phase 3 (tri-modal AOI+EL+PL fusion with generative augmentation, v4.5, Dec 2025 - present). A companion bulletin in August 2025 documented Phase 1 in detail; the present work extends the engineering account through Phases 2 and 3 and records the operational outcomes of the multimodal and generative-augmentation methodologies that distinguish the v4.5 production system from its single-modal predecessor. The principal evolution from v3.2 to v4.5 has been the transition from a single-modal classifier operating on visible-spectrum AOI images to a tri-modal classifier jointly operating on AOI (visible morphology), EL (carrier recombination), and PL (sub-surface defect signatures) images. The tri-modal classifier achieves 99.4 percent top-1 classification accuracy on a 24,860-image production validation set, against 97.4 percent for the v3.2 single-modal classifier on a 14,420-image set. The accuracy gain is concentrated in three categories of defect that single-modal classifiers cannot fully resolve: subsurface cracks invisible to AOI, edge shunts producing only weak EL signatures, and emerging PID-risk patterns visible only through PL contrast. The accuracy gain comes at the cost of a modest median latency increase (380 ms → 450 ms), well within the 1-second cadence target. ◆ v4.5 top-1 accuracy: 99.4% on a 24,860-image production validation set across 16 defect classes - a 2.0 percentage-point absolute improvement over v3.2. ◆ AUC trajectory: 0.872 (v1.0, Mar 2025) → 0.987 (v3.2, Aug 2025) → 0.996 (v4.5, Mar 2026), supported by training-set growth from 12,000 to 124,000 labeled images. ◆ Multimodal fusion: AOI + EL + PL three-channel fusion adds 4 new defect classes (PID risk, subsurface crack, edge shunt, bulk shunt) over v3.2’s 12-class taxonomy. ◆ Synthetic augmentation: Diffusion-model-generated training images for rare classes lift per-class F1 by 60–95 percent, reducing labeling burden for naturally-rare defects. ◆ Active learning: Uncertainty-driven sample selection achieves target validation AUC 0.985 with 62 percent less labeling effort than random sampling. ◆ Federated training: Joint training across Greenwood, Phase-2 (commissioning), and Mexico sites lifts per-site AUC by 0.005–0.020 absolute without raw image sharing. ◆ Yield trajectory (12-month): Top-bin yield 17.2% (Mar 2025) → 22.8% (Jul 2025) → 24.5% (Jan 2026) → 25.7% (Mar 2026) - cumulative 8.5 percentage-point absolute uplift. ◆ Annualized commercial value: Approximately $15–18 million per year run-rate in Phase 3, against cumulative deployment investment of $9.4 million - sustained payback under 8 months across all phases. ◆ Inference latency: Median 450 ms (v4.5 tri-modal) versus 380 ms (v3.2 single-modal). Both well within the 1-second per-wafer cadence target. p95 below 720 ms. ◆ Capability maturity: v4.5 reaches ≥ 0.85 maturity across all 8 capability dimensions (multimodal, synthetic, active, federated, latency, interpretability, edge-case, adoption) - v3.2 reached ≥ 0.85 on only two. |
| AUTHOR | SEKHAR TATINENI Vice President, Technology, Greenwood, South Carolina, USA |
| VOLUME | 183 |
| DOI | DOI: 10.15680/IJIRCCE.2026.1404148 |
| pdf/148_Automated Optical Inspection (AOI) and AI-Based Defect Classification in Silicon Solar Cell Manufacturing Inline Implementation and Yield Correlation.pdf | |
| KEYWORDS | |
| References | [1] Ho, J., Jain, A., & Abbeel, P. (2020). Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33, 6840–6851. [2] Saharia, C., Chan, W., Saxena, S., et al. (2022). Photorealistic text-to-image diffusion models. NeurIPS 2022. [3] McMahan, B., Moore, E., Ramage, D., et al. (2017). Communication-efficient learning of deep networks from decentralized data. AISTATS 2017, 1273–1282. [4] Settles, B. (2010). Active learning literature survey. University of Wisconsin-Madison Computer Sciences Technical Report. [5] Tan, M., & Le, Q. V. (2019). EfficientNet: Rethinking model scaling for convolutional neural networks. ICML 2019, 6105–6114. [6] He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. IEEE CVPR Proceedings, 770–778. [7] Hu, J., Shen, L., & Sun, G. (2018). Squeeze-and-excitation networks. IEEE CVPR Proceedings, 7132–7141. [8] Vaswani, A., Shazeer, N., Parmar, N., et al. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30. [9] Lundberg, S. M., & Lee, S. I. (2017). A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems, 30, 4765–4774. [10] Deitsch, S., Christlein, V., Berger, S., et al. (2019). Automatic classification of defective photovoltaic module cells in electroluminescence images. Solar Energy, 185, 455–468. [11] Mehta, V., Pradhan, S., & Gupta, R. (2021). Application of computer vision and machine learning in solar manufacturing. Solar Energy, 213, 86–104. [12] Trupke, T., Bardos, R. A., Schubert, M. C., & Warta, W. (2006). Photoluminescence imaging of silicon wafers. Applied Physics Letters, 89(4), 044107. [13] Bedrich, K., Bokalic, M., Bliss, M., Topic, M., & Betts, T. R. (2018). Electroluminescence imaging of PV devices: Advanced vignetting calibration. IEEE Journal of Photovoltaics, 8(5), 1297–1304. [14] Tatineni, S. (2025). Automated Optical Inspection and AI-Based Defect Classification (Phase 1). Solar Manufacturing Forum, 7(3). [15] Tatineni, S. (2024). Defect Root Cause Methodology in High-Volume Solar Cell Manufacturing. Solar Manufacturing Engineering Review, 6(3). [16] Tatineni, S. (2025). Digital Twin Implementation for Solar Cell Process Lines. ES Foundry Technical Brief Series, TB-2025-01-027. [17] Tatineni, S. (2024). Electroluminescence-Based Inline Quality Gating in a US PERC Solar Cell Fab. Journal of PV Manufacturing Engineering, 11(11). [18] Tatineni, S. (2026). AI-Powered Real-Time Yield Forecasting in Silicon Solar Cell Manufacturing. AI Yield Forecasting White Paper Series, WP-2026-01-022. [19] Preu, R., Lohmüller, E., Lohmüller, S., et al. (2020). Passivated emitter and rear cell - devices, technology, and modeling. Applied Physics Reviews, 7(4), 041315. [20] Dullweber, T., & Schmidt, J. (2016). Industrial silicon solar cells applying the passivated emitter and rear cell concept. IEEE Journal of Photovoltaics, 6(5), 1366–1381. [21] Green, M. A., Dunlop, E. D., Hohl-Ebinger, J., et al. (2025). Solar cell efficiency tables (Version 65). Progress in Photovoltaics, 33(1), 3–15. [22] Bonawitz, K., Eichner, H., Grieskamp, W., et al. (2019). Towards federated learning at scale: System design. Proceedings of the 2nd SysML Conference. [23] Kingma, D. P., & Ba, J. (2015). Adam: A method for stochastic optimization. ICLR 2015. [24] Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press. [25] Inflation Reduction Act of 2022, Public Law 117-169. Section 45X Advanced Manufacturing Production Credit. [26] US Department of Energy (2025). Solar Energy Technologies Office: Annual Technology Baseline. NREL. |
