Inside the Sub-2nm Fab: Deconstructing the Indispensable Tools of the 3D IC Era

Introduction

For more than five decades, semiconductor progress was driven largely by planar scaling. Shrinking transistors and interconnects enabled successive generations of faster and more energy‑efficient devices. At dimensions below 2 nm, however, the assumptions that supported conventional scaling begin to break down. Quantum tunneling increasingly affects transistor behavior, while shrinking interconnect dimensions introduce substantial resistance and signal‑integrity challenges.

In response, the industry has shifted toward three-dimensional integration. Advanced packaging, chiplet architectures, gate-all-around nanosheet transistors, and vertically stacked memory structures such as 3D NAND all represent different expressions of the same strategy: extending integration into the vertical dimension rather than relying solely on lateral scaling [1].

This transition introduces a new set of manufacturing demands. Fabrication facilities must control material deposition, removal, alignment, and inspection across complex three-dimensional structures with tolerances measured in fractions of a nanometer [2]. Features buried deep within stacked architectures must be manufactured and inspected without compromising throughput or yield.

As a result, sub‑2 nm manufacturing cannot be achieved simply by refining existing process tools. It requires a coordinated set of advanced lithography, etch, deposition, metrology, inspection, and testing technologies that collectively enable high‑volume production [3].

1. Lithography: The Transition to High‑NA EUV

Lithography determines the minimum feature size that can be patterned on a wafer. For current leading-edge manufacturing, EUV systems operating at a wavelength of 13.5 nm and a numerical aperture (NA) of 0.33 have become the standard production platform.

Physical Constraints

At the sub‑2 nm node, required pitches fall below 26 nm. Under these conditions, conventional 0.33 NA EUV systems approach their practical resolution limits. Diffraction prevents sufficient collection of high-angle light from densely packed features, making single-exposure patterning increasingly difficult.

Manufacturers can compensate through EUV multi-patterning, where a single layer is divided across multiple exposure and etch steps. Although effective, this approach increases process complexity, manufacturing cost, and overlay sensitivity. It also amplifies stochastic defects arising from photon statistics and photoresist chemistry.

High‑NA EUV

High‑NA EUV addresses this limitation by increasing the numerical aperture from 0.33 to 0.55. Systems such as the ASML EXE:5000 series capture higher-angle diffracted light, extending single-patterning capability to approximately an 8 nm half-pitch [4].

The higher numerical aperture introduces additional optical challenges because light reaches the wafer at steeper angles. To avoid requiring substantially larger reticles, High‑NA systems employ anamorphic optics that magnify patterns asymmetrically [5]. This design preserves compatibility with existing mask infrastructure while improving resolution.

Fig. 1. A 3D isometric scientific schematic diagram of an advanced High-NA EUV lithography lens path, illustrating how standard circular light is compressed asymmetrically (4× on one axis, 8× on the other) through specialized lenses to resolve sub-10nm patterns without clipping the reticle mask. In the center, a series of sleek, anamorphic optical glass lenses are suspended in a setting. A bright, focused beam of Extreme Ultraviolet light (13.5nm wavelength) enters from the top. As it passes through the anamorphic lenses, the light ray is bent and compressed asymmetrically, changing from a circular beam into a precise, sharp ellipse. The elliptical beam projects onto a silicon wafer (gray), printing a series of ultra-dense, perfect parallel nanoscale lines (sub-10nm half-pitch).

Outlook

High‑NA EUV is widely viewed as a necessary technology for continued logic scaling below 2 nm. Current development efforts focus on managing anamorphic field stitching, improving overlay performance, and engineering photoresists capable of maintaining pattern integrity at extremely small dimensions.

2. Etch: High‑Aspect‑Ratio Cryogenic Processing

Three-dimensional integration depends heavily on deep vertical structures, including through-silicon vias and memory channel holes [6]. Creating these structures requires etching features with aspect ratios that may exceed 50:1 or even 100:1.

Physical Constraints

As aspect ratios increase, conventional reactive ion etching encounters aspect-ratio-dependent etching (ARDE) effects [7]. Ions lose directional energy through repeated sidewall interactions, while reaction byproducts become increasingly difficult to remove from deep trenches.

These transport limitations can produce incomplete etching, profile bowing, and directional deviations that compromise alignment with underlying structures [7]. The problem becomes particularly severe as feature dimensions continue to shrink while required depths increase.

Fig. 2. A schematic of semiconductor etching in silicon. On the left side, labeled “Standard RIE”: showing a deep micro-trench with chaotic red chemical radicals attacking the side walls. On the right side, labeled “Cryogenic HAR Etch”: showing a perfectly vertical, deep trench with ultra-straight walls, accented with subtle blue frost and ice cyan glow (indicating -100°C).

Cryogenic and Pulse‑Shaped RF Approaches

Advanced etch platforms from Lam Research, Applied Materials, and Tokyo Electron address these issues through a combination of cryogenic processing and sophisticated plasma control [8].

Cryogenic etching cools the wafer to temperatures below –100°C, suppressing unwanted chemical reactions along trench sidewalls. This promotes highly anisotropic material removal and improves profile control.

Pulse-shaped RF operation further enhances ion transport by reducing electron shading effects. By carefully synchronizing source and bias power, these systems maintain more uniform electric fields within deep structures, helping ions reach the bottom of the trench efficiently.

Outlook

Future generations of high-aspect-ratio processing are expected to incorporate atomic-layer etching sequences. Combining self-limiting surface chemistry with directional ion removal offers a path toward near-atomic control of profile formation while minimizing lateral erosion [9].

3. Deposition: Seam‑Suppressed Metal Fill

After vertical structures have been etched, they must be filled with conductive materials to form reliable electrical connections [10].

Fig. 3. Schematic illustrating the multi-stage mechanism of atomic layer nucleation followed by bottom-up chemical vapor deposition to achieve void-free vertical interconnects. Step 1: “Conformal ALD Nucleation” showing a single-atom-thick, uniform layer of bright metallic silver atoms (tungsten) perfectly lining the inner walls of a deep U-shaped trench. Step 2: “In-Situ Top Passivation” showing the upper lip of the trench with a protective inhibitor layer. Step 3: “Bottom-Up CVD Fill” showing the metallic tungsten filling cleanly from the bottom of the trench upward, leaving no empty spaces or seams.

Physical Constraints

Traditional chemical vapor deposition often struggles to fill narrow, high-aspect-ratio features. Because precursor concentration is highest near the trench opening, deposition tends to proceed more rapidly at the top than at the bottom.

This behavior can seal the opening before the feature is fully filled, trapping voids or seams within the deposited metal. Such defects increase resistance and may become sites of mechanical failure during thermal cycling.

Seam‑Suppressed Tungsten Deposition

Platforms such as Applied Materials’ Centura iSprint Seam‑Suppressed Tungsten system combine ALD and CVD techniques to improve fill quality [11].

The process begins with a conformal ALD nucleation layer that uniformly coats internal surfaces. Because ALD reactions are self-limiting, coverage remains highly consistent throughout the feature.

Subsequent surface treatments slow deposition near the opening, while carefully controlled CVD growth promotes filling from the bottom upward. The result is a substantially lower probability of void formation and improved electrical performance.

Outlook

As feature dimensions continue to shrink, interest is growing in ruthenium and cobalt deposition processes [12]. At nanometer scales, these materials may offer advantages over copper by reducing resistivity penalties associated with electron scattering and diffusion-barrier requirements.

4. Inspection and Metrology: Hybrid Optical and Multi‑Beam E‑Beam Strategies

Defects capable of causing device failure at the sub‑2 nm node may be only a few nanometers in size. Detecting such defects across high-volume production environments presents a major challenge.

Physical Constraints

Optical inspection systems are fundamentally limited by diffraction and cannot directly resolve features at the smallest relevant scales. Electron-beam systems provide substantially higher resolution but are generally too slow for comprehensive wafer inspection when used alone.

The industry therefore faces a persistent trade-off between throughput and sensitivity [13].

Hybrid Inspection Workflows

Modern inspection strategies combine multiple techniques within a coordinated workflow [13].

High-speed optical systems first screen entire wafers for potential anomalies. Machine-learning software then evaluates the resulting data, correlating inspection signatures with layout information to identify the most likely failure-inducing defects.

Multi-beam e-beam tools subsequently perform targeted high-resolution review of selected locations [14]. This hierarchical approach preserves throughput while maintaining the ability to identify extremely small defects.

Voltage-contrast inspection adds another layer of capability by revealing buried opens and shorts within three-dimensional structures through differences in charge accumulation behavior [15].

5. Test and Monitoring: Adaptive ATE with Embedded Telemetry

Manufacturing variation becomes increasingly significant as device dimensions shrink. Two chips produced from the same design may exhibit measurably different electrical characteristics due to small process variations.

Physical Constraints

Traditional testing methodologies rely on fixed guard bands that assume a common operating margin for all devices. At advanced nodes, this approach can unnecessarily reduce yield or fail to identify marginal devices that may experience reliability issues later in life.

At the same time, test complexity continues to increase, with projected test counts approaching one million measurements per chip.

Adaptive Testing Approaches

Adaptive automated test equipment, combined with embedded telemetry, offers a more flexible alternative [16].

Telemetry circuits integrated within the chip continuously monitor parameters such as timing margins, voltage behavior, and thermal conditions. During testing, ATE systems analyze this information and adjust test conditions dynamically rather than relying on static limits.

This approach improves classification accuracy by identifying genuinely weak devices while reducing the rejection of functional parts. It also accelerates yield-learning cycles during process development [17].

6. Yield-Enabling Analytics and AI-Enhanced Process Control

At the sub-2nm node, the challenge of manufacturing extends beyond individual process steps. Lithography, etch, deposition, planarization, inspection, and testing are no longer independent optimization problems. Variations introduced at one stage can propagate through the entire process flow and interact with the complex geometries of 3D integrated circuits. As Edge Placement Error (EPE) budgets approach 0.5 nm, maintaining yield requires continuous coordination between design, manufacturing, metrology, and test systems.

For this reason, modern fabs increasingly rely on AI-enhanced analytics as a unifying layer that links physical process tools with design and yield-management workflows.

Fig. 4. An infographic of a closed-loop semiconductor process control flow. The diagram forms a rectangular loop connected by thin gray and blue arrows indicating clockwise flow. Top-Left: a chip layout (Design Data); Top-Middle: an abstract AI brain node (ML-DFM); Top-Right: a simplified lithography chamber silhouette (Manufacturing); Bottom-Right: a testing bench with wavy signal lines (Test & Telemetry); Bottom-Middle: a stylized microscope lens probing a grid (Metrology); Bottom-Left: a 3D bar graph with an upward trend (Yield Analytics).

6.1 AI-Driven Design-for-Manufacturability as the Yield Backbone

Traditional rule-based Design-for-Manufacturability (DFM) methodologies were developed for process nodes where manufacturing variability could be managed through deterministic design rules [18]. At sub-2nm dimensions, however, the interaction between lithography, CMP, thermal effects, and 3D stacking complexity makes purely rule-based approaches increasingly insufficient.

Modern DFM platforms therefore incorporate machine-learning models directly into the designclosure process [19].

ML-Based Hotspot Prediction

One of the most important applications is hotspot prediction. Machine-learning models are trained using historical manufacturing data, process simulations, and layout information to identify regions that are likely to experience:

• Lithographic patterning failures

• CMP non-uniformity

• Thermal hotspots

• Overlay-sensitive geometries

Because these risks can be identified before tape-out, engineers can modify layouts proactively rather than discovering manufacturability issues during process qualification.

Generative AI for Pattern Repair

Advanced DFM systems increasingly combine hotspot prediction with generative AI. Instead of merely flagging problematic patterns, the software can propose or automatically generate alternative geometries that maintain electrical functionality while improving manufacturability.

This capability is particularly valuable in dense standard-cell libraries and complex routing regions where manual optimization becomes impractical.

Reinforcement Learning for Design-Space Exploration

Three-dimensional ICs introduce competing requirements involving power consumption, performance, thermal behavior, routing density, and manufacturing robustness. Reinforcement-learning techniques allow design teams to explore these tradeoffs more efficiently by evaluating large numbers of possible configurations and converging toward solutions that satisfy both design objectives and manufacturing constraints [20].

Edge Placement Error as a Unifying Metric

A notable trend is the emergence of Edge Placement Error (EPE) as a common metric linking design and manufacturing [21].

At sub-2nm dimensions, EPE budgets may fall below 0.5 nm. Achieving such accuracy requires lithography, etch, and CMP variation to be treated as a single optimization problem rather than separate process modules. Modern AI-driven DFM frameworks increasingly use EPE as a holistic yield metric that bridges design intent and process execution.

Physics-ML Hybrid Verification

Although machine-learning models can identify complex patterns that are difficult to capture using conventional rules, they do not inherently provide formal guarantees. Consequently, fabs are increasingly adopting hybrid methodologies that combine machine learning with physics-based process simulation [22].

These hybrid approaches provide greater confidence during design sign-off and are expected to become increasingly important for high-reliability applications where yield predictions must be both accurate and explainable [22].

6.2 ML-Enhanced Metrology and Root-Cause Analytics

Machine learning is also transforming metrology and inspection. Rather than functioning solely as post-processing software, AI is increasingly embedded within measurement systems themselves, improving both throughput and defect sensitivity [23].

ML-Powered Instrumentation

Several metrology techniques now rely on machine-learning-assisted data interpretation.

Examples include:

• Optical scatterometry systems that use ML to improve correlation between optical signatures and physical dimensions [24].

• CD-SEM platforms that employ AI to improve signal-to-noise performance [25].

• Optical inspection systems that use ML to determine which regions require higher-resolution SEM review.

These approaches increase measurement accuracy while reducing the inspection burden on slower high-resolution tools.

Design-Aware Defect Classification

The effectiveness of defect inspection depends not only on identifying anomalies but also on determining whether those anomalies are likely to affect circuit functionality.

By correlating inspection signatures with layout information, modern analytics systems perform design-aware defect classification [26], distinguishing between nuisance defects and yield-limiting defects. This substantially improves engineering efficiency by directing review resources toward the most critical failure mechanisms.

Root-Cause Deconvolution (RCD)

A major challenge in advanced manufacturing is distinguishing random defectivity from systematic process drift.

To address this problem, fabs increasingly employ Root-Cause Deconvolution (RCD) techniques [27]. These methods correlate:

• In-line metrology measurements

• Inspection data

• Process histories

• Layout characteristics

to isolate the underlying source of variation.

For example, copper dishing measurements obtained through White-Light Interferometry may be linked to specific pattern-density conditions or CMP process parameters. Once the source is identified, process recipes can be adjusted before yield degradation becomes significant.

Unit-Level Traceability

Another emerging capability is unit-level traceability.

Individual dies are increasingly tracked throughout manufacturing and testing, allowing engineers to correlate electrical performance with spatial location on the wafer. Such analysis often reveals characteristic yield patterns — such as radial or “donut-shaped” performance distributions — that indicate process-tool drift or material non-uniformity.

This traceability transforms yield management from a wafer-level activity into a die-level optimization problem, providing earlier visibility into process excursions and accelerating yield ramp for new technologies.

Analytics for the Test Data Explosion

The amount of manufacturing and test data generated at sub-2nm nodes continues to increase rapidly. Test programs may involve hundreds of thousands or even one million measurements per chip, making manual analysis impractical.

AI-driven analytics platforms are therefore becoming essential for:

• Failure clustering

• Yield prediction

• Adaptive test optimization

• Correlation of telemetry data with manufacturing history

These capabilities allow manufacturers to extract actionable information from increasingly complex datasets while maintaining production throughput.

7. Future Directions

The evolution of sub-2nm manufacturing is likely to be driven as much by analytics and automation as by advances in process hardware. Several technological trajectories appear particularly significant.

7.1 Certifiable Physics-ML Hybrid Models

Current AI-driven DFM systems can identify manufacturing risks but cannot formally guarantee correctness. Future frameworks are likely to combine machine-learning prediction with physics-based constraints and uncertainty quantification [22].

Such systems could provide certifiable verification of hotspot fixes, EPE compliance, and reliability margins, making them suitable for automotive, aerospace, and other high-reliability applications.

7.2 AI-Driven Design-Technology Co-Optimization (DTCO)

Future DTCO workflows are expected to integrate design environments, process simulation, inspection systems, and production equipment into a unified feedback loop.

Inspection results generated in the fab — such as voltage-contrast defects or CMP topography measurements — may eventually be fed back into EDA environments in near real time, enabling adaptive design-rule tuning and faster process development cycles.

7.3 Real-Time Adaptive Manufacturing

Inspection and process equipment are becoming increasingly interconnected.

Future fabs are expected to move toward closed-loop operation in which:

• E-beam inspection identifies emerging defect signatures.

• AI systems classify root causes.

• Process tools automatically modify lithography, etch, deposition, or CMP parameters.

Such adaptive manufacturing could significantly reduce yield loss by responding to process excursions within production timescales rather than after offline analysis.

7.4 Predictive Yield Management

Yield engineering is gradually shifting from reactive troubleshooting to predictive control [28].

By correlating:

• Inspection results

• Equipment-maintenance records

• Material-batch histories

• Environmental conditions

• Test telemetry

Machine-learning models may be able to forecast yield excursions before measurable losses occur. This would enable proactive intervention and improve overall fab stability [28].

7.5 Standardization of Trustworthy AI

As AI assumes a greater role in manufacturing decisions, industry-wide standards will become increasingly important.

Organizations such as IRDS and SEMI are expected to develop benchmark datasets, validation methodologies, explainability requirements, and acceptance criteria for AI-based yield-management systems. These standards will play a role analogous to existing reliability standards used throughout semiconductor manufacturing.

7.6 Cross-Physics Analytics for 3D ICs

Three-dimensional integration introduces tightly coupled electrical, thermal, and mechanical effects that are difficult to evaluate independently.

Future analytics platforms are expected to combine:

• Process-control measurements

• Thermal simulation

• Mechanical-stress modeling

• Electrical characterization

into unified predictive frameworks [28]. Such cross-physics analysis may enable early detection of complex failure mechanisms, including thermal-stress-induced delamination, vertical alignment errors, and degradation of high-density interconnect structures.

Conclusion: The Automated Future of 3D IC Fabrication

As the semiconductor industry transitions to the sub-2nm node, the traditional boundaries between design, manufacturing, metrology, inspection, and testing continue to erode. The industry’s ability to sustain scaling increasingly depends on its capacity to coordinate these domains through a tightly integrated ecosystem of advanced process equipment and data-driven analytics.

The critical tools discussed throughout this review report — High-NA EUV lithography, cryogenic highaspect-ratio etch systems, seam-suppressed deposition platforms, hybrid inspection architectures, adaptive test systems, and AI-enhanced process-control frameworks — each address a distinct physical bottleneck. Their true value, however, emerges when they operate as components of a unified manufacturing system capable of continuous feedback and optimization.

Looking forward, the most significant advances may not come from any single fabrication tool. Instead, they are likely to arise from increasingly sophisticated interactions among AI-driven DFM, ML-enhanced metrology, adaptive process control, and predictive yield analytics. The integration of physics-based models with machine learning, coupled with closed-loop design-to-fab feedback, points toward a future in which manufacturing decisions are continuously informed by real-time operational data.

In this environment, the fab evolves from a collection of isolated process modules into a self-correcting cyber-physical system. Such an approach is likely to be essential for maintaining yield, reliability, and economic viability as semiconductor manufacturing advances deeper into the sub-2nm era and increasingly complex 3D integrated architectures.

References

[1] Imec. “Imec’s chip scaling roadmap: smaller, better, faster.” https://www.imec-int.com/en/articles/smaller-better-faster-imec-presents-chip-scaling-roadmap

[2] Semiconductor Engineering. “Metrology Advances Step Up To Sub-2nm Device Node Needs.” https://semiengineering.com/metrology-advances-step-up-to-sub-2nm-device-node-needs/

[3] IEEE IRDS. “International Roadmap for Devices and Systems (IRDS), 2022 Edition — More Moore.” https://irds.ieee.org/images/files/pdf/2022/2022IRDS_MM.pdf

[4] ASML. “TWINSCAN EXE:5000 — EUV lithography systems.” https://www.asml.com/en/products/euv-lithography-systems/twinscan-exe-5000

[5] ASML. “5 things you should know about High NA EUV lithography.” https://www.asml.com/en/news/stories/2024/5-things-high-na-euv

[6] Semiconductor Engineering. “Cryogenic Etch: A Key Enabler Of 3D NAND.” https://semiengineering.com/cryogenic-etch-a-key-enabler-of-3d-nand/

[7] Shen, M. et al. “Progress report on high aspect ratio patterning for memory devices.” Japanese Journal of Applied Physics, 62, SI0801 (2023). https://iopscience.iop.org/article/10.35848/1347-4065/accbc7

[8] Lam Research. “Lam Research Introduces Lam Cryo™ 3.0 Cryogenic Etch Technology to Accelerate Scaling of 3D NAND for the AI Era.” https://investor.lamresearch.com/2024-07-31-Lam-Research-Introduces-Lam-Cryo-TM-3-0-Cryogenic-Etch-Technology-to-Accelerate-Scaling-of-3D-NAND-for-the-AI-Era

[9] “Atomic Layer Etching: The Key to High Aspect Ratio Semiconductor Manufacturing.” https://drmanager.github.io/2026/02/06/ale-high-aspect-ratio-etching-semiconductor/

[10] U.S. Patent 9,082,826. “Methods and apparatuses for void-free tungsten fill in three-dimensional semiconductor features.” https://image-ppubs.uspto.gov/dirsearch-public/print/downloadPdf/9082826

[11] Applied Materials. “Centura iSprint SSW ALD/CVD.” https://www.appliedmaterials.com/us/en/product-library/centura-isprint-ssw-ald-cvd.html

[12] Semiconductor Engineering. “Ruthenium Liners Give Way To Ruthenium Lines.” https://semiengineering.com/ruthenium-liners-give-way-to-ruthenium-lines/

[13] Semiconductor Engineering. “E-Beam Inspection Proves Essential For Advanced Nodes.” https://semiengineering.com/e-beam-inspection-proves-essential-for-advanced-nodes/

[14] Semiconductor Engineering. “Finding Defects With E-Beam Inspection.” https://semiengineering.com/finding-defects-with-e-beam-inspection/

[15] “Voltage contrast inspection of an electrical defect on a semiconductor wafer with X-ray photoelectron spectroscopy and an electron beam.” Applied Surface Science (2025). https://www.sciencedirect.com/science/article/abs/pii/S0169433225000492

[16] Tessolve. “Automated Test Equipment (ATE) Modernization: Enhancing Efficiency in Electronics Testing.” https://www.tessolve.com/blogs/automated-test-equipment-ate-modernization-enhancing-efficiency-in-electronics-testing/

[17] Advantest. “What Is ATE (Automated Test Equipment)? Roles, Systems & Feature Outlook.” https://www.advantest.com/en/semiconductor-basics/automated-test-equipment/

[18] U.S. Patent Application. “Machine learning-based hotspot prediction in electronic design automation (EDA) applications.” https://image-ppubs.uspto.gov/dirsearch-public/print/downloadPdf/12561507

[19] “Computational Lithography Using Machine Learning Models.” Journal of Micro/Nanolithography, MEMS, and MOEMS. https://www.researchgate.net/publication/349386359_Computational_Lithography_Using_Machine_Learning_Models

[20] “AI for Computational Lithography: ILT/OPC Acceleration via Reinforcement Learning.” International Journal of Research Publication and Reviews, 6(9), 490–504 (2025). https://ijrpr.com/uploads/V6ISSUE9/IJRPR52523.pdf

[21] Science.gov. “Edge placement error: Topics.” https://www.science.gov/topicpages/e/edge+placement+error

[22] “A Review of Multiscale Thermal Modeling in Heterogeneous 3D ICs.” arXiv:2604.03290. https://arxiv.org/pdf/2604.03290

[23] SPIE. “Metrology, Inspection, and Process Control for Semiconductor Manufacturing XXXV” (2021). https://spie.org/Publications/Proceedings/Volume/11611

[24] “Exploring Machine Learning for Semiconductor Process Optimization: A Systematic Review.” TechRxiv. https://www.techrxiv.org/users/689925/articles/1184255/master/file/data/Exploring_Machine_Learning_for_Semiconductor_Process_OptimizationA_Systematic_Review1_/Exploring_Machine_Learning_for_Semiconductor_Process_OptimizationA_Systematic_Review1_.pdf

[25] Dey, B. et al. “Deep Learning-Based Defect Classification and Detection in SEM Images.” arXiv:2206.13505. https://arxiv.org/pdf/2206.13505

[26] U.S. Patent 10,679,333. “Defect detection, classification, and process window control using scanning electron microscope metrology.” https://image-ppubs.uspto.gov/dirsearch-public/print/downloadPdf/10679333

[27] Semiconductor Engineering. “Root Cause Deconvolution.” https://semiengineering.com/root-cause-deconvolution/

[28] “Review of Applications of Regression and Predictive Modeling in Wafer Manufacturing.” Electronics, 14(20), 4083 (2025). https://www.mdpi.com/2079-9292/14/20/4083