Lithography Overlay Control: How AI Enhances Alignment Accuracy
Anyone who has worked in lithography knows that overlay — the alignment between successive patterning layers — may seem like a small number (just a few nanometers of deviation), but the consequences can be severe. I have personally witnessed, more than once, entire lots of wafers scrapped because an overlay drift was not caught in time. This article explores the fundamentals of overlay control and what becomes possible when AI enters the picture.
What Overlay Really Is, and Why It Matters So Much
Semiconductor manufacturing is a layer-by-layer stacking process. Each lithography layer must be precisely aligned to the previous one, and the misalignment between layers is the overlay error. To use a rough analogy: imagine drawing a circuit on a sheet of paper, then placing a transparent sheet on top to draw the second layer. If the two sheets are not aligned, the circuits will not connect.
At mature process nodes, such as 28nm, the overlay tolerance is roughly 8-10nm — still providing some margin. But at advanced nodes the situation is completely different. The overlay budget at the 5nm node is only about 2-3nm, and leading customers at the 3nm node are running on-product overlay of 2nm to 2.5nm. To put this in perspective, a single atom is approximately 0.1-0.3nm in diameter — we are talking about precision on the order of just a dozen atoms.
The consequences of overlay exceedance are immediate: vias fail to land on metal lines, transistor gates misalign with source/drain regions, and contacts land where they should not. In milder cases this leads to degraded device performance and yield loss; in severe cases it results in functional failure and complete wafer scrap. On a production line with 50,000 wafers per month capacity, even a 1% increase in scrap rate translates to tens of millions in losses each month.
How Overlay Is Measured and Corrected
Overlay measurement relies primarily on two methods: IBO (Image-Based Overlay) and DBO (Diffraction-Based Overlay). IBO is relatively straightforward — an optical microscope examines alignment marks on both the exposed layer and the reference layer, and image analysis computes the offset. DBO is based on diffraction principles, deducing overlay values from changes in diffraction intensity of grating targets. Each method has its strengths and weaknesses: IBO is more intuitive but limited in resolution, while DBO offers higher precision but imposes stricter requirements on target design and is susceptible to interference from underlying film stacks in complex 3D NAND structures.
Once overlay values are measured, the APC (Advanced Process Control) correction flow takes over. The most basic approach is lot-wise correction — after each lot is processed, metrology data is used to calculate correction values that are fed back to the scanner for the next lot. A more refined approach is wafer-wise correction, where each wafer receives its own correction. Going further, site-by-site correction is possible, though it places very high demands on metrology throughput.
The key physical parameters being corrected include translation (global shift), rotation, magnification, and various higher-order terms. During scanning exposure, the scanner compensates for these errors in real time by adjusting the relative positions of the reticle stage and wafer stage. ASML scanners have an internal sensor system that collects extensive alignment data during exposure, combining it with metrology feedback to apply corrections.
Why Linear Models Are No Longer Sufficient
Traditional overlay correction models are linear, using a set of polynomials to fit the overlay distribution across measurement points on the wafer. The industry started with 6-parameter models (translation x/y, rotation, magnification, plus some low-order terms), then expanded to 10-parameter, 20-parameter, and even higher-order corrections.
At mature process nodes, this approach worked reasonably well. But at advanced nodes, the limitations have become apparent.
Let me give a concrete example. On a 14nm production line, we observed that overlay distributions in the wafer edge region exhibited clearly nonlinear patterns, and these patterns were related to the thermal history the wafer experienced in prior process steps. The film stress distribution after CMP, localized deformation from ion implantation, even the wafer’s dwell time in the FOUP — all of these factors combined to form a highly nonlinear overlay fingerprint. Fitting such distributions with traditional polynomial models, even at very high orders, could not fully capture the pattern.
Higher-order corrections also present an awkward dilemma: the more parameters you use, the more metrology data you need. You require a sufficient density of measurement points to reliably fit a high-order model; otherwise you end up overfitting, mistaking noise for genuine overlay signal. But increasing the number of measurement points means longer metrology time, reduced throughput, and increased cycle time — all of which represent real costs to the production line.
Another often-overlooked issue is that traditional feedback correction inherently lags behind. You use this lot’s metrology data to correct the next lot, but process conditions may have already drifted between the two lots. Chuck thermal deformation is changing, photoresist coating thickness is fluctuating, and upstream CMP uniformity is not constant. This paradigm of “using yesterday’s data to correct today’s problems” naturally becomes less effective at advanced nodes where process sensitivity is ever increasing.
AI Methods: From Reactive Compensation to Proactive Prediction
This is the core value proposition of AI/ML in overlay control — using machine learning models to capture nonlinear relationships that traditional linear models cannot handle, and shifting from “post-hoc compensation” to “preemptive prediction.”
How does this work in practice? A typical approach looks like this:
All the metrology data accumulated across upstream process steps — film thickness, CD, prior-layer overlay values, alignment sensor signals, even equipment sensor readings — are fed into an ML model (typically gradient boosting or a neural network), which learns the mapping between these input variables and the final overlay outcome. Once the model is trained, it can predict the overlay distribution for each wafer before it enters the lithography step, based on the metrology data from its preceding layers, and proactively adjust exposure parameters to compensate.
This is what is known as feedforward compensation. Its advantage is that corrections are already tailored before exposure, rather than waiting until after exposure to measure and feed back. ASML refers to this concept as “computational overlay” — not measuring overlay, but computing and predicting it. Their results show that combining scanner internal sensor data with ML models can identify both systematic and random overlay errors that conventional metrology methods miss.
From my own project experience, the two areas where ML models provide the most value in overlay prediction are: first, their ability to capture wafer-to-wafer variation — since each wafer has a different upstream process history, the ML model translates these differences into individualized overlay predictions; and second, their ability to fit nonlinear patterns, particularly the complex distributions in the wafer edge region that traditional polynomials struggle with.
Of course, ML models are not a silver bullet. Data quality is critical — if the metrology data itself contains noise or bias, the model will learn incorrect relationships. The model’s generalization capability also requires validation; when process conditions change or new products are introduced, the model may need retraining or fine-tuning. These are all practical hurdles that must be addressed in engineering implementation.
R2R Closed-Loop Control: Keeping Corrections in Step with Process Changes
Feedforward compensation addresses the “prediction” challenge, but feedforward alone is not enough. The process itself is continuously drifting, equipment states are changing, and these are things the feedforward model cannot fully anticipate. A feedback loop is also needed — this is R2R (Run-to-Run) closed-loop control.
The basic concept of R2R is straightforward: after each batch of wafers, ADI (After Develop Inspection) or AEI (After Etch Inspection) overlay metrology data is collected, compared against model predictions, residuals are calculated, and then model parameters or correction recipes are updated. This way, the model continuously adapts to process changes rather than remaining frozen at its initial training state.
Traditional R2R uses statistical methods such as EWMA (Exponentially Weighted Moving Average) — simple and effective, but slow to respond to complex drift patterns. With ML, R2R can become more intelligent — for example, using time-series models to capture equipment drift trends, anomaly detection algorithms to identify sudden overlay excursions, or even adaptively adjusting model weights based on equipment PM (Preventive Maintenance) cycles.
On actual production lines, a mature R2R system can improve Cpk (process capability index) by 40% to 70%. What does that mean? It means the overlay distribution becomes more concentrated and stable, with dramatically reduced probability of out-of-spec events. The contribution to yield is very direct.
However, the engineering challenges of deploying R2R in production are substantial. Data pipelines must be established — metrology data, equipment data, MES data, all scattered across different systems, must be aggregated into the R2R system in real time. Computation must be fast enough — the wafer is waiting on the scanner for the correction value to be calculated, and excessive latency directly impacts throughput. Model robustness must be sufficient — a single anomalous data point should not send the correction value wildly off course. These are all system-level engineering challenges.
Engineering Practices for Online Control
This brings us to a practical question: how are all these AI models and R2R systems actually deployed on a production line?
This is a real dilemma that many fabs face. Algorithm teams can produce impressive results in offline environments using Python, but deploying models into a production line’s real-time control loop involves engineering challenges far more complex than the algorithms themselves — real-time data ingestion, model inference latency, exception handling, integration with MES/EAP systems, model lifecycle management, and more.
NeuroBox E3200 for Overlay Control Applications
This is precisely the problem that MST Semiconductor’s NeuroBox E3200 product line is designed to solve. E3200 is an online AI platform for production lines, with core capabilities including VM (Virtual Metrology) and R2R real-time control. In the overlay control use case, it delivers several key functions:
First, it unifies data pipelines. E3200 connects directly to the scanner’s SECS/GEM interface and metrology tool data streams, aggregating previously siloed alignment data, overlay metrology data, and equipment sensor data onto a unified data platform for real-time collection and cleansing.
Second, it provides online VM prediction. Based on upstream metrology data, trained ML models estimate each wafer’s overlay distribution, providing the basis for feedforward compensation. This eliminates the need for full overlay metrology on every wafer, enabling reduced metrology sampling rates while maintaining control precision and freeing up metrology capacity.
Third, it enables R2R closed-loop correction. By combining metrology feedback with model predictions, correction recipes are continuously updated to keep overlay control aligned with dynamic process changes. E3200 supports multiple R2R strategies and can be flexibly configured to meet the needs of different products and processes.
In essence, E3200 bridges the engineering gap between “algorithm validated in the lab” and “stable operation on the production line.” Those who have been through this journey know just how wide that gap can be.
Closing Thoughts
Overlay control is a topic where the technical details run deep, but the overarching trend is clear: as process nodes advance, traditional linear models and simple lot-to-lot feedback are no longer adequate. AI/ML brings stronger nonlinear modeling capabilities and a paradigm shift from reactive compensation to proactive prediction, while R2R closed-loop control ensures these capabilities can deliver sustained, stable results on the production line.
Of course, putting technology into practice is always far more complex than theoretical analysis. Data quality, system integration, model operations, personnel training — any of these can become a bottleneck. But the direction is certain, and an early start yields early benefits.
Learn How NeuroBox E3200 Supports Overlay Control
NeuroBox E3200 provides online VM/R2R solutions for semiconductor production lines, covering key applications including lithography overlay control, etch CD control, and thin film process optimization.
