Using Test And Metrology Data For Dynamic Process Control

Advanced packaging is transforming semiconductor manufacturing into a multi-dimensional challenge, blending 2D front-end wafer fabrication with 2.5D/3D assemblies, high-frequency device characterization, and complex yield optimization strategies.

These combinations are essential to improving performance and functionality, but they create some thorny issues for which there are no easy fixes. For example, variability — whether in material properties, bonding precision, or even test parameters — can show up anywhere in the manufacturing process, reducing yield and impacting long-term reliability.

Managing all of this requires a data-centric approach that allows real-time adjustments whenever problems show up in test and metrology. This is complicated by the fact that modern multi-die systems rely on an array of different materials, fine-pitch interconnects, and high-frequency components, which introduce entirely new sources of variability. These include edge defects from lithographic pattern breaks, subtle parametric shifts in analog and mixed-signal devices, and challenges in parallel testing due to electrical crosstalk or probe-card design.

The solution lies in integrating test and metrology data into dynamic process control systems, which provide a closed-loop framework that links early-stage measurements, in-line inspections, and final test results. The resulting data can be used to pinpoint and rectify process deviations before they propagate downstream.

“The semiconductor industry can no longer rely on static process windows to ensure yield and reliability,” says Monita Pau, strategic marketing director at Onto Innovation. “Dynamic process control, powered by metrology and defect inspection, provides the actionable insights needed to make real-time adjustments and manage the complexity of advanced packaging.”

Improving process control
Advanced packaging has elevated the role of inspection and test tools from passive defect monitors to active components in dynamic process control systems. These tools now provide real-time data that feeds back into the manufacturing process, enabling precise and timely adjustments that preserve yield and reduce waste. By integrating inspection and test results into a closed-loop system, manufacturers can address excursions before they propagate downstream, safeguarding both product reliability and profitability.

“Test data serves as the dashboard for your process, showing key performance indicators (KPIs) and statistical distributions,” says David Vondran, wireless product strategist at Teradyne. “When you see a spread in the data that doesn’t match expectations, it could point to calibration issues, or it could be environmental or thermal issues.”

The push toward mmWave devices adds other challenges involving signal integrity, calibration, and cost-effective throughput. The solution involves adaptive test strategies, such as continuous-wave or loopback methods, which help reduce costs while enabling real-time process adjustments. For example, by continuously monitoring signal reflections during testing, manufacturers can identify variations in interconnect impedance and adjust process parameters to maintain optimal signal integrity.

“By detecting and addressing these anomalies — such as variations in substrate warpage or interconnect impedance — in real time, you protect not only the wafer but also the high-value die that get attached later,” says Pau. “This ability to catch and correct errors early is essential to maintaining yield as advanced packaging tolerances grow tighter.”

Finely-tuned process control hinges on the precise measurement of critical parameters, such as line/space uniformity in redistribution layers (RDL), film thickness for plating and photoresist steps, and bump co-planarity for bonding success.

“Measuring thin films is extremely important to our customers,” says John Hoffman, computer vision engineering manager at Nordson Test & Inspection. “They really want to measure those films in situ, in the middle of the process, to catch failures early before continuing the manufacturing process.”

Deviations, such as uneven bump height or irregular film thickness, are immediately fed back in order to initiate corrective actions. That high precision enables manufacturers to address subtle issues that might otherwise escape traditional inspection methods, preventing defects from compromising subsequent stages and safeguarding yield.

To further enhance real-time control, advanced process flows increasingly are incorporating in-situ metrology into process equipment. For instance, CMP (chemical mechanical planarization) systems can measure polishing depth and uniformity in real-time. If wafers are not sufficiently planarized, they can be reworked immediately rather than finding a problem later in the flow.

This in-line metrology closes the gap between inspection and correction, allowing manufacturers to intervene at the earliest possible stage and to avoid propagating defects through later steps. Similar in-situ techniques are being applied in other critical advanced packaging processes, such as die placement, bonding, and underfill encapsulation to monitor parameters like alignment accuracy, bond strength, and voiding.

“Metal oxide thickness correlates directly with non-wet open failures,” explains Christopher Claypool, senior R&D director at Bruker. “To get an accurate thickness measurement, you must characterize the optical constants simultaneously. Multi-angle, multi-modal technology, which combines ellipsometry and reflectometry, achieves better-than-angstrom dynamic repeatability. This level of precision is critical for controlling the thickness of metal oxide films, ensuring strong process control and minimizing the risk of non-wet open failures.”

While not every back-end step can integrate metrology in situ, more packaging vendors see the benefit of real-time control loops, especially as advanced packaging margins narrow. By embedding metrology into critical steps and leveraging the data for immediate adjustments, manufacturers can maintain tighter tolerances and adapt dynamically to process variations.

Limitations of dynamic process control
Dynamic process control, while transformative for advanced packaging, faces significant challenges that can complicate efforts to maintain high yields and profitability. These challenges often stem from the inherent complexity of managing data across multiple test steps, the economic and technical hurdles of parallel testing, and the unique difficulties of identifying defects in analog and mixed-signal devices.

“Maintaining tight control over critical parameters like temperature, pressure, flow rates, and material composition is essential for achieving high yield and product quality,” explains Andras Vass-Varnai, 3D-IC solutions engineer at Siemens Digital Industries Software. “However, customers are facing increasingly complex challenges with precise process control, particularly as device geometries continue to shrink and manufacturing processes become more complex with heterogeneous integration.”

One of the primary limitations in dynamic process control lies in the integration and latency of data from different test stages, such as wafer sort, parametric test, final test, and optical inspection. The timeline for collecting and correlating this data can stretch over days or weeks, making it difficult to implement real-time corrective actions.

“An event might occur at litho, but the relevant test data only comes a week later,” says Dieter Rathei, CEO of DR Yield. “In analog or mixed signal, you can’t always see a neat pattern. It’s often a parametric drift that’s partially out of spec. It demands thorough correlation.”

Parallel testing adds other hurdles. While testing multiple devices simultaneously can significantly reduce test time, it also increases the complexity of probe-card design and data analysis. Electrical crosstalk, alignment constraints, and the inherent mismatch between round wafers and rectangular probe arrays all contribute to potential inaccuracies.

“We often see stripy patterns and ask, ‘Is it a process flaw or just a site in the probe card?’” says Rathei. “Parallel test means you might have 16 or 32 devices under contact at once. You have to watch parametric slopes or site-to-site trends to catch incipient probe issues.”

Another challenge is the variability in test data itself, which can obscure actionable insights. Even with automation, real-time decision-making in dynamic manufacturing processes can be difficult due to the sheer volume of data and the complexity of analysis required.

“When the statistical distribution spreads out, you’re likely failing more units—or over-testing and filtering out good ones,” explains Vondran. “The challenge is identifying outliers in a way that’s mutually beneficial to the engineers trying to adjust the process and to the supplier trying to increase yield and reduce costs.”

The unique characteristics of analog and mixed-signal devices further complicate things. Unlike digital devices, which often fail definitively and are easy to classify, analog devices can exhibit gradual degradation or parametric drift while still passing specifications. This makes identifying outliers more difficult and introduces the risk of marginal parts escaping detection.

“These ‘defect-induced parametric losses’ can slip by if specs are wide,” Rathei explains. “In analog, it’s often not a total device fail, but a transistor mismatch that shifts an output stage near the spec limit.”

Dynamic process control must account for these nuanced failure modes, balancing broad test coverage with overkill. Yield-management algorithms are critical in this effort, as they track near-spec anomalies and correlate them with process data to identify risks early. Insufficient test coverage increases the likelihood of field failures and costly RMAs (return material authorizations), while overly broad testing inflates costs and slows production.

“A lot of fabs deal with false alerts from SPC (Statistical Process Control) and FDC (Fault Detection and Classification) systems—sometimes as high as 40%,” says David Park, vice president of marketing at Tignis. “AI can cut down on those false positives while also shrinking the size of the haystack when diagnosing root causes, helping engineers identify real issues faster and more accurately.”

These limitations underscore the need for smarter, more adaptive solutions in dynamic process control. In response, the integration of AI and machine learning is transforming how manufacturers leverage test and metrology data to address these challenges. By enabling real-time analysis, predictive insights, and automated decision-making, AI-driven tools offer a path forward—tackling issues like latency, variability, and defect detection with unprecedented precision.

AI-driven process control
The sheer number of variables, from material inconsistencies to parametric shifts in analog devices, requires an approach that can adapt to continually changing conditions. This is where AI-driven process control fits in, leveraging advanced algorithms and real-time data analysis to refine workflows, predict defects, and improve yield in ways that static process windows and manual interventions cannot.

Fig. 1: By monitoring film thickness in real time, fabs can optimize their processes, minimize variability, and improve yield. Source: Tignis

Unlike traditional manufacturing, where process deviations often manifest as straightforward dimensional or electrical issues, advanced packaging introduces a broader spectrum of variabilities. These include temperature gradients, heterogeneous material responses, and subtle defects that can escape detection until final test stages, or worse, show up in the field.

“There are defects in the process — not many — but you need to find them. They will affect you,” says Evelyn Landman, CTO of proteanTecs. “Sometimes the problems aren’t even inside the chip. They come from the system feeding the chip, like a DC-to-DC converter issue or poor assembly quality, where you lose connections or see power issues that make everything worse.”

AI and machine learning are uniquely positioned to address this complexity because they excel at identifying patterns and correlations within vast datasets. These tools can detect subtle anomalies that human operators or traditional statistical process control (SPC) might overlook, providing insights that enable faster and more targeted interventions.

“Drift is inevitable in every process step, whether it’s lithography, CMP (Chemical Mechanical Planarization), or PVD (Physical Vapor Deposition),” explains Tignis’ Park. “AI can predict when drift is going to take you out of spec by analyzing multivariate correlations that humans often miss. This lets fabs manage drift more effectively, keeping critical dimensions within spec longer and reducing unnecessary cleanings or maintenance.”

AI-driven process control operates by creating a continuous feedback loop between metrology, test, and process adjustments. By analyzing data from in-line measurements, wafer sort, and final test, AI algorithms can identify trends, classify defects, and correlate process deviations with yield-impacting outcomes.

“The faster you can identify the root cause of an issue, the sooner you can adjust the process,” says Pau. “Analytics can turn raw data into actionable insights, making process control more dynamic and effective.”

For example, if metrology data reveals a slight deviation in copper pad flatness during hybrid bonding, AI algorithms can flag the issue and recommend adjustments to bonding pressure or temperature to prevent misalignment or void formation downstream.

In advanced packaging, meanwhile, high-frequency devices generate enormous data streams during testing, making manual analysis impractical. AI tools are designed to sift through this data, filtering noise and identifying meaningful trends that inform process improvements.

“AI can recognize patterns, classify defects, and provide real-time feedback during testing,” explains Adrian Kwan, business development manager at Advantest. “This helps manage data processing and ensures we quickly identify defects before they propagate through the process.”

A significant advantage of AI-driven process control is its ability to predict defects before they occur, allowing engineers to implement corrective actions early. By analyzing historical and real-time data, machine learning models can identify process conditions that correlate with future failures, enabling proactive interventions.

“AI-driven tools help track marginal defects that may cause parametric shifts, allowing fabs to address quality risks early,” adds Rathei. “These systems can correlate subtle process variations with eventual test outcomes, improving yield predictability.”

This is particularly important for analog and mixed-signal devices, where defects often manifest as gradual parametric drifts rather than outright failures. For example, an AI system might detect that a slight change in a deposition process leads to increased resistance in a critical interconnect, allowing engineers to adjust the process before the issue impacts yield.

“Dynamic parametric test (DPT) automates manual debugging tasks that previously required significant human involvement,” explains Preet Paul Singh, director of electronic R&D at Advantest. “Using programmed automation, DPT acts as a virtual engineering staff, making decisions on parametric test measurements in real-time and executing tasks like retesting, cleaning probe cards, or even initiating additional data collection.”

AI identifies defects and classifies them more accurately than traditional methods. By combining data from metrology tools, such as X-ray fluorescence or atomic force microscopy, with test results, AI can differentiate between benign variations and yield-critical defects. This enhanced classification capability reduces false positives, ensuring that only genuinely defective parts are flagged for rework or scrap.

“We generate massive data at these frequencies, so AI or machine learning can really help,” says Kwan. “It can spot defect signatures quickly, track statistical trends, and push real-time dashboards to the floor.”

AI-driven process control also accelerates yield ramp in new product introductions. By analyzing early prototype data and correlating it with production outcomes, AI systems can identify the root causes of variability and refine process recipes more quickly. This capability improves time-to-market and reduces costly rework or scrap during the early stages of production.

The future of dynamic process control
Dynamic process control’s greatest strength lies in its ability to bridge the gap between test, inspection, and actionable process adjustments. By creating a continuous feedback loop, manufacturers can detect issues early and correct them before they propagate through production.

The importance of AI and machine learning in this evolution cannot be overstated. These tools bring new levels of precision and efficiency to dynamic process control.

“There are many small defects in mission-critical applications that are difficult to find with traditional pass/fail tests, like ATPG or memory BIST,” says proteanTecs’ Landman. “Our approach uses parametric measurements and machine learning to go beyond pass/fail testing by providing a model of expected results for different parameters. If a test result deviates from the model, we can detect outliers.”

As test and metrology tools become more sophisticated, they also are driving advancements in areas such as in-situ metrology and adaptive test methods. In-line tools are now capable of providing real-time measurements of critical parameters such as film thickness, bump co-planarity, and redistribution layer (RDL) uniformity, feeding this data into automated systems that implement dynamic corrections.

“AI becomes one of the few appreciating assets in a fab,” adds Park. “With AI tools you can document the rules and insights from senior engineers, so even when they retire or move on, their knowledge is retained and accessible to train the next generation.”

The value of dynamic process control also extends beyond yield protection. It drives innovation by enabling faster yield ramps in new product introductions and supports the scalability of advanced packaging technologies.

“DPT builds a foundation for fab-to-electrical and parametric-to-wafer-sort correlations,” says Advantest’s Singh. “It identifies fab deviations using upstream process data and real-time electrical measurements, accelerating response times to protect product yield and implement corrective actions consistently across factories.”

Conclusion
Looking ahead, dynamic process control will be critical to addressing the challenges posed by emerging technologies like 6G, AI-driven hardware, and terahertz-level devices. As performance demands continue to grow, the ability to adapt to variability in real time will determine which manufacturers thrive in an increasingly competitive landscape.

The integration of test and metrology data into dynamic process control systems enables a holistic approach to managing the complexities of advanced packaging. By linking in-line inspections, high-frequency testing, and metrology results, manufacturers can identify trends, correlate data across process steps, and implement adjustments in real time. This synergy is indispensable for maintaining yield when even minor process deviations can have cascading effects. And as advanced packaging technologies evolve, the ability to adapt dynamically will remain essential to achieving high yields, preserving profitability, and ensuring reliability in increasingly complex device architectures.