The semiconductor industry is currently navigating a transition toward sub-2nm process nodes, where the precision of measurement tools must match the atomic-scale dimensions of the transistors they monitor. In a significant advancement for the field of nanometrology, researchers from Purdue University, Intel Corporation, and Bruker Corporation have published a joint study detailing the complex physical interactions that occur when measuring extreme ultraviolet (EUV) lithography features. The paper, accepted for publication in March 2026, provides a critical analysis of how high-aspect ratio (HAR) atomic force microscopy (AFM) tips behave when encountering the delicate sidewalls of photoresist patterns.
As chip manufacturers push toward higher transistor densities, the ability to measure three-dimensional profiles without damaging the sample has become a primary bottleneck. While scanning electron microscopy (SEM) remains the industry standard for high-throughput inspection, it is known to cause "shrinkage" in photoresist materials due to electron beam exposure. The AFM offers a non-destructive alternative, but as this new research demonstrates, it introduces its own set of physical artifacts that can skew measurement data if not properly accounted for.
The Evolution of EUV Metrology and the 40 nm Pitch Challenge
The shift to EUV lithography has allowed the industry to print features at pitches of 40 nm and below. However, at these dimensions, the photoresist structures—the temporary patterns used to etch circuits into silicon—are incredibly fragile and narrow. Measuring the height, width, and sidewall angle of these features is essential for process control. If the measurement is off by even half a nanometer, it can lead to catastrophic failures in the final integrated circuit.
Historically, the industry assumed that the primary limitation of AFM was the physical size of the probe. The logic followed that if one could manufacture a "spike" tip—a high-aspect ratio needle—one could reach into the narrow trenches of a 40 nm pitch pattern and map the surface accurately. The Purdue-Intel-Bruker study challenges this conventional wisdom, suggesting that the mechanical dynamics at the tip-sample interface are far more influential than the geometric shape of the tip alone.
Chronology of the Research and Technical Framework
The collaboration began as part of a multi-year effort to refine metrology for the "Angstrom Era" of computing. The research team utilized advanced diamond-like carbon (DLC) spike tips, which are engineered for extreme hardness and high-aspect ratios, to probe EUV photoresist features.
- Phase I: Data Acquisition (2024–2025): The team performed extensive force mapping-based AFM on 40 nm pitch EUV patterns. Unlike standard AFM, which scans across a surface, force mapping involves taking discrete "taps" or indentations at every pixel to record the force-distance relationship.
- Phase II: Machine Learning Integration (2025): Recognizing that the volume of data was too vast for manual inspection, the researchers trained a random forest machine learning algorithm. This AI model was designed to classify thousands of individual force curves based on specific physical characteristics: linearity, adhesion, and hysteresis.
- Phase III: Comparative Analysis (Late 2025): The AFM data was compared against cross-sectional SEM images. This allowed the team to isolate where the AFM was seeing "ghost" structures or where the SEM was causing the resist to contract.
- Phase IV: Publication and Peer Review (March 2026): The findings were synthesized into the final manuscript, providing a roadmap for future metrology tool development.
Technical Analysis: The Physics of Tip-Sample Interaction
The study’s most significant contribution is the identification of "stick-slip" friction and tip bending as primary sources of measurement error. When a high-aspect ratio tip descends into a narrow trench, it does not simply touch the bottom and retract. Instead, as the tip nears the sidewall of the photoresist, long-range Van der Waals forces and capillary forces begin to pull the tip toward the wall.
The researchers observed that the tip often "sticks" to the sidewall during indentation. As the AFM scanner continues to move, the tip undergoes a "slip" event, creating a jump in the force curve. To a standard AFM controller, this might look like a change in the topography of the sample, when in reality, it is a mechanical vibration of the probe itself.
Furthermore, the study quantified "geometric dilation." Because an AFM tip has a finite radius—even the sharpest DLC spikes are several nanometers wide—the resulting image is a "convolution" of the tip shape and the sample shape. The research found that at the 40 nm pitch, the tip is often nearly as wide as the trench it is trying to measure, leading to a significant loss of fidelity in the bottom corners of the features.
Supporting Data and Comparative Metrics
The study provided a rigorous comparison between AFM and SEM measurements, highlighting the discrepancies that engineers must navigate in a high-volume manufacturing environment.
- Photoresist Shrinkage: The SEM measurements consistently showed features that were 5% to 10% smaller than those measured by the AFM. The researchers attributed this to the electron beam "curing" or collapsing the polymer chains in the EUV resist.
- Sidewall Angle Distortion: Due to tip bending, the AFM tended to report sidewall angles that were more slanted than they actually were. The random forest algorithm identified that force curves on the sidewalls exhibited higher hysteresis (energy loss), a clear indicator of the tip dragging against the resist.
- Tip Longevity: The DLC spike tips showed remarkable durability, but the study noted that even microscopic wear on the tip apex changed the machine learning classification of the force curves, suggesting that real-time tip monitoring is necessary for metrology accuracy.
Industry Implications: Beyond the 2nm Node
The implications of this research are far-reaching for companies like Intel, TSMC, and Samsung as they prepare for High-NA (High Numerical Aperture) EUV lithography. High-NA EUV will enable even smaller features, but it also results in a shallower depth of focus and thinner photoresists. In this environment, the "mechanical artifacts" identified by the Purdue and Intel team will become even more pronounced.
Metrology tool manufacturers, such as Bruker, are expected to use these findings to develop next-generation AFM controllers. By integrating machine learning directly into the tool’s firmware, future AFMs could potentially "subtract" the effects of tip bending and stick-slip friction in real-time, providing a corrected 3D profile that more accurately reflects the physical silicon.
Official Responses and Strategic Perspectives
While official corporate statements often remain guarded regarding specific metrology secrets, the collaboration itself speaks volumes. Intel’s involvement underscores the company’s "IDM 2.0" strategy, which relies heavily on maintaining a lead in process control to regain transistor density leadership.
"The integration of machine learning into physical metrology is no longer an option; it is a necessity," noted a source familiar with the research. "As we move into the Angstrom Era, we are no longer just measuring shapes; we are measuring the interaction of forces. This paper proves that the probe is part of the system, not just an observer."
Academic contributors from Purdue University emphasized that this study bridges the gap between theoretical nanomechanics and industrial application. The use of a random forest algorithm to classify force curves represents a shift toward "data-driven physics," where the complexity of the interaction is managed through computational power rather than just better hardware.
Future Outlook: The Road to Autonomous Metrology
Looking ahead, the semiconductor industry is moving toward a "holistic metrology" approach. This involves combining data from AFM, SEM, and optical scatterometry into a single "digital twin" of the wafer. The Purdue-Intel-Bruker study provides the missing physical model for the AFM component of this triad.
By understanding that high-aspect ratio tips are subject to bending and friction-induced artifacts, engineers can now develop "error budgets" that account for these variables. This will be vital for the production of the next generation of gate-all-around (GAA) transistors, where the 3D geometry of the "nanosheet" stacks must be controlled with sub-nanometer precision.
The research concludes that while HAR tips are a powerful tool, they are not a "silver bullet." The future of metrology lies in the sophisticated interplay between advanced materials (like DLC), high-speed sensing, and machine learning algorithms that can see past the physical limitations of the probe. As the industry approaches the limits of Moore’s Law, the ability to accurately characterize the "interaction dynamics" at the nanoscale will be the deciding factor in the viability of future computing architectures.