AI in Semiconductor Inspection Is Becoming a Two-Layer Game

//php echo do_shortcode(‘[responsivevoice_button voice=”US English Male” buttontext=”Listen to Post”]’) ?>

Artificial intelligence is rapidly reshaping semiconductor defect detection, but not along a single trajectory. Instead, the industry is beginning to coalesce around two distinct, and potentially complementary, architectural approaches. One embeds AI tightly within advanced inspection hardware to extract new physical insight from wafers; the other sits above the manufacturing data stack, correlating signals across tools to flag emerging yield risks.

Exclusive interviews with Kevin Berghoff, CEO of QuantumDiamonds, and Dieter Rathei, CEO of DR Yield, illustrate how these two schools of thought are taking shape as fabs struggle to manage both the physics complexity of advanced nodes and the data deluge of modern production lines.

Together, they point toward a future in which defect detection becomes a multi-layer intelligence problem rather than a single-tool capability.

Physics-informed AI at the sensor level

QuantumDiamonds represents the first camp: companies building AI deeply coupled to proprietary sensing hardware. The firm’s quantum diamond microscope generates magnetic field maps of ICs, which its software then converts into current-flow information inside the chip.

“We built completely in-house the software that classifies defects,” Berghoff said. The company’s core challenge is reconstructing current distributions from magnetic field measurements—an inverse problem that requires physics-informed machine learning rather than purely statistical analysis.

The workflow begins with customer layout data. QuantumDiamonds simulates expected current flow for a known-good die, then compares that baseline with measured magnetic field data from the actual device. AI models classify anomalies that may indicate shorts, opens, or material defects.

Because the approach is rooted in a new sensing modality, the AI is tightly bound to the hardware. “The core of what we do is interpreting those magnetic fields to detect defects,” Berghoff said.

This deep coupling is both a strength and a constraint. On the positive side, the system can surface failure mechanisms that may be invisible to conventional inspection. Berghoff said the company has demonstrated the ability to find defects that would otherwise remain hidden, though he declined to quantify yield improvements.

However, the current generation of systems is optimized for sample-based analysis rather than full inline inspection. Wafers are pulled from the line and analyzed offline, with full inline capability still several years away. Even so, Berghoff sees the long-term trajectory clearly. The “golden end state,” he said, is to fuse multiple physical data sources—optical, X-ray, and magnetic—into a unified defect classification framework.

That vision hints at the second major AI architecture now gaining traction.

A fab-wide data intelligence layer

DR Yield represents a very different starting point. Rather than building new sensing hardware, the company focuses on aggregating and analyzing the vast quantities of data already generated across semiconductor fabs.

Founded more than two decades ago, the company’s YieldWatchDog platform was designed to automatically scan manufacturing data and flag anomalies that engineers might miss. “The idea was to integrate all data from different sources,” Rathei said. “There are so many data in the industry, and they are siloed into different data warehouses.”

The platform ingests electrical test results, inline metrology, defect inspection data, and equipment signals—essentially any numerical dataset available in the fab. From there, statistical and AI models look for early warning signals for yield excursions.

Unlike QuantumDiamonds’ physics-heavy approach, DR Yield’s models are primarily statistical, though adapted specifically for semiconductor manufacturing geometries and workflows. “We provide a yield analytics solution,” Rathei said. “As long as the data are numerical, we can look for anomalies in trends or spatial patterns on wafers.”

One of the system’s key strengths is multivariate correlation. Rathei described cases in which individual parameters remained within control limits, but their relationship broke down—an early indicator of tool degradation.

“In one study, pressure and flow inside a tool were both within control,” he said. “However, the correlation between those two broke down because a valve was getting leaky. You didn’t see that if you looked at each parameter separately.”

Such cross-signal intelligence is increasingly important as fabs deploy dozens of specialized inspection tools, each generating its own data stream.

Different layers, not direct competitors

At first glance, the two approaches might appear to compete. In practice, they operate at different layers of the manufacturing intelligence stack. QuantumDiamonds is focused on extracting new physical insight from individual devices; DR Yield is focused on connecting signals across the fab to detect systemic risk.

Rathei acknowledged the distinction directly when asked about defect classification: “The defect classification itself is not something that we try to achieve with YieldWatchDog.” Instead, the company concentrates on high-volume data ingestion, fast access, and anomaly detection across datasets. That positioning makes the technologies more complementary than competitive.

Hardware vendors typically provide analytics specific to their own tools, Rathei noted. DR Yield’s role is to “put this layer on top to connect everything.”

Another key difference lies in deployment timing. QuantumDiamonds today operates primarily in sample-based workflows, with inline capability still under development; DR Yield positions its software as “near-real-time,” with data typically available to engineers within seconds or minutes of generation. In modern fabs, both timescales matter.

Deep physical inspection can uncover subtle failure mechanisms that statistical monitoring might miss. Conversely, fab-wide analytics can surface systemic drift long before individual devices fail catastrophically. The result is an emerging multi-timescale intelligence model.

Both companies also reflect the broader evolution of AI in semiconductor manufacturing. QuantumDiamonds uses physics-informed ML to solve inverse electromagnetic problems; DR Yield employs a mix of statistical models, supervised learning, and unsupervised clustering for wafer pattern analysis.

More recently, DR Yield has begun integrating large language models into its platform. The company’s Yield AIssistant allows engineers to query wafer data conversationally and generate automated reports. “We developed an interface that allows people to analyze the data and ask a language model questions about what they see,” Rathei said.

Even here, however, the company emphasizes human oversight, a theme that resonates across the industry. “It’s an assistant,” Rathei stressed, not an autonomous decision-maker.

Toward a hybrid future

If the two approaches start from different philosophies, both executives ultimately converge on a similar long-term view: Fabs will likely need multiple layers of AI working together. Berghoff’s vision of fusing magnetic, optical, and X-ray data points toward increasingly rich physical inspection stacks. Rathei argues that fab-level correlation will remain essential as process complexity grows.

When asked how he would architect a greenfield fab, Rathei was unequivocal: “I would definitely do both.” At the tool level, specialized AI is needed for signal extraction and physics interpretation. At the factory level, a broader intelligence layer is required to connect the dots across processes and tools. That dual-layer model may ultimately define the next phase of semiconductor manufacturing analytics.

The backdrop to this evolution is the sheer scale of data that modern fabs must manage. As inspection sensitivity improves and process steps multiply, engineers are increasingly overwhelmed by information. Rathei recalled that many major yield excursions historically showed early warning signs in the data—signals that simply went unnoticed because no one had time to sift through the noise. “There are just not enough people in any given factory to deal with all the data,” he said.

At the same time, new sensing modalities, such as quantum diamond microscopy, are expanding the universe of measurable physical effects inside chips. Together, these trends are pushing defect detection toward a layered AI architecture in which no single tool or model has full visibility.

Defect detection is evolving from a point-tool capability into a systems-level intelligence challenge. Physics-aware inspection will continue to push the boundaries of what can be measured on individual devices. Data-centric analytics will increasingly determine how quickly fabs can respond to subtle process drift. The winners in advanced manufacturing will likely be those who can integrate both.

As AI moves deeper into the fab, the question is no longer whether to deploy machine intelligence in defect detection; it’s where—and at how many layers.

---

See also: