Detecting Chemical Variability At Advanced Nodes

Key Takeaways

• Yield loss is increasingly driven by molecular variability in thin films, interfaces, and contamination rather than visible defects.
• Reliability issues often appear first as parametric drift or margin erosion under workload and thermal stress.
• Detection requires correlating molecular metrology, embedded electrical telemetry, and AI-driven wafer inspection.

As semiconductor manufacturing pushes toward angstrom-scale devices and heterogeneous integration, the nature of defects affecting yield and reliability is changing. At advanced nodes and in multichip schemes, a growing share of yield loss originates in subtle variations within materials rather than in visible defects.

For decades, the industry relied on inspection and electrical test strategies to identify structural failures, such as particles bridging metal lines, patterning defects, or open interconnects. Those failures remain important, but they are no longer the only sources of variability in leading-edge processes.

Thin-film composition changes, bonding irregularities at interfaces, process residues, and microscopic shifts in chemical structure can all influence device behavior. These effects often remain invisible to traditional inspection systems and may not immediately produce a functional failure. Instead, they appear as a gradual parametric drift that erodes electrical margins over time. This kind of degradation may only reveal itself under specific workloads or thermal conditions, long after the device has left the fab. Detecting them requires tools and strategies operating at three distinct levels — molecular characterization at the material surface, electrical monitoring at the circuit level, and AI-driven correlation at the wafer scale.

“Traditional defectivity tests focus on clear, binary failures,” said Nir Sever, senior director of business development at proteanTecs. “You either have a stuck-at fault, a resistive open, a bridge short, or a functional failure. But material or interface instability behaves differently. It manifests as parametric drift, intermittent marginality, or workload-dependent degradation long before it becomes a permanent defect.”

In high-performance computing systems and AI accelerators operating close to their power and thermal limits, even small shifts in materials behavior can alter timing stability, signal integrity, or long-term reliability. Catching them requires a different kind of analysis.

Materials complexity is rising
One reason these problems are becoming more pronounced is the rapid expansion in materials used throughout semiconductor manufacturing and advanced packaging flows. Traditional front-end processes relied primarily on silicon and a relatively small set of well-understood thin films.

Today’s heterogeneous integration schemes introduce polymer dielectrics, bonding metals, adhesives, encapsulation materials, redistribution layers, and thermal interface compounds. Each has unique mechanical and chemical characteristics that interact with the materials around them. But the complexity arises not simply from the materials themselves, but from how they interact at every stage. A small variation in chemical bonding early in the process flow may not become visible until multiple downstream steps have amplified its effect.

“A lot of these defects are things that you cannot really see,” said Tiago Tavares, program and project manager at Critical Manufacturing. “They are either hidden or they are small deviations that you neglect. But they need to be captured. We need to find new ways of measuring this, ways of early detecting this, so that we can plug it into the equation.”

At the same time, the films at the center of these interactions are themselves shrinking toward scales where their behavior changes fundamentally.

“When you deposit a very thick film, the interface means nothing,” said Hichem M’Saad, CEO of ASM. “Ten angstroms out of microns is negligible, but when you deposit an ALD film that is only five angstroms thick, that interface becomes a big deal.”

At such dimensions, slight differences in precursor chemistry, atomic bonding, or surface preparation directly influence how atoms organize within the film. Those differences may be too small to detect through conventional inspection methods, yet they can shift transistor threshold voltages, alter carrier mobility, or compromise gate oxide integrity. Those effects accumulate silently, surfacing only after a device has left the fab.

When defects don’t look like defects
The difficulty of detecting chemical variability stems from the fact that these effects rarely produce visible defects. Instead, they alter electrical characteristics gradually and statistically.

Small changes in film composition may shift threshold voltages or reduce carrier mobility. Variations in bonding interfaces can increase interconnect resistance or degrade signal propagation across high-speed logic paths, and residues may trigger aging mechanisms such as electromigration or bias temperature instability (BTI). These small degradations can accumulate over time and may not manifest as functional failures until the device is operating in the field.

“Parametric drift is often the electrical manifestation of underlying physical changes in the silicon or package,” said Sever. “Contamination and variability during manufacturing can introduce traps or weak interfaces in dielectric or metal layers. Over time, those defects may not cause immediate failure, but they can accelerate aging mechanisms such as BTI  or electromigration. The electrical signature becomes gradual margin erosion or abnormal degradation slope compared to the population.”

In all cases, the physics changes first, and the electrical parameters shift next. What makes these effects particularly difficult to manage is that they often appear only under specific combinations of workload, thermal load, and supply conditions, which are characteristic of sustained operation in production AI systems and high-performance compute clusters.

These effects typically surface as subtle shifts in distribution across a population of devices, manifesting as an increasing spread in timing margins, unexplained leakage, or intermittent failures tied to specific workloads. One reason these variations remain difficult to isolate is that conventional electrical characterization often averages away the very outliers engineers need to find. Many test structures were originally designed to confirm connectivity or detect gross failures rather than resolve subtle variations within individual interconnects or bonding regions. When thousands of structures are chained together in a single measurement, localized variability can disappear into the aggregate result.

“Traditional data chain methods have thousands of vias connected together. When you divide the chain by ten thousand, you’re not going to find the outliers,” said Jesse Ko, COO at Modus Test. “The averaging effect hides the variability. So, what we’re trying to do is focus on specific critical structures so customers can characterize those regions more accurately.”

Standard test regimes are designed to answer a binary question: Does the device function? They are not designed to trace electrical anomalies back to their chemical origins, or to distinguish between a film whose composition drifted during deposition and one that formed a weak interface due to surface contamination. Making that connection requires correlating electrical data across populations and over time, with enough resolution to separate material-induced degradation from ordinary process spread.

“Classic defect detection answers, ‘Is it broken?’” said Sever. “Deep telemetry answers ‘Is it becoming unstable, and why?’ When we correlate timing margin degradation with voltage noise, workload stress, thermal gradients, and interconnect behavior, we can differentiate between random transient stress, intrinsic aging, latent production defects, and material or interface instability.”

Detecting at the molecular level
Identifying chemical variability before it becomes an electrical problem requires characterization tools capable of resolving material differences at the molecular scale — determining what chemical or bonding condition produced it rather than just detecting whether a defect is present. That distinction matters because two defects with identical electrical signatures may have entirely different physical origins, and the corrective action for each is different.

Conventional techniques for elemental analysis, such as energy-dispersive X-ray spectroscopy (EDS), work reasonably well for heavier elements, but they struggle with the light elements that dominate the bonding chemistry of advanced dielectrics and surface contaminants. Carbon, oxygen, nitrogen, and hydrogen are central to the variability problems engineers face in high-k and low-k dielectric films, organic surface adders, and fluoropolymer contamination residues, yet they fall below the reliable detection thresholds of many standard analytical methods.

“The window that our tool looks at is in the space of lighter elements,” said Cassandra Phillips, nanoIR product manager at Bruker. “EDS tends to work on heavier elements that have a larger X-ray response. When you start to talk about light elements — carbon, oxygen, nitrogen, hydrogen — those techniques struggle, and they only provide elemental identification, which can be insufficient. When you’re looking at surface adders, small amounts of grease, lubricants, fluoropolymers, and polyethylene, other techniques aren’t able to distinguish them.”

The bonding state matters as much as the elemental composition. Silicon dioxide, for example, is chemically simple but structurally complex. It can exist in single-crystal, polycrystalline, or amorphous forms, each with different bonding densities and surface compositions. At the surface, the specific bonds expressed (silicon hydride versus silicon oxyhydride, for instance) produce large differences in downstream electrical performance, even when bulk elemental analysis shows identical composition.

“You don’t just look at the direct, short-scale chemical bonding,” said Phillips. “You also look at the chemical environment of the material around it. How close are these oxygens to other oxygens? How close are these silicons to other silicons, even if they’re not directly bonded? When you get into amorphous domains, these problems scale exponentially. You no longer have a known surface composition with a known bonding density. Infrared responses let you get some order from all of that.”

Nanoscale infrared spectroscopy, or nano-IR, is the technique that makes this level of analysis possible at semiconductor-relevant dimensions. The method combines an atomically sharp AFM probe with a tunable infrared light source. When the IR light illuminates the sample surface, materials that absorb at that wavelength convert the energy to heat, producing a minuscule thermal expansion on the order of bond radii. The AFM probe detects that expansion as a mechanical force on its cantilever. By pulsing the laser at a frequency matching a cantilever resonance, the signal is further amplified, enabling detection of sub-nanometer-thick materials.

Fig. 1: How nano-IR spectroscopy works. Source: Bruker

The critical advantage over conventional IR microscopy is resolution. Standard infrared techniques are limited by optical diffraction. The smallest resolvable feature is roughly half the wavelength of the light, which puts the floor at several microns. Nano-IR uses the AFM probe as the detector rather than collecting the scattered light directly, which means lateral resolution is determined by the probe geometry rather than the wavelength. The result is sub-10nm lateral resolution from a technique that would otherwise be limited to features orders of magnitude larger.

“There has been significant progress in detecting smaller and thinner amounts of material,” said Bruker’s Phillips. “Our nano-IR systems can detect sub-nanometer-thick materials with chemical identification capabilities and less than 10 nanometers laterally. That translates to attograms of material.”

An important practical advantage is that nano-IR is non-destructive. Techniques such as secondary ion mass spectrometry and atom probe tomography provide chemical information, but they consume the sample in the process, making them unsuitable for in-line monitoring or iterative analysis of the same region. Nano-IR can be applied before and after a process step without sacrificing the material under study, which is a significant consideration as the industry moves these capabilities from the failure analysis lab toward inline or near-line deployment.

For visible defects or particle adders with known coordinates, established scattering techniques can localize the target and direct subsequent chemical identification to the right site. The harder problem is thin-film inhomogeneity — variability distributed through the bulk of a film rather than concentrated at a discrete location. There is no coordinate to navigate to. The variation is the film itself. That is where nano-IR provides detection capability that localized techniques cannot.

Hidden differences in “identical” materials
One of the most difficult aspects of chemical variability is that materials that appear identical under standard process monitoring may still behave differently at the molecular level. The bonding network at the surface can differ in ways that bulk analysis does not capture, but that downstream electrical performance reflects.

“We often see cases where wafer one works electrically and wafer two doesn’t,” said Phillips. “The only difference we know between the two of them is the vendor who supplied the material. They’re supposed to be molecularly homogeneous between the two samples, but when we analyze them, we can identify the molecular reason for why it’s not working.”

The supply chain is significant here. Suppliers are often unaware of, or unwilling to disclose, subtle variations in their own processes, and the problem is difficult to model without that data. The characterization burden falls entirely on the manufacturer who encounters the failure, often well downstream from where the root cause originated.

Molecular characterization identifies the chemical condition. Detecting its electrical consequences in production before they accumulate to the point of failure requires a different approach. Embedded monitoring circuits distributed across the die provide a continuous readout of electrical behavior under actual operating conditions, capturing the parametric signatures of material instability in real time.

That approach answers questions that end-of-line testing cannot. A device that passes functional testing may still be operating with degraded margin in a specific clock domain or interconnect region, close enough to its limits that sustained workload will push it over the edge. Without monitoring those margins continuously, the degradation is invisible until it becomes a failure.

“By analyzing patterns across devices, locations, workloads, and time, you can separate systematic design headroom limitations from physical instability mechanisms,” said Sever. “Design margin issues are present immediately and remain relatively stable. Material-related variability often evolves. You may observe abnormal aging acceleration, increased sensitivity to voltage or workload stress, or progressive timing compression that was not visible at characterization.”

In advanced packaging environments, where microbumps, through-silicon vias, and hybrid bond interfaces connect dies in close thermal and mechanical proximity, parametric monitoring is especially valuable. Variability at any one interface can produce effects that propagate across the assembly, such as increased interconnect resistance, lane-to-lane asymmetry, or localized timing margin compression, without producing the kind of clean binary failure that structural test is designed to catch.

Simulation and the modeling gap
Closing the gap between materials-level variability and its electrical consequences also requires the ability to model how molecular-level variations propagate through a device’s electrical behavior. That is where current simulation approaches face a structural limitation. A model is only as accurate as its material property inputs. When two nominally identical films differ at the molecular level in bonding state, interface chemistry, or precursor purity, those differences are rarely captured in the parameters the model depends on.

“Variability includes the realistic ranges and combinations of temperature, process, voltage, pressure, and material properties that can come into play,” said Marc Swinnen, director of product marketing at Synopsys. “The millions of possible combinations of these variable factors are difficult to comprehensively validate.”

The deeper challenge is that individual effects are not independent. Mechanical stress alters electrical parameters. Thermal gradients shift material properties. Chemical variations at an interface change how electrons propagate across it. Simulating any one of these in isolation produces results that diverge from reality when the others are active simultaneously, which they always are in production.

“Individual effects are inter-related and require a multi-physics approach to simulate accurately,” said Swinnen. “Simulators often simulate effects separately and don’t capture how one quantity affects the others. Mechanical stress affects not only reliability but also changes the electrical parameters of stressed devices, but mechanical and electrical are rarely considered together.”

This disconnect is especially consequential for chemical variability, where the effects being modeled are themselves poorly characterized. Engineers cannot simulate the impact of a bonding irregularity they have not yet detected or model the electrical behavior of a contamination layer that existing inspection tools did not flag. The simulation and the metrology must advance together, each informing what the other looks for.

Yield implications at scale
The case for investing in detection at all three levels — molecular, electrical, and wafer scale — ultimately rests on economics. Although chemical and molecular variations may appear small, their consequences can scale dramatically across high-volume production. Semiconductor devices at leading nodes can sell for tens of thousands of dollars per unit. At that price point, even fractional yield improvements translate directly into substantial gains.

“This is one of the central questions in semiconductor metrology and failure analysis right now,” said Phillips. “Even small percentage point increases in yield can lead to millions or billions in profit or loss right on either side of that line.”

At the wafer scale, AI-driven inspection is beginning to close the correlation gap between what inspection tools see and what actually affects yield. The historical weakness of wafer inspection, large volumes of flagged defects with poor correlation to electrical performance, is precisely the kind of noise problem that makes chemical variability hard to extract from the signal.

“Traditionally, inspection tools produced large volumes of nuisance defects that had little correlation with yield loss,” said Woo Young Han, product marketing director at Onto Innovation. “With AI-driven image recognition, inspection results can now be brought into closer alignment with electrical test outcomes. Beyond fixed defect-detection thresholds, AI can dynamically adjust sensitivity based on defect type, wafer location, and prior lot history, improving capture of yield-impacting defects while filtering noise.”

Conclusion
Detection tools now exist at each level of the problem. Nanoscale infrared spectroscopy can resolve bonding states and surface chemistry at sub-10nm resolution, identifying the molecular differences between films that look identical but perform differently. Embedded parametric monitors can track timing margin degradation, interconnect resistance drift, and workload-dependent instability across populations of devices in real time, separating material-induced effects from design margin limitations. AI-driven inspection at the wafer scale is narrowing the gap between what inspection tools flag and what actually affects electrical performance, filtering noise while improving capture of the variations that matter. Each of those capabilities represents a genuine advance over where the industry stood even five years ago.

What remains incomplete is the integration between them. Molecular characterization generates data about bonding states and surface chemistry. Electrical monitoring generates data about parametric behavior. Wafer-scale inspection generates data about defect populations. None of those datasets is particularly useful in isolation. The bonding anomaly that nano-IR identifies at a specific surface location means little without the electrical history that tells engineers whether that condition ever produced a yield loss. The timing margin degradation that embedded monitors detect means little without the materials characterization that traces it back to a specific interface condition or film composition. Closing the loop requires not just better tools at each level, but a shared analytical framework that lets data from all three inform one another.

The economics make the urgency clear. At leading-edge nodes, yield losses measured in fractions of a percent translate into hundreds of millions of dollars across a production program. The variations that produce those losses are increasingly molecular in origin, invisible to structural inspection, and increasingly capable of evading end-of-line test until they surface as field failures. Detecting them is no longer optional. The question is whether manufacturers can build the connected data infrastructure fast enough to stay ahead of the complexity they are simultaneously introducing.