Plasma OES Monitoring: Etch Endpoint Detection and Process Anomaly Diagnosis

Author: MST Semiconductor | Category: Semiconductor Process Optimization | Keywords: OES, Plasma Optical Emission Spectroscopy, Etch Endpoint Detection, FDC

What OES Actually Measures

In simple terms, OES works by directing an optical fiber at the observation window of a plasma chamber, collecting the light emitted by the plasma, and sending it to a spectrometer for dispersive analysis. Gas molecules in the plasma, excited by RF energy, emit characteristic light — different atoms and molecules correspond to different emission wavelengths. For example, fluorine atoms have a strong emission peak at 703.7nm, oxygen atoms near 777nm, CN radicals at 388nm, and silicon reaction products such as SiF show signal around 440nm.

A typical OES spectrometer covers the 200-800nm range at roughly 0.3nm resolution, capturing 2,048 data points per acquisition. These data fundamentally reflect relative concentration changes of various reactive species in the plasma — I say “relative” because OES is strictly semi-quantitative: it can tell you that “a particular species concentration is increasing,” but cannot directly provide absolute concentration values.

The greatest advantage of OES is that it is completely non-invasive. The optical fiber attaches to an observation window on the chamber sidewall; no probe needs to be inserted into the chamber, and the plasma state itself is not disturbed. This advantage is extremely important on production lines — you do not need to modify the chamber design or interrupt production to install a sensor.

The Classic Battleground: Etch Endpoint Detection

The most mature and widely used application of OES is EPD (End Point Detection). The principle is straightforward: when you etch through material layer A and reach the interface with layer B, the chemical environment in the plasma undergoes an abrupt change. For example, when etching silicon oxide with a fluorine-based chemistry and you break through to the underlying silicon nitride layer, the types and concentrations of fluorine-containing reaction byproducts change, and the corresponding spectral signals shift accordingly. By monitoring the intensity of specific wavelengths over time and identifying the inflection point, you find the endpoint.

When I was working on poly-Si gate etch, I typically monitored the SiF signal at 440nm — once the poly-Si is etched through, the underlying gate oxide barely reacts with fluorine, causing the SiF signal to drop abruptly. This transition is very distinct with a good signal-to-noise ratio. But in high-selectivity etch scenarios, such as SAC (Self-Aligned Contact) etch where silicon oxide must stop on silicon nitride, the monitored wavelength needs to switch to the CN radical signal at 388nm.

Small Open Area Ratio: A Perennial Challenge

The biggest nemesis of EPD is small open area ratio. The logic is simple: if less than 1% of the wafer surface is exposed to the plasma for etching, the total amount of reaction byproducts generated is very small, and the spectral signal change may be only a fraction of the background noise. I recall one instance doing via etch with roughly 0.5% open area — single-wavelength monitoring simply could not identify the endpoint. We eventually switched to a multi-wavelength joint detection algorithm, simultaneously tracking 4-5 correlated wavelengths and applying weighted normalization, which barely allowed us to extract the endpoint. This was one of the driving forces behind the emergence of full-spectrum analysis methods.

Beyond Endpoint Detection: OES for Chamber Diagnostics

Starting around 2010, the industry began using OES not just for EPD, but as a chamber condition monitoring tool. The reasoning is as follows: the full optical spectrum of the plasma is essentially a “fingerprint” of the chemical environment inside the chamber. If the chamber is in normal condition, the spectrum produced by the same recipe should be highly consistent. If something is wrong inside the chamber — excessive wall residue buildup, a micro-leak in an O-ring, or a drift in actual MFC flow rate — the plasma’s chemical composition will be affected, and the spectrum will shift accordingly.

This application no longer focuses on just one or two characteristic wavelengths, but requires analyzing changes in the shape of the entire spectrum. Here is a real example: on an ICP etcher, we noticed that after running more than 800 wafers consecutively, the relative intensity of the O atom peak at 777nm was slowly rising while the F atom peak at 703nm was slightly declining. Looking at conventional sensor data in the FDC system (RF power, pressure, temperature), nothing appeared abnormal. But the spectral data clearly revealed the trend. A subsequent PM inspection found that the upper electrode had worn beyond expectations, with micro-cracks in the Al2O3 surface causing trace amounts of oxygen-containing contaminants to be released into the chamber.

This type of slow drift is very difficult to catch with conventional sensors, because the RF matching network automatically compensates for impedance changes, making reflected power look perfectly normal. But plasma chemistry cannot lie — OES spectroscopy truly “sees” what is happening inside the chamber.

Traditional PCA Dimensionality Reduction vs. AI Full-Spectrum Analysis

When facing 2,048 wavelength points of full-spectrum data, the classic approach is PCA (Principal Component Analysis). The high-dimensional spectral data is projected onto a few principal components, and statistical process control is performed in this low-dimensional space — plotting T-squared statistics or Q-residual control charts. This method has been in the academic literature for over twenty years and is indeed practical in engineering.

However, PCA has several practical limitations. First, it is linear. Plasma chemical processes are highly nonlinear, especially in multi-step etch recipes where the spectral change patterns are completely different between steps. Using a fixed set of principal components to describe all steps produces suboptimal results. Second, PCA has weak anomaly localization capability — it can tell you “this wafer’s spectrum is abnormal,” but not “the abnormality is likely due to elevated Cl2 flow” or “chamber temperature drift.”

In recent years, AI methods have brought improvements in two main areas. First, convolutional neural networks or autoencoders process the full spectrum without assuming linearity, capturing more complex spectral pattern changes. Second, explainable AI techniques such as gradient backtracking or SHAP values attribute anomalies to specific wavelength regions or even specific chemical species, helping engineers rapidly identify root causes. In our testing, XGBoost achieved over 95% classification accuracy for OES anomaly detection, specifically for MFC flow drift and chamber leak faults — the PCA plus T-squared approach achieved only about 78% on the same dataset.

OES + FDC Combined Diagnostics: The Value of Data Fusion

Although OES spectral data is information-rich, it has blind spots. For example, ESC (Electrostatic Chuck) helium backside pressure leaks barely affect plasma chemistry, so the spectrum will not reveal them. Similarly, abnormal wafer backside temperature is not directly reflected — the spectrum represents the gas-phase environment in the chamber, not the thermal state of the wafer itself.

Therefore, the truly effective approach is to fuse OES data with other sensor data from the FDC system. A typical etch tool’s FDC can collect dozens of signals: RF forward/reflected power, chamber pressure, MFC flow feedback for each gas line, ESC temperature and helium backside pressure, matchbox capacitor positions, and more. Placing these time-series signals alongside synchronously collected OES spectral data within a unified analytical framework — where they cross-validate and complement each other — significantly improves both detection sensitivity and false alarm rates.

To give a concrete figure: on a customer’s Si etch platform, we conducted a comparative study. Using only traditional FDC sensor data for anomaly detection yielded a miss rate of approximately 12% and a false alarm rate of approximately 8%. After incorporating full-spectrum OES data for fusion analysis, the miss rate dropped below 3% and the false alarm rate was controlled to around 2%. The most significant improvements were in chamber contamination faults and slow drifts in gas delivery systems — precisely the fault types that traditional sensors are insensitive to but OES spectroscopy can capture.

Practical Challenges You Cannot Avoid in Deployment

Fiber Aging and Window Contamination

This is one of the most persistent headaches for OES on production lines. Optical fiber transmission efficiency slowly degrades over time, particularly in the ultraviolet range. Meanwhile, after processing several thousand wafers, the inner surface of the chamber observation window accumulates a thin film deposit (polymer or reaction byproducts), reducing overall transmittance. The combined effect of these two factors is that the absolute intensity of the spectral signal continuously drifts, and this drift has nothing to do with the process itself.

Without correction, over time EPD sensitivity will decline and full-spectrum analysis models will generate excessive false alarms due to baseline drift. Several solutions exist: one is to perform periodic spectral intensity calibration, using the background spectrum during chamber idle as a reference; another is to use intensity ratios between wavelengths rather than absolute intensities — for example, the 703nm/777nm ratio can largely cancel out the effect of overall transmittance decline; a third is to incorporate temporal features into the ML model, allowing the model itself to learn and compensate for this drift.

Data Volume and Real-Time Processing

Full-spectrum OES typically samples at 10-100 times per second, with 2,048 data points per acquisition. If a single wafer’s etch process lasts 2 minutes, that produces 1.2 million to 12 million data points. Across an entire production line with dozens of tools running 24/7, the data volume is substantial. Balancing real-time analysis requirements against algorithm complexity and hardware compute resources requires careful engineering trade-offs.

NeuroBox E3200: Bringing OES Analysis to the Production Floor

The technical directions discussed above are not new in academia — papers have been published on them for over a decade. But very few of these have actually made it into production, precisely because the engineering details of deployment are so extensive: data interfacing, signal calibration, model maintenance, integration with MES/OCAP systems, automated response workflows after anomaly detection — every link in the chain must hold.

MST Semiconductor’s NeuroBox E3200 online FDC system was designed from the ground up to treat OES data as a first-class citizen. The system natively supports data ingestion from mainstream OES instruments (HORIBA, Ocean Optics, etc.), includes built-in spectral preprocessing pipelines (baseline correction, wavelength alignment, intensity normalization), provides multiple analysis algorithms ranging from traditional PCA to deep learning, and unifies OES analysis results with other FDC sensor data within a single diagnostic framework. Engineers can view correlated analysis of spectral anomalies and sensor anomalies on the same interface, without switching between different software tools.

More critically, E3200’s adaptive baseline correction feature automatically tracks signal drift from fiber aging and window contamination, maintaining detection sensitivity without requiring frequent manual calibration. In real-world production operations, this eliminates a significant maintenance burden.