Results shown here reflect actual Atomscale deployments. Specific metrics are reported where available. Customer identities are confidential.
Automatic Detection of Atomic-Scale Defects
Atomscale automatically detects and flags killer defects during MBE growth by extracting quantitative atomic spacing measurements from streaming RHEED data in real time.Setup
Setup
Streaming RHEED video data was ingested from an MBE chamber during epitaxial growth. Atomscale’s adapter models transformed raw diffraction pattern images into quantitative time series of atomic spacing values. Anomaly thresholds were established relative to baseline spacing (1x), with deviations beyond 2x triggering automatic alerts.
Result
Result
The system identified atomic spacing deviations indicative of defect nucleation and flagged them automatically during active growth runs. This enabled in-process adjustments and run-to-run insight extraction from data that was previously only reviewed qualitatively, if at all.
Why it matters
Why it matters
Defect detection during MBE growth has historically depended on operator experience and visual interpretation of RHEED patterns. Critical defects can nucleate and propagate before a human observer recognizes the signature. Automated quantitative tracking converts a subjective, attention-dependent task into a systematic one, catching defects that would otherwise reach downstream characterization as unexplained yield loss.
“I have been hoping somebody would be able to do this for 20 years.” — Staff Scientist
Early Detection: 40 Seconds Ahead of Expert Operators
Atomscale detected a nucleating surface reconstruction 40 seconds earlier than an experienced operator observing the same RHEED data, with higher consistency and no false positives from image processing artifacts.Setup
Setup
During an MBE growth run, both Atomscale and an expert operator monitored live RHEED pattern evolution. Detection timestamps were compared for the onset of a surface reconstruction change. Atomscale processed the streaming diffraction data continuously, while the operator observed the same feed under normal working conditions.
Result
Result
Atomscale detected the reconstruction change at approximately 41 seconds into the relevant growth phase. The expert operator identified the same transition at approximately 81 seconds, a 40-second lag. Atomscale’s detection was consistent across repeated observations with no noise or artifacts from classical image processing.
Why it matters
Why it matters
In thin film growth, surface reconstruction transitions signal changes in growth mode, stoichiometry, or defect formation. A 40-second detection advantage represents a meaningful intervention window: time to adjust flux ratios, substrate temperature, or shutter sequences before a defect propagates through subsequent layers. For processes with narrow recipe windows, this is the difference between a correctable deviation and a scrapped run.
In-Process Composition Prediction from Diffraction Data
Atomscale’s task-specific AI models predicted chemical composition during active growth by correlating real-time RHEED fingerprints with ex-situ XPS composition measurements from previous runs.Setup
Setup
A training dataset was assembled by pairing RHEED diffraction features captured during growth with post-growth XPS composition measurements from the same samples. Task-specific models were trained to map the real-time diffraction signal to chemical composition values. The models were then evaluated on held-out test samples.
Result
Result
Predicted composition values closely tracked measured values across both training and test datasets, with train and test points following the ideal diagonal on a predicted-vs-measured correlation plot. The model operated across a composition range of 0.0 to 0.4, demonstrating accuracy on new samples not seen during training.
Why it matters
Why it matters
Chemical composition is typically measured only after deposition is complete, using ex-situ techniques like XPS that add hours or days of latency to the feedback loop. By creating in-situ proxies for ex-situ measurements, Atomscale converts post-hoc quality checks into real-time process signals. This means composition drift can be detected and corrected during growth rather than discovered after the fact, saving the time and cost of lost trials.
Real-Time ALD Feedback from Raw Ellipsometry
Atomscale automatically extracted quantitative layer-by-layer comparisons of doped and undoped hafnia films from raw ellipsometry data during ALD, differentiating growth conditions and doping composition without manual analysis.Setup
Setup
ALD runs of HfO₂ and Zr-doped HfO₂ were deposited on SiO₂ with identical recipes except for substrate temperature (200°C, 250°C, 300°C). Raw in-situ ellipsometry data was ingested by Atomscale and processed into material fingerprints at the resolution of individual deposition cycles.
Result
Result
Atomscale’s fingerprints clearly differentiated temperature conditions across the undoped HfO₂ runs and showed pronounced divergence when Zr doping was introduced, even within the single-phase regime where surface roughness is the primary distinguishing signal. Layer-by-layer tracking showed the point at which doped and undoped films began to diverge in material state as thickness increased.
Why it matters
Why it matters
ALD process development for advanced dielectrics like hafnia depends on precise control of composition and crystallization behavior, which vary with temperature and dopant concentration. Ellipsometry data is collected during nearly every ALD run but is rarely used for real-time decision-making because extracting quantitative comparisons from raw optical data requires significant manual effort. Atomscale automates this extraction, making every ALD cycle a source of actionable process intelligence rather than archived raw data.
MOCVD Outlier Detection for Wafer Uniformity
Atomscale trained surrogate models on CVD recipe parameters and sensor data to detect growth outliers resulting in non-uniform wafers, achieving greater than 96% accuracy while tuned for high detection certainty.Setup
Setup
A model was trained to map MOCVD recipe parameters and instrument sensor logs to ex-situ measurements of wafer uniformity. Uniformity was quantified using a distance metric between photoluminescence and Raman spectra taken at the wafer center and edge. The model was deliberately tuned to favor detection certainty over recall — minimizing false positives at the cost of occasionally missing borderline anomalies.
Result
Result
The model achieved greater than 96% accuracy in identifying growth runs that produced non-uniform wafers. This enables real-time drift and outlier detection using data that is already collected during standard MOCVD production runs.
Why it matters
Why it matters
Wafer uniformity is a critical yield metric in compound semiconductor manufacturing, but it is typically assessed only after growth is complete, often only after additional characterization steps. A surrogate model that flags uniformity outliers from in-process data allows engineers to catch drifting processes before an entire cassette of wafers is affected. The >96% accuracy at high certainty means alerts are trustworthy enough to act on without requiring manual verification of every flag.