Heidi Olson, Michael De Zeeuw, Barry Wissman, k-Space Associates Inc, Dexter, MI
Modern thin film deposition processes generate rich streams of in-situ metrology data that remain largely underutilized. Translating that data into actionable process decisions requires more than measurement hardware; it demands intelligent data infrastructure. This work describes k-Space’s effort to build a framework combining high-throughput data streaming, multivariate analysis, and custom analytical algorithms to transform in-situ metrology into closed-loop process control.
Central to this framework is the recognition that modern computing power, when paired with purpose-built algorithms, can extract process-relevant information from complex metrology signals in real time. A high-throughput image streaming architecture delivers full-frame RHEED data — up to 1.58 MP at 70 fps — directly to user analysis environments or back into the deposition tool for real-time correction, expanding the volume and fidelity of data for process inference.
Principal Component Analysis (PCA) plays a key role by reducing high-dimensional RHEED image data to a compact set of physically meaningful components while preserving dominant process information. Applied to representative MBE growth datasets, PCA projection coefficients resolve intensity oscillations, identify surface phase transitions and reconstructions as changepoints, and reveal crystallographic symmetry in rotating RHEED patterns. PCA reprojection error serves as a sensitive, real-time anomaly indicator: when surface behavior departs from the learned baseline, error increases sharply, providing an automated flag for unexpected process excursions.
kSA FitTool illustrates how custom algorithms can be embedded directly into this infrastructure: performing iterative optical model fitting on spectral reflectivity and transmissivity data (375-900 nm) in real time to extract multilayer film thickness, with incoherent layer support for transparent substrates and fitting times ranging from ~10 ms for single-layer films to under 100ms for complex multi-layer stacks.
Together these capabilities define a scalable intelligence layer for in-situ metrology with AI/ML-ready data outputs, resulting in a clear path toward autonomous closed-loop control.