Bipolar-CMOS-DMOS (BCD) devices require advanced metrology for their Deep Trench Isolation (DTI) structures, which provide electrical separation between circuit blocks. The primary need is to monitor post-etch trench depth and bottom width. While cross-section imaging is viable, it is time consuming and destructive. Alternatively, scatterometry systems that perform Optical Critical Dimension (OCD) measurements are fast and non-destructive.
Measurements of etch depth and trench widths are standard outputs for OCD systems, but the large DTI dimensions lead to sensitivity problems. However, new algorithms and machine learning can enable measurements, when accompanied with optical data sensitive to the critical parameters. Fitting this description, the n&k OptiPrime-CD couples advanced modeling with a wide spectral range of polarized reflectance (190-1000nm), covering the Deep-Ultraviolet (DUV) to Near-Infrared (NIR). Standard OCD measurement results are based on fitting the data to an analytical model, using Rigorous Coupled Wave Analysis (RCWA) or a Scalar formulation. The use of machine learning adds new capabilities for dimensions which are not easily characterized with standard methods.
For the case of a DTI structure with 8μm pitch, the measured reflectance is not fully coherent, resulting in inaccurate RCWA-calculated reflectance. To supplement RCWA, we utilize an approximate model, known as Scalar, where the reflectance of each feature (line and space) is calculated independently in a plane-wave approximation. This allows for both coherent and incoherent formulations, and one can integrate the Scalar model with RCWA and introduce a coherence parameter to blend the two models for better data fitting.
Trench depth, top width and film thickness measurements are reliable and accurate, for the blended RCWA and Scalar model. Bottom width, however, can be measured well within a certain range of values but appears unreliable when testing a large range and higher values of bottom width, approaching 2μm. The solution is to incorporate machine learning into the analysis for the bottom width measurement.
A decision tree machine learning model is implemented for the bottom width. Inputs to the algorithm include standard modeling results from RCWA and Scalar, reflectance spectra and SEM comparison values at given locations. The data set is divided into two groups, with 80% used for model training and 20% for testing of untrained data. With several hundred data points, the correlation to reference values is quite high (>0.98 RSQ) for both training and test groups, validating the method.
The RCWA, Scalar and machine learning methods are combined into a single model, which has been tested with a large set of test wafers, covering a range of trench depths and bottom widths. Results are evaluated for quality in several ways: 1) Correlation with the nominal conditions and cross-section SEM, 2) Realistic mapping uniformity, and 3) Reproducibility tests.
With a single model, combining RCWA, Scalar and machine learning, there is a clear match with the intended conditions and the SEM measurement results. Mapping results show typical center to edge wafer uniformity, without any significant outliers. Finally, reproducibility tests demonstrate the measurement precision is well within the process control limits. Under all conditions, the measurement results meet the process control requirements for accuracy and stability.
