Machine learning model training

Inventors

Hong, Jeong Jin • Cheon, Sejune • KIM, Sang Hong • LI, Thomas Ho Fai • Nottrott, Anders Andelman • Zhu, Zhaozhao • Jang, Mihyun

Assignees

Applied Materials Inc

Abstract

A method includes receiving spectral data of a substrate and metrology data corresponding to the spectral data of the substrate. The method further includes determining a plurality of feature model configurations for each of a plurality of feature models, each of the plurality of feature model configurations including one or more feature model conditions. The method further includes determining a plurality of feature model combinations, where each feature model combination of the plurality of feature model combinations includes a subset of the plurality of feature model configurations. The method further includes generating a plurality of input datasets, where each input dataset of the plurality of input datasets is generated based on application of the spectral data to a respective feature model combination of the plurality of feature model combinations. The method further includes training a plurality of machine learning models, where each machine learning model is trained to generate an output using an input dataset of the plurality of input datasets and the metrology data. The method further includes selecting a trained machine learning model from the plurality of trained machine learning models satisfying one or more selection criteria.

Core Innovation

The disclosed metrology workflow receives spectral data of a substrate and metrology data corresponding to the spectral data of the substrate. It determines a plurality of feature model configurations for each of a plurality of feature models, where each feature model configuration comprises one or more feature model conditions, and determines a plurality of feature model combinations, where each feature model combination comprises a subset of the plurality of feature model configurations.

Based on the feature model combinations, the method generates a plurality of input datasets, with each input dataset generated by applying the spectral data to a respective feature model combination. The metrology data are used together with the input datasets to train a plurality of machine learning models, where each machine learning model generates an output using an input dataset and the metrology data.

After training, the method selects a trained machine learning model from the plurality of trained machine learning models satisfying one or more selection criteria. The selection criteria include at least one selection value criterion based on a selection value such as RMSE, R-squared (R2), or error variance, and may also include processor utilization, processing speed, and memory utilization criteria.

The selected trained machine learning model is tied to critical dimension (CD) profile prediction and is further described as used to generate synthetic microscopy images, including CD profile prediction images and virtual SEM (VSEM) style synthetic imagery, to support more consistent synthetic data.

Claims Coverage

The independent claims cover four core features: determining feature model configurations and combinations, generating multiple input datasets from spectral data, training multiple machine learning models with metrology data, and selecting a trained machine learning model using one or more selection criteria. The claims also specify selection criteria based on RMSE, R-squared, error variance, and resource/performance measures, and allow the output to include critical dimension (CD) profile prediction.

Across the independent claims, the inventive coverage centers on building feature model configurations and combinations to generate input datasets from spectral data, training multiple machine learning models using those datasets with corresponding metrology data, and selecting a trained model using one or more selection criteria. The claims further specify RMSE, R-squared, error variance, and resource/performance measures, and they permit CD profile prediction output.

Stated Advantages

Documented Applications