While applying machine learning (ML) to semiconductor manufacturing is prevalent, an efficient way to sample the search space has not been explored much in key processes such as lithography, annealing, deposition, and etching. The aim is to use the fewest experimental trials to construct an accurate predictive model. Here, we proposed a technology computer added design (TCAD)-assisted meta-learned sampling approach. The meta-learner adjusts the way of sampling in terms of how to hybridize the TCAD with ML when selecting the next sampling point. While an advanced semiconductor process is expensive, efficient sampling is indispensable. Using laser annealing as an example, we demonstrate the effectiveness of the proposed algorithm where the mean square error (MSE) at the first 100 sampling steps using TCAD-assisted meta-learned sampling is significantly lower than the pure ML approach. Besides, with reference to the pure TCAD approach, the TCAD-assisted sampling prevents the MSE degradation at 200−400 sampling steps. The proposed approach can be used in other manufacturing or even any applied machine intelligence fields.
Promoted model architectures or algorithms are crucial for intelligent manufacturing since developing them takes a lot of trial and error to embed the domain knowledge into the models correctly. Especially in semiconductor manufacturing, the whole processes depend on complicated physical equations and sophisticated fine-tuning. Therefore, we use a neuroevolution-based model to search the optimized architecture automatically. The collector current value at a particular bias of the silicon–germanium (SiGe) heterojunction bipolar transistor, generated by technology computer-aided design (TCAD), is used as the target dataset with six process parameters as the inputs. The processes include oxidation, dry and wet etching, implantation, annealing, diffusion, and chemical–mechanical polishing. Our work can build a suitable model network with a fast turnaround time, and practical physical constraints are fused in it without domain knowledge extraction. Take the case with 3840 data and one output as an instance. The mean square errors of the train set and validation set, as well as the mean absolute percentage error of the test set, are 1.317 × 10–6, 7.215 × 10–7, and 0.216 while using multilayer perceptron (MLP) and they are 3.285 × 10–7, 1.661 × 10–7, and 0.097 while using NE. The consequences show that the work in this vein is promising. According to the trend plot and results, the ability to extract physic is much better than the traditional (MLP) model.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.