Human-machine collaboration for improving semiconductor process development

The development of advanced semiconductor chips, crucial for AI systems, relies heavily on complex chemical plasma processes like etching and deposition. These processes are currently developed manually by highly skilled engineers using trial and error, a costly and inefficient approach. While AI algorithms have demonstrated superiority in complex tasks like chess and Go, their application to semiconductor process development faces the challenge of "little data." The high cost of experiments on silicon wafers limits the data available for training accurate predictive models at the atomic scale. This research explores the potential of AI, specifically Bayesian optimization algorithms, to significantly reduce the cost of developing these complex processes. The central research question is whether AI can reduce the cost-to-target (the number of experiments required to achieve the desired process outcome) compared to experienced human engineers. The importance of this research stems from the potential to revolutionize semiconductor manufacturing, a cornerstone of modern technology, and accelerate advancements in AI itself, creating a synergistic feedback loop. The study's impact extends beyond the semiconductor industry; insights gained could be applicable to other complex engineering problems characterized by a scarcity of experimental data.

The paper references existing literature on the International Roadmap for Devices and Systems, highlighting the increasing cost of semiconductor process development. It cites work on plasma processing and the challenges associated with it. Previous research on AI's success in complex tasks such as games is mentioned, contrasting it with the high cost of data acquisition in semiconductor processing. The "little data problem" is discussed, referencing studies on applying machine learning to limited datasets in materials science. The paper also acknowledges the use of Bayesian optimization in other semiconductor industry applications, citing relevant publications. Finally, it touches upon existing research on algorithm aversion and the challenges of human-computer collaboration in AI.

This study employed a novel virtual process game to compare the performance of human engineers and computer algorithms in developing a semiconductor fabrication process. The game simulates a single-step plasma etch of a high-aspect-ratio hole in a silicon dioxide film. A proprietary simulator, calibrated using existing data, converts input tool parameters ("recipe") into output etch results. Six professional process engineers (three senior, three junior) participated, along with three inexperienced individuals as a control group. The engineers developed their experiments based on mechanistic hypotheses and their experience. Three different Bayesian optimization algorithms (Algol, Algo2, Algo3) were used, each employing different sampling methods, surrogate models, and acquisition functions. The algorithms operated without prior training data. The key metric was cost-to-target, which includes wafer, metrology, and tool operation costs. A human first-computer last (HF-CL) strategy was also evaluated. In this hybrid approach, the expert engineer provided initial data to guide the algorithm, after which the algorithm took over. The experiment was conducted multiple times for each algorithm and human participant to ensure statistical significance. Data points were transferred from human to algorithm at various stages of process development, allowing for the examination of the optimal transfer point.

The results showed significant differences between human engineers and algorithms. Senior engineers achieved the target at roughly half the cost of junior engineers, emphasizing the importance of experience and domain knowledge. Computer algorithms, starting without prior knowledge, performed poorly compared to the expert human, with a success rate of less than 5%. This underscores the challenge posed by the "little data problem." The HF-CL strategy, however, demonstrated a significant improvement. The optimal transfer point from human to algorithm resulted in a median cost-to-target of $52,000, almost half the cost achieved by the expert human alone. The optimal transfer point showed a V-shaped dependence on the amount of expert data provided. Early transfers resulted in insufficient guidance for the algorithm, while late transfers made the human's contribution a costly burden. The study observed that algorithms adopted multivariate parameter changes, contrasting with the human engineers' preference for univariate or bivariate changes. The algorithms also used batch sizes of one, deemed inefficient by the engineers. These differences highlight potential cultural challenges in integrating human and AI approaches to process development.

The findings directly address the research question by demonstrating that a hybrid human-AI approach significantly reduces the cost-to-target in semiconductor process development compared to human experts working alone. The success of the HF-CL strategy indicates that AI algorithms can excel in the fine-tuning stage of process development, capitalizing on the initial insights provided by human experts. The V-shaped cost-to-target curve illustrates the importance of optimal transfer points from human to AI, revealing a tradeoff between the benefit of human expertise and the cost of additional data. The contrasting approaches of human engineers and algorithms highlight the need to address cultural and behavioral differences to successfully integrate AI into semiconductor process development. The results suggest that the HF-CL strategy could be generalizable to other "little data" problems, drawing parallels with other AI applications like chess and protein folding.

This study demonstrates the potential of combining human expertise and AI algorithms to drastically reduce the cost and time required for semiconductor process development. The HF-CL strategy proves highly effective but requires careful selection of the transfer point from human to AI. Future research should investigate methods to optimize the transfer point and to encode domain knowledge into algorithms for better performance. Addressing cultural challenges in integrating AI into established engineering workflows is crucial for realizing the full potential of this collaborative approach.

The study was conducted in a virtual environment, which may not perfectly capture all the complexities and nuances of real-world semiconductor fabrication. The number of human participants and algorithms tested was relatively limited. The generalizability of the findings to other semiconductor processes or to different types of AI algorithms requires further investigation. The proprietary nature of the simulator used limits the transparency and reproducibility of the results to some extent.