Plasma etching, extensively used in semiconductor manufacturing for pattern transfer, suffers from macroscopic non-uniformity (at the wafer level) and batch to batch shifts due to reactor wall changes necessitating frequent seasoning; both mishits increase the production cost. A dominant demand is also “greener” recipes, i.e., processes with lower global warming potential (GWP), non-toxic gases, and efficient energy use. Additionally, the quest for chemistries and the design of recipes suitable for plasma etching of new materials entering the stack is a difficult task and is usually based on minimal literature data and/or a trial and error procedure. Finally, although the development of new recipes may be performed in reasonable time scales, introducing a process in the manufacturing pipeline and assuring that it is stable for processing thousands of wafers requires longer time scales.

We envision a universal solution with the potential to tackle all modern challenges of plasma etching process: A computationally fast and accurate data driven model of the process will be developed and, through optimization algorithms, will be utilized to provide answers to important questions regarding the cost, the uniformity, the drifts, the environmental footprint, and the design of new recipes. The development of the model will be based on data computed from first-principles models (physics-based models) and collected from apt experimental measurements; these data can be used to train a used to train a machine learning system to approximate the outputs of the physical process. Artificial intelligence models (AI-models) are essential as physics-based modeling, although theoretically feasible to deal with all the above-mentioned challenges, entails a high computational cost (e.g., for the numerical solution of partial differential equations describing conservation laws in reactor scale models). Although using machine learning to train a model is also computationally expensive, once trained, such models are orders of magnitude faster than physics-based models. This computational efficiency allows the quick exploration of the complete parameter space, which is critical when performing optimization studies to address the plasma etching challenges. The case study will be silicon dioxide etching with fluorine chemistry in an inductively coupled plasma (ICP) reactor.

The critical questions to be answered by plasmAI, through the coupling of experimental measurements, physics-based modeling, AI-models, and optimization algorithms, are summarized below:

We envisage the proposed approach as a paradigm, not only for plasma etching processes with different chemistries and substrate materials but also for other nanoelectronics processes such as chemical vapor deposition processes.

The research team of plasmAI, bringing together researchers from two Universities (National Technical University of Athens and University of Patras) and two Research Institutes (Institute of Nanoscience & Nanotechnology and Institute of Informatics and Telecommunications of NCSR Demokritos), is not only highly motivated in the proposed approach and invested in the outcome of plasmAI, but also carries a high level of expertise in complementary fields, machine learning and AI methods, plasma processing, physics-based modeling, and optimization, making the success of plasmAI almost inevitable.