Semiconductors enable the modern world we live in by selectively controlling the conductivity of materials and ultimately the flow of information. As a whole, the need for getting the right information in the hands of the right people continuously grows year over year, despite recent challenges. During the Covid-19 pandemic in 2020, the global sales of semiconductors grew 5.4% and the market is forecasted to grow an additional 7.7% to a total of $476 billion in 2021. Currently, the United States excels in the design and innovation of new semiconductor technologies. The United States holds 47% (as of 2019) of the global sales market share but in order to keep that position spends around 40 billion a year in research and development. Alternatively, semiconductor supply chains tend to be more global to drive value and efficiency gains for the industry. With 80% of semiconductor foundries and assembly/test operations now concentrated in Asia, there is a clear strategic weakness for the United States and a need to consider investments in the semiconductor manufacturing sector.
Semiconductors are manufactured in a sequence of material patterning steps on a base substrate— usually a silicon wafer. The semiconductor industry uses silicon as a substrate due to its natural electrical insulating properties and its ability to absorb dopants that alter the properties of the silicon to their specific requirements. Many semiconductors must go through a 50 or more step manufacturing process in order to produce functioning devices. Transistors, contacts, etc., all made of different materials are laid down that, in concert, function as the core of all modern computing devices. The main steps of this manufacturing process are photo masking, etching (selectively removing material of the desired circuit patterns), ionic implantation (to introduce a dopant at a given depth into the material using a high energy electron beam), and metal deposition.
From an electronics consumer perspective the industry regularly comes out with new hardware that is astounding, however within the industry these advancements were planned and considered for decades. For example, Samsung just released a 512GB DDR5 RAM module which uses High-K Metal Gate (HKMG) technology. This technology has about twice the transfer speed and more than an order of magnitude more capacity than the current generation of DDDR 4 RAM modules. In this product, hafnium was used in the gates of the transistors of the device in order to achieve an extremely low leakage current. Ultimately, this means the HKMG DDR5 RAM has significantly reduced power consumption and more reliability. Although it is being marketed as a newly released product, the technology used in this module was demonstrated over 20 years ago and has been in use for more than a decade in some applications. Maturing a technology to be ready for mass production is challenging and time consuming. In the end, new technologies that improve semiconductor performance are only valuable when manufacturing is viable at scale.
So what makes semiconductor manufacturing difficult and what tools do semi fabs have at their disposal? As semiconductor manufacturing processes shrink to the micron level, sources of process variance are becoming increasingly harder to control. This lack of control can lead to significant decreases in process yield, the quantitative measure of the quality of a semiconductor process. In addition, the significant number of steps in order to manufacture a semiconductor can have compounding effects on yield. For example, if each manufacturing step in the semiconductor manufacturing process is ninety nine percent successful then after 50 similar manufacturing steps the yield will be below sixty percent. In this example, 2 in every 5 semiconductors would be discarded due to production issues. The industry employs various advanced process control methods to monitor process variance. For example, feedforward control algorithms may use metrology from the previous step to recommend control schemes for the next step of semiconductor manufacturing. These methods of process control are well understood, have been around for decades, and largely rely on physical simulations of the process in order to make control decisions. Unfortunately, physical simulations have their limitations. Each simulation is either an overly simplified representation of the real world process or they are extremely computationally expensive and cannot be run in real time during the manufacturing process.
As the semiconductor manufacturing industry looks forward to smaller critical dimension sizes and higher process yield, a larger number of process parameters need to be controlled with more precision. Rapid high-fidelity simulations at each step of the manufacturing process will be needed in order to optimize process control. The newly required quality and speed of simulations cannot be achieved with current advance process control technologies, therefore AI Process Control (AI-PC) must be utilized. AI-PC simulates semiconductor manufacturing processes hundreds of thousands times faster than conventional simulation methods enabling real time process control. AI-PC’s ability to model more possible states and apply the best process parameters in real time significantly improves process yield and opens up new semiconductor technologies to mass production.
Inevitably, without significant investment into AI-PC and the advancement of semiconductor manufacturing, all those dollars spent on the new semiconductor technologies will just be a waste of sand.
Alexander Fry is a Data Scientist & Machine Learning Engineer at Tignis Inc. He holds degrees in physics and astronomy including a PHD from the University of Washington. He currently serves as a key subject matter expert for Tignis in the Semiconductor Industry.
 “IDC forecast $476 billion in 2021, a 7.7% year-over-year growth rate.”