News: Suppliers
12 July 2021
Argonne uses AI to optimize ALD in real time
Manufacturers increasingly rely on atomic layer deposition (ALD) to make new types of films, but figuring out how to tweak the process for each new material takes time. Part of the problem is that researchers primarily use trial and error to identify optimal growth conditions. But a recent study suggests that using artificial intelligence (AI) can be more efficient (N Paulson, ‘Intelligent Agents for the Optimization of Atomic Layer Deposition’, ACS Applied Materials & Interfaces vol.13 (2021) issue 14, 17022).
Researchers at the US Department of Energy’s (DOE) Argonne National Laboratory describe multiple AI-based approaches for optimizing the ALD processes autonomously, detailing the relative strengths and weaknesses of each approach, as well as insights that can be used to develop new processes more efficiently and economically.
“All of these algorithms provide a much faster way of converging to optimum combinations because you’re not spending time putting a sample in the reactor, taking it out, doing measurements etc, as you typically would today. Instead you have a real-time loop that connects with the reactor,” says principal materials scientist Angel Yanguas-Gil, co-author.
Cutting edge, but with challenges
ALD excels at growing precise, nanoscale films on complex, 3D surfaces such as the deep and narrow trenches patterned into silicon wafers. This has motivated scientists worldwide to develop new thin-film ALD materials for future generations of semiconductor devices.
However, developing and optimizing these new ALD processes is challenging and labor-intensive. Researchers have to consider many different factors that can alter the process, including:
- the complex chemistries between the molecular precursors;
- reactor design, temperature and pressure; and
- the timing for each dose of their precursors.
To find ways of overcoming these challenges, Argonne scientists evaluated three optimization strategies — random, expert system and Bayesian optimization — the latter two utilizing different AI approaches.
Set it and forget it
Researchers evaluated their three strategies by comparing how they optimized the dosage and purge times of the two precursors used in ALD.
The goal: find the conditions that would achieve high and stable film growth in the shortest time. Scientists also judged the strategies on how quickly they converged on the ideal set of timings using simulations that represented the ALD process inside a reactor.
Linking their optimization approaches to their simulated system let them measure film growth in real time after each cycle, based on the processing conditions that their optimization algorithms generated.
“All of these algorithms provide a much faster way of converging to optimum combinations because you’re not spending time putting a sample in the reactor, taking it out, doing measurements etc, as you would typically. Instead you have a real-time loop that connects with the reactor,” says principal materials scientist Angel Yanguas-Gil, another co-author.
This set up also made the process automatic for the two AI approaches by forming a closed-loop system.
“In a closed-loop system, the simulation performs an experiment, gets the results, and feeds it to the AI tool. The AI tool then learns from it or interprets it in some way, and then suggests the next experiment,” says computational scientist Noah Paulson (lead author).
Despite some weaknesses, the AI approaches effectively determined the optimal dose and purge timings for different simulated ALD processes. It is reckoned that this makes the study among the first to show that thin-film optimization in real time is possible using AI.
“It opens up the possibility of using these types of approaches to rapidly optimize real ALD processes, a step that could potentially save manufacturers precious time and money when developing new applications in the future,” concludes senior chemist and co-author Jeff Elam.
The scientists used Argonne’s Blues cluster in its Laboratory Computing Resource Center. This research was funded by the Laboratory Directed Research and Development (LDRD) program at Argonne.
