The future of microelectronics will depend on more than making today's chips smaller. Researchers are searching for new materials and new ways to combine them - advances to support the faster computing, reduced energy use and secured domestic production that will guarantee continued U.S. technological leadership. But that search is slow and extraordinarily complex, requiring scientists to perform many experiments with multiple competing parameters and laboriously interpret the huge streams of resulting data.
At the Center for Nanophase Materials Sciences, a Department of Energy Office of Science user facility at Oak Ridge National Laboratory, researchers are meeting this challenge with a new generation of scientific technology: autonomous workflows for microelectronics research. By combining advanced instruments, artificial intelligence and theoretical know-how, they are creating systems that can do more than automate routine steps. CNMS systems are increasingly capable of interpreting incoming data and determining next steps in real time, helping researchers quickly navigate vast experimental spaces and achieve results that are beyond reach with conventional methods.
For Rama Vasudevan, who leads CNMS's data analytics efforts for autonomous science, the value of these systems goes well beyond speed. "It's not as simple as a performance enhancement from one-fold to ten-fold," he said. "It's really a qualitative difference in capability."
At CNMS, that qualitative shift is beginning on two fronts. First, autonomous systems are helping researchers more intelligently grow and refine materials for microelectronics, using data from each experiment to guide the next one. Second, they are helping scientists connect measurements, simulations and AI into a fuller picture of how those materials behave - and why they succeed or fail under real operating conditions.
Autonomous science in action
Autonomous systems are already changing how researchers create materials for microelectronics. Growing these materials means navigating a maze of interacting variables, from temperature and pressure to substrate conditions and laser settings. Each of those choices can affect the final material, and changing one often alters how the others interact. Instead of relying only on trial and error or a researcher's intuition to discover the best way to proceed, automated workflows can access a wealth of stored information alongside each experimental run to guide the process in real time.
Adaptive synthesis methods like these are especially valuable for creating the delicate nanomaterials used in next-generation microelectronics. In one ongoing CNMS project led by staff scientist Sumner Harris, researchers are working to grow oxide membranes that can be lifted off one surface and transferred onto another, giving scientists a way to stack and combine materials in new configurations. These configurations offer a lever to alter the electronic properties of the resulting composite, creating opportunities for dramatically improved memory and transistor performance.
That promise comes with a difficult materials challenge. Oxide membranes are often grown on a surface material such as graphene so they can be later separated and transferred to their destination surface, but there is a catch: the oxide material needs oxygen to form, but too much oxygen will degrade the graphene, making it impossible to transfer the target material. Autonomous workflows are especially useful for this challenge because AI systems can digest and apply a vast body of prior research and experimental information to complement human expertise as researchers search for the conditions that preserve both materials.
Vasudevan describes this project and others like it as a kind of human-AI conference, where researchers and autonomous systems each contribute where they are strongest. "Humans do really well in sparse-data regimes," he said. "But once a problem becomes too large and high-dimensional, AI can find patterns and simplify it in ways people can't."
Autonomous workflows are not just iterative loops of synthesis and measurement. By placing theory in the loop and leveraging ORNL's high-performance computing, these systems can generate and test new hypotheses based on discrepancies between simulation and experiment. This is particularly valuable for complex problems like materials synthesis, where intuition alone may be insufficient. "These autonomous workflows seamlessly integrate knowledge from literature, multi-modal experiments and simulations to accelerate materials discovery and build foundational knowledge," said Panchapakesan Ganesh, leader of the Theory and Computation section at CNMS.
Teaching instruments to choose
Autonomous experiments at CNMS are built step by step, not switched on all at once. Researchers first need to automate their tasks by teaching an instrument to independently and reliably perform a set of tasks, whether that means running a synthesis recipe or scanning a sample in a repeatable way. Then they begin adding AI: training models on data from prior runs, relevant literature and synthetic datasets so the system can recognize patterns, interpret what it is seeing and choose how to respond.
"Automated is necessary for autonomous," explained Marti Checa Nualart, an R&D staff scientist in the CNMS Functional Atomic Force Microscopy Group. "The switch happens when we empower automated instruments with properly trained models to make autonomous decisions, adapting as the experiment unfolds."
Autonomous workflows are changing more than operations of individual instruments. They are also helping researchers connect what they learn across many tools, datasets and simulations, adding up isolated measurements into a more complete picture of how microelectronic materials behave. In a project led by Argonne National Laboratory, Alphafold for Microelectronics, ORNL researchers are helping develop AI tools that combine experiments and simulations to better understand materials for memory devices and transistors, including investigating why some promising ferroelectric memory materials degrade with repeated use.
"The memory degradation problem is difficult because no single measurement can explain it," said Vasudevan. Researchers are working to integrate several disparate data sources from multiple facilities and experimental approaches, while simulations add still another layer of information. CNMS is using autonomous workflows to combine those data streams in a process called multimodal fusion - in which AI models are recruited to sort through incoming information and identify patterns that are nearly undetectable by human-level intelligence - guiding experimental approaches to zero in on the mechanisms behind memory fatigue.
Supporting users worldwide
CNMS is structured to support exactly that kind of work. As one of DOE's five Nanoscale Science Research Centers, it supports users from around the world. The constant flow of outside researchers and variable scientific challenges they present has helped push CNMS toward more streamlined and increasingly autonomous workflows, while Vasudevan's embedded Data NanoAnalytics Group helps connect those capabilities across synthesis, microscopy and theory.
For microelectronics, that integration could prove just as important as any single breakthrough material. The field is increasingly faced with problems that are too complex, too data-rich and too interconnected for traditional workflows alone. At CNMS, researchers are building systems that can learn from each step quickly enough to guide the next one - not only speeding discovery but also expanding what discovery can look like.
The DOE Office of Science and ORNL Laboratory Directed Research and Development program supported this research.
ORNL is supported by the DOE Office of Science. The Office of Science is the single largest supporter of basic research in the physical sciences in the United States and is working to address some of the most pressing challenges of our time. For more information, please visit science.energy.gov . - Galen Fader
