Closed-Loop Autonomous Experimentation Discovers Cutting-Edge Phase Change Memory Material
In this work, we used closed-loop autonomous experimentation at the beamline to discover a new phase change memory material, Ge4Sb6Te7 (GST467) which was found to be significantly better than the well-known Ge2Sb2Te5 (GST225) in a side-by-side photonic device comparison in terms of optical contrast.
In this work, we used autonomous experimentation to discover a new phase change memory material. In particular, we developed a machine learning algorithm called CAMEO, which stands for closed-loop autonomous materials exploration and optimization. CAMEO allows rapid mapping of phase diagrams and utilization of the phase distribution information as a guide to quickly optimize physical properties. The material we discovered Ge4Sb6Te7 (GST467) was found to be significantly better than the well-known Ge2Sb2Te5 (GST225) in a side-by-side photonic device comparison.
As a central part of the autonomous experiment, CAMEO implements active learning to dynamically determine the sequence of synchrotron diffraction experiment across a combinatorial thin film library. Rather than measuring every single point on the library, the algorithm self-navigated the compositional space of the Ge-Sb-Te ternary system, and a clear picture of the phase distribution across the library was obtained with only a handful of measurements. A fundamental tenet of CAMEO is that materials compositions with maximum functional “responses” are often located at the structural phase boundaries. By incorporating analysis of the bandgap of the material in the closed loop, we were able to arrive at GST467.
We actually had the idea for CAMEO-based autonomous experimentation for rapid discovery of materials for a long time. We had earlier demonstrated on-the-fly analysis of synchrotron diffraction data for rapid phase mapping [https://doi.org/10.1038/srep06367]. Clearly, the next step was to incorporate active learning through Bayesian optimization, so that the algorithm controls the sequence of experiments. Compared to other reported demonstration of autonomous experiments, remotely and dynamically controlling the sequence of diffraction at a synchrotron endstation presented a major instrumentation challenge. This is where the longstanding collaboration with Apurva Mehta at SLAC really made a big difference. After all, we had been doing measurements of combinatorial libraries at SLAC for over ten years. When the idea for a self-driving run at a beamline emerged, we could envision different pathways to making it a reality.
The first time we actually got active learning driven experiment to work at a beamline was five years ago. For this original demonstration, we had used combinatorial libraries we had already characterized before. When it worked, we immediately got the sense that the strategy can actually go much further and that it can be unleashed on a real search-and-discover mode.
Many materials systems we have explored using the high-throughput approach had their functionalities derived from being at structural phase boundaries. A good example is morphotropic phase boundary piezoelectric materials. On the other hand, there are also many systems where having pure phases, away from phase boundaries is important to avoid poisoning the functionality. The birth of CAMEO, as embodied in a machine learning platform, was when the concept of rapid phase screening was combined with the use of the structural phase diagram as a blue print to home in on the path of steepest functionality ascent.
Because of the complexity associated with required characterization techniques, exploration of new phase change memory materials seemed like a good challenging topic to tackle with CAMEO. The target is a photonic switching device for neuromorphic computing which we have been pursuing together with Mo Li at the University of Washington. Thus, we focused on the bandgap difference between the amorphous and crystalline states as the key metric to maximize, so that the larger the bandgap difference, we would see the larger optical contrast between the states in a device. Actually, extracting bandgap from ellipsometry data is a tedious nontrivial task, and we decided to feed unprocessed raw ellisometry data as the prior for Bayesian active learning in CAMEO.
To our amazement, only after sampling 19 compositions, the algorithm found GST467, the “best” composition out of 180 spots within a large ternary region. This composition possesses bandgap difference almost 3 times that of GST225 which results in almost 50% more intermediate states which can be used for neuromorphic computing implementation. GST467 is indeed located at the boundary between the single-phase fcc Ge-Sb-Te region and the region consisting of Ge-Sb-Te and Sb-Te phases. What is particularly interesting is that electron microscopy of GST467 revealed coherent epitaxial nanocomposite of Ge-Sb-Te and Sb-Te phases, and despite this structure, photonic switching test carried out in Mo Li’s lab showed that the material is stable over at least 30,000 cycles.
We have been developing high-throughput characterization techniques for a variety of materials systems, but as is the case with phase change memory materials, some measurement tasks are just too difficult to implement in a true high-throughput manner. This is where CAMEO really shines, as it can be used to reduce the overall number of measurements by up to an order of magnitude. We are now actively applying CAMEO to combinatorial search of other types of materials.
For more details on the work, please see Nature Communications 11, Article number: 5966 (2020)
A. Gilad Kusne and Ichiro Takeuchi