Session F60: Autonomous Systems and Control

Live

The last few decades have seen significant advancements in materials research tools, allowing scientists to rapidly synthesis and characterize large numbers of samples - a major step toward high-throughput materials discovery. Autonomous research systems take the next step, placing synthesis and characterization under control of machine learning. For such systems, machine learning controls experiment design, execution, and analysis, thus accelerating knowledge capture while also reducing the burden on experts. Furthermore, physical knowledge can be built into the machine learning, reducing the expertise needed by users, with the promise of eventually democratizing science. In this talk we will discuss autonomous systems being developed at NIST with a particular focus on autonomous control of X-ray diffraction and neutron scattering for materials characterization, exploration and discovery.   

Live

At the world’s first nanocar race at CEMES-CNRS, in France, participants had to direct a nanocar across a “racetrack” [1]. In order to control their nanocar, they had to move it using the tip of a STM, albeit without making direct contact with the nanocar.

The physics that govern the molecule’s movement and rotation is complex and involves the interaction between the molecule and the tip as well as the molecule and the substrate [2]. Thus, it requires some expertise from humans to manoeuvre the nanocar and predict the outcome of a performed action.

Here, we show how an artificial intelligence (AI) based on reinforcement learning (RL) can be implemented to manipulate single molecules. The AI is implemented in the form of an off-policy RL algorithm, known as the Q-Learning. Being off-policy enables the AI to learn without the necessity of a physical model. In a prime example, the AI manoeuvres the nanocar with a success rate of 89%.

Our results can be the basis for more sophisticated techniques of molecular manipulations which allow identification and relocation of single molecules at will, building the basis for future bottom-up constructions of nanotechnology.


[1] G. Rapenne et al., Nature Rev. Mater. 2, 17040 (2017)
[2] G. J. Simpson et al., Nature Nanotech. 12, 604 (2017)   

Live

Recent advances in the areas of machine learning and reinforcement learning have led to
tremendous increases in accuracy and efficacy of ‘AI’ in fields such as computer vision, natural
language processing and gameplay. Here, we will focus on the applications and developments of
these methods as pertaining to physical sciences with several textbook cases. First, we show that
utilizing non-parametric models in conjunction with Bayesian optimization can substantially
increase the efficiency of data collection with scanning probe microscopy instruments. This is
illustrated for an example of finding areas of high electromechanical response in a thin
ferroelectric film. At the same time, improvements in machine learning and edge devices can
also be leveraged to accelerate the acquisition by performing parameter estimation in near real-
time, obviating the need for more expensive post-processing and enabling faster feedback for the
user. Finally, recent advances in the field of reinforcement learning can be leveraged to
dramatically alter the prevailing materials synthesis paradigm. A multi-agent reinforcement
learning algorithm was modified to work at scale and was utilized to tackle problems of
materials synthesis from the bottom-up in simulated environments of dopants in a graphene
lattice, as well as adjusting deposition parameters for a simulated kinetic Monte-Carlo
environment for thin-film deposition.   

Live

Material science problems are often information rich, but sparsely sampled. Here we present how material science knowledge can be encoded into active learning frameworks to efficiently navigate such search spaces. The physics-based constraints limit the solution space to only relevant solutions. We present the example of electron backscatter diffraction tomography, where each location in 3D space has orientation described by a 4D set of quaternions. As data is acquired, the goal is to: 1. Discover the structure of the data set with some measure of cluster membership probability, 2. Predict the grain map in unmeasured locations, and 3. Choose the best measurement to take next. This method explicitly handles uncertainty at each step, demonstrating the importance of uncertainty propagation for appropriate confidence in the predictions. By building physical knowledge and uncertainty propagation into these AI’s, this method takes advantage of the richness of the data, allows for more accurate, physically sound predictions with less data, and facilitates interpretability.   

Live

The first step to understanding the microscopic origins of the properties of a material is to determine the crystal structure. This can be accomplished with neutron diffraction. However, there are a small number of neutron sources in the world and thus it is critical to perform measurements as optimally as possible. We use reinforcement learning to address this problem. We compare several approaches within this framework including epsilon-greedy, Q-learning, and actor-critic. We find that in toy models, it is possible to measure a significantly smaller fraction of measurements than would commonly be performed to determine structural properties with the same accuracy.   

Live

High-fidelity ab-initio simulations of atomistic dynamics are limited to small systems and short times, and development of surrogate machine learning models for force fields is an emerging promising direction to access long-time large-scale dynamics of complex materials systems. However, the main challenges are high accuracy, reliability, and computational efficiency of these models, which critically depend on the training data sets. We develop ML interatomic potential models that are interpretable and uncertainty-aware, and orders of magnitude faster than reference quantum methods. Principled Bayesian uncertainty quantification built into these models enables the construction of autonomous data acquisition schemes using active learning. We demonstrate on-the-fly learning of machine learning force fields and use them to gain insights into previously inaccessible physical and chemical phenomena in ion conductors, catalytic surface reactions, 2D materials phase transformations, and shape memory alloys [1,2,3].
1. J. Vandermause, et al, NPJ Computational Materials, 6, 20 (2020)
2. J. S. Lim, et al, JACS. 2020, 142, 37, 15907–15916 (2020)
3. Y. Xie et al, https://arxiv.org/abs/2008.11796