Session G51: Autonomous Science

A Cloud-based AI-driven Autonomous Lab

Creating and designing new molecules is one of chemistry's most significant outcomes. The application of domain knowledge gained over decades of laboratory experience has been critical in the synthesis of numerous new molecular structures. Nonetheless, most synthetic success stories are preceded by lengthy hours of unproductive repetitive experiments. Automation systems, which were developed less than two decades ago to assist chemists with repetitive laboratory operations, have helped to reduce this problem. While these technologies have demonstrated exceptional efficacy in a select fields, such as high-throughput chemistry, automating general-purpose jobs remains an extremely complex issue even today. It requires chemical operators to develop unique software for various operations, each of which codifies a distinct sort of chemistry.

Natural language processing models have emerged as one of the most effective, scalable approaches for capturing human knowledge and modelling chemical processes in organic chemistry. Its use in machine learning tasks demonstrated high quality and ease of use in problems such as predicting chemical reactions [1-2], retrosynthetic routes [3], digitizing chemical literature [4], predicting detailed experimental procedures [5], designing new fingerprints [6] and yield predictions [7]. In this talk, I'll talk about the impact of language models in chemistry by highlighting the critical role of NLP architectures in implementing the first cloud-based AI-driven autonomous laboratory [8].

The remote laboratory is made accessible to chemists through the cloud [6] and is equipped with automation technologies. The AI assists remote chemists with several tasks: designing retrosynthetic trees and suggesting the correct sequence of operational actions (reaction conditions and procedures), or ingesting literature on synthetic procedures to convert them into an executable program. Following supervision by synthetic chemists, the AI self-programs the automation layer and makes decisions on the synthesis execution using feedback loops from analytical chemistry instruments. I will present the platform architecture and its performance across various classes of synthetic tasks.   

Magnetic control of tokamak plasmas through deep reinforcement learning

Nuclear fusion using magnetic confinement, in particular in the tokamak configuration, is a promising path towards sustainable energy. A core challenge is to shape and maintain a high-temperature plasma within the tokamak vessel. This requires high-dimensional, high-frequency, closed-loop control using magnetic actuator coils, further complicated by the diverse requirements across a wide range of plasma configurations. Traditionally, shape and position control have been handled by linear controllers designed using control engineering methods. In this work [1], we demonstrate the first successful application of Deep Reinforcement Learning (RL) to the magnetic control of tokamaks. The Deep RL algorithm learns magnetic controllers for a target TCV plasma equilibrium solely by interacting with a free-boundary evolution model, which simulates the plasma equilibrium evolution coupled to the circuit equations for external conductors. These controllers generate and maintain a plasma of the desired shape by receiving only magnetic measurements, without the need for equilibrium reconstruction, and actuating directly the coil power supply voltages in real-time. A variety of controllers have been implemented and successfully tested in TCV experiments, achieving several different target shapes including elongated plasmas, negative triangularity and snowflake configurations. This control design method has the potential to reduce the effort required to obtain a new target shape in a tokamak. Moreover, this work represents the first use of reinforcement learning for feedback control on a tokamak and paves the way for combining physics models and machine learning for improving other aspects of control of fusion plasmas.

 

[1] J. Degrave, F. Felici, J. Buchli et al. Magnetic control of tokamak plasmas through deep reinforcement learning. Nature 602, 414–419 (2022). https://doi.org/10.1038/s41586-021-04301-9   

Autonomous Materials Science

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 I will discuss autonomous systems being developed between NIST and collaborators. Examples include the first autonomous discovery of a best-in-class solid-state material, an autonomous system that merges live experiments and live computation, and a scalable operating system for the autonomous laboratory   

Rheostats and toggles switches for modulating protein function: A Cautionary Tale

Proteins are heteropolymers – each assembled from a unique, linear sequence of amino acids – that fold into 3D shapes to perform their functions. Changes in protein sequences, such as the SARS-CoV2 changes frequently in headlines, can alter protein function. Some changes are biologically relevant, whereas other changes are neutral. To predict the outcomes of these changes, decades of biological, biochemical, and biophysical mutation experiments – along with bioinformatic studies of evolutionarily-related protein sequences – have been used to derive several common assumptions that underlie most computer prediction algorithms. However, most historical studies were biased to positions that are conserved during evolution. Most mutations at conserved positions are catastrophic (they “toggle off” structure/function), which is reliably predicted by most computer algorithms. In contrast, mutations at evolutionarily-changing positions (nonconserved) were largely overlooked, despite their critical roles in evolving functional variation. As a consequence, computer predictions about amino acid changes at functionally-important, nonconserved positions are poor. Our studies demonstrated one source of this failure: Amino acid substitutions at a special class of nonconserved protein positions do not follow the same substitution rules as conserved positions. These special “rheostat” positions are present in a wide range of protein types; in some proteins, rheostat positions comprise >40% of the protein positions. In crystallo and in silico structural studies of functional rheostat substitutions showed only local perturbations to the protein 3D structure. Dynamic coupling calculations for rheostat substitutions showed promising correlations with measured functional changes. Combined, results suggest that emergent properties of coupled amino acid networks could produce the complex outcomes observed for rheostat substitutions that must be understood to advance predictions.