Day 3: Applications

Dr. Masafumi Fukuto

Brookhaven National Laboratory

Autonomous X-ray Scattering Experiments at Synchrotron Beamlines

With the recent advances in bright x-ray sources, detectors, and automation, the data rates at light source facilities have been increasing rapidly. However, a key challenge facing these facilities is to maximize the value of each measurement while increasing the data throughput. Addressing this challenge is particularly important for studies of modern materials whose structure and properties depend on a large number of material composition and processing parameters. In order to efficiently probe the vast and complex material parameter spaces and promote the discovery and optimization of novel materials, intelligent experiment guidance methods are critically needed wherein the information content of each measured data point is maximized. In order to help tackle these challenges, a collaboration between LBNL CAMERA, BNL Center for Functional Nanomaterials, and National Synchrotron Light Source II has been developing autonomous experimentation capabilities, with a focus on applications to x-ray scattering experiments at synchrotron beamlines. The talk will summarize our recent progress with these efforts, including the successful integration of decision-making algorithms based on Gaussian processes into the experimental workflow to enable autonomous x-ray scattering experiments at beamlines. To illustrate the potential of the autonomous experimentation paradigm to expand the scope of beamline experiments, we will highlight examples of user-science experiments in which the implemented autonomous feedback loop was applied to the characterization of material heterogeneities in real space and mapping of material parameter spaces based on combinatorial sample libraries. The talk will end with discussions of our ongoing efforts and possible future directions.

Prof. Philip A. Romero

UW-Madison

Autonomous Systems to Rapidly Navigate Sequence-Function Landscapes for Protein Engineering

Proteins are capable of performing a wide variety of complex biological and chemical functions with exceptional proficiency, accuracy, and specificity. Protein engineering leverages this potential by creating new proteins with customized functions to address challenges in energy, medicine, agriculture, and industrial chemistry. Despite its broad utility and widespread success, the protein engineering process is laborious, time-consuming, and driven by ad hoc decisions that limit its overall efficiency. We are building systems that mimic the human protein engineering process, but can operate continuously with full autonomy. Our system consists of an intelligent agent that learns from data, generates hypotheses, and proposes new experiments; and a tightly integrated robotic platform that provides empirical feedback from the real world. Iteratively cycling between the agent and the robot allows for rapid navigation of the sequence-function landscape to discover optimized sequences. We demonstrate the capabilities of our autonomous system by engineering cellulase enzymes with enhanced tolerance to biomass pretreatment conditions. Our basic framework can be readily extended to new biochemical assays and detection modalities (e.g. LCMS), and systems beyond enzymes such as mammalian cells, metabolic pathways, and microbial communities. Autonomous systems’ continuous operation and AI-based decision-making enable them to engineer proteins faster and more efficiently than humans, and will push the boundaries of protein engineering.

Prof. Keith A. Brown

BU College of Engineering

Autonomous Experimentation for Structural Design and Additive Manufacturing

Nature has taught us that intricate structure from the molecular scale to the macroscale can lead to exquisite material properties. While inspiring, the extraordinarily vast number of permutations of composition, processing conditions, and structures spanning these scales makes brute force exploration intractable. Further, simulation cannot accurately and rapidly predict many important material properties such as non-linear mechanical properties such as toughness, indicating that experiments are necessary to explore this space. Autonomous experimentation (AE) presents a unique opportunity to address the complexity presented by hierarchical materials through their combination of automation to perform experiments rapidly and the use of machine learning to select experiments to achieve specific goals. In this presentation, we detail our work developing and utilizing AE systems for studying the mechanical properties of structures constructed using additive manufacturing. Specifically, we discuss our recently reported Bayesian experimental autonomous researcher (BEAR) that combines 3D printing and automated testing to realize structural materials that have highly tuned non-linear mechanical performance. While this approach was initially a black-box process, mechanics insight in the form of finite element analysis (FEA) contains critical insight that can in principle be useful, despite not completely capturing experimental performance. We continue to describe recent efforts to make the BEAR physics-informed through the incorporation of FEA. The interplay of high fidelity experimentation and comparatively low fidelity but high throughput simulation makes this an interesting case study for how best to efficiently use these disparate data streams for structural optimization.

Dr. Paolo Mutti

Institut Laue Langevin

Improving Measurement Strategy in Neutron Spectroscopy with Machine Learning

During the last reactor cycle in 2020, a combined team from the Institut Laue-Langevin and Berkeley National Lab has commissioned and tested a self-learning algorithm capable to perform autonomous measurements. For the first time, the computer took control of the three-axis neutron spectrometer ThALES, without any human intervention. The algorithm was able to explore the reciprocal space and fully reconstruct the signal without any prior knowledge of the physics case under study. Thanks to autonomous learning gpCAM --developed by Marcus Noack of the CAMERA team at Berkeley Lab -- estimates the posterior mean and covariance and uses them in a function optimization to calculate the optimal next measurement point. The posterior is based on a prior Gaussian probability density function, which is repeatedly retrained on previously measured points. The main advantage of such an approach is clearly the possibility to drastically reduce the number of measurements with respect to a classical grid scan and therefore optimize the beam-time usage. In the present paper, the excellent results obtained will be discussed as well as the opportunities for further improve this technique.

Dr. Aram Amassian

North Carolina State University

Accelerating Semiconductor Research through Robotic Automation and ML-Guided Experimentation

Solution-processing of semiconductors promises a new manufacturing paradigm by enabling ambient, high-throughput, and therefore low-cost fabrication of electronic and optoelectronic materials. These materials include a large area of chemically synthesizable conjugated organic molecules, colloidal quantum dot nanocrystals, and hybrid inorganic-organic metal halide perovskites, all of which are making tremendous progress in electronics, optoelectronics, and solar energy applications. However, the vast chemical, processing, and device design universe of these materials and technologies can be overwhelming to current research practices. The sheer scale of the design problem and the emergent nature of many phenomena can easily overwhelm the research infrastructure, benefitting from data-driven, and more specifically, from machine learning- (ML) guided research paradigms.

This presentation will introduce attendees to the complexity of emerging ink-based semiconductors and will discuss our recent efforts to leverage robotic automation and ML-guided experimentation towards accelerating parameter space exploration and discovery, as well as enabling efficient exploitation and autonomous experimentation. Examples will include the Artificial Chemist, an autonomous synthesis companion bot for photonic materials development, as well as the development of a robotic materials maker platform designed to accelerate materials research into solution-processed semiconductors.

Prof. Mykel J. Kochenderfer

Stanford University

Probabilistic Models and Algorithms for Discovery

The process of scientific discovery involves modeling the physical world, understanding and generalizing those models, and then using them to design and control future experiments. In this talk, we present probabilistic models and algorithms that can be used to automate parts of the scientific discovery process. Traditional machine learning approaches are generally data-driven but are often brittle to noise and extrapolation, while analytical approaches generalize well but can be intractable for complex phenomena. We study ways to combine data with structured probabilistic representations to improve data efficiency and generalizability. We demonstrate this idea by modeling autonomous driving, aircraft trajectories, and other dynamical systems. To arrive at more interpretable models, we developed a framework for optimizing modeling languages represented as a grammar. We demonstrate interpretable modeling for three tasks: equation discovery, time series categorization, and control. Lastly, we demonstrate how tools from decision-making can aid in the design of experiments and control of experimental instruments. We show how reinforcement learning can be applied to adaptive sampling for spectroscopy and how modeling with deep neural networks can be used to control a high-dimensional nonlinear fluid system.

Dr. Jason Hattrick-Simpers

NIST

How Robots Can Teach Us To Trust A.I.

Autonomous materials systems are becoming increasingly prevalent and promise to increase the rate at which research is performed, new materials are discovered, and long-held beliefs are challenged. But this only works if we trust that the AI/MLs directing the experiments. Here I will start by discussing the autonomous scanning electrochemical droplet cell system developed at NIST and its capabilities. We will then take a hard left and discuss the different varieties of ground truth, the importance of capturing scientific consensus and variance in data sets, and the importance of preserving the AI explainability even at the expense of traditional training scores.