Accelerated Materials Discovery

How to Use Artificial Intelligence to Speed Up Development

Phil De Luna

• Publisher: De Gruyter
• Year: 2022
• Print ISBN: 9783110738049
• eBook ISBN: 9783110733259
• Edition: 1
• Topic: Computer Science
• Subtopic: Chemistry
• Index: Chemistry

Typical timelines to go from discovery to impact in the advanced materials sector are between 10 to 30 years. Advances in robotics and artificial intelligence are poised to accelerate the discovery and development of new materials dramatically. This book is a primer for any materials scientist looking to future-proof their careers and get ahead of the disruption that artificial intelligence and robotic automation is just starting to unleash. It is meant to be an overview of how we can use these disruptive technologies to augment and supercharge our abilities to discover new materials that will solve world's biggest challenges.

• Written by world leading experts on accelerated materials discovery from academia (UC Berkeley, Caltech, UBC, Cornell, etc.), industry (Toyota Research Institute, Citrine Informatics) and national labs (National Research Council of Canada, Lawrence Berkeley National Labs).

• 
  
  

1 An overview of accelerated materials discovery

Robert Black

Energy, Mining, and Environment Research Centre, National Research Council of Canada, Canada

Isaac Tamblyn

Security and Disruptive Technologies Research Centre, National Research Council of Canada

Disclaimer: © Her Majesty the Queen in Right of Canada, as represented by the National Research Council of Canada, 2021.

1.1 Accelerated materials discovery and the promise of autonomous material discovery

New materials are core to forward progress in technology. We refer to historical periods based on the prevalent working material of the time: the Stone age, the Iron Age, the Bronze Age, and so on. Major technological leaps are often closely related to new approaches to material synthesis, processing, or creation. Record setting vehicles often operate at the extreme of material limits (e.g. titanium in the world’s fastest manned aircraft (SR-71), carbon fiber and Kevlar in high performance racing sails, pristine and ultra-pure materials for semiconductors for solar cells in spacecraft). As the rate of technological improvement increases, so does the demand for research to find new and better materials and material processing capabilities. Human-centric research is becoming increasingly expensive to perform, and scientific breakthroughs are becoming less efficient in terms of researcher productivity and costs compared to past innovations [1]. Factors pushing research in more expensive directions (both monetary and time requirements) are numerous: the magnitude of data being generated is increasing substantially year after year, expenses associated with sophisticated and specialized equipment to understand more complex and ground-breaking ideas becomes increasingly more expensive, and the resources required to know and understand all the information available is reaching a critical point. This increased complexity has resulted in timeframes of 20+ years for a product to go from conception to market [2]. Naturally, with the need for further investments to push R&D efforts forward, it is of no surprise that investors shy away from “deep technology” for investment return, that which focuses on material and/or hardware production, and instead is moving towards a more digital space [3].

1.1.1 Accelerating experimental workflow

To accelerate research efforts, traditional human-centric experimentation has a significant opportunity to undergo a paradigm shift. In the world of materials, this can mean designing, developing, and deploying automated experimental platforms (laboratories) that perform synthesis and characterization, with the goal of using artificial intelligence and machine learning (AI/ML) to optimize a specific material property, either through a more efficient “exploratory” method, or through capabilities to explore new design space that was previously not considered. This shift will come through the application and integration of one or multiple of the following areas into the experimental workflow:

1. Automated experimentation – Robotic manipulation of benchtop experimentation, data handling and operation of equipment, integration of various experimental pieces into a unified platform capable of serial or high-throughput operation
2. Artificial intelligence/machine learning – Utilization of data to train statistics-based models to determine the optimal experimental space towards a specific material function, including concepts of inverse material design
3. Data – A unified protocol for the collection and storage of data. This includes data acquisition from the measurements, post-processing and curation, storage of relevant information into various databases, and assurance of data accuracy and integrity
4. Computer simulation – Augmenting the experimentation and AI with first principles or simulation-driven aspects of the material of interest

The combination of these concepts enables accelerated material discovery, with the exact degree of integration of each concept dependent on the desired goals and the level of autonomy required of the experimental platform. We take an analogy to that of the Society of Automotive Engineers (SAE) International definition of autonomy for self-driving cars [4] where different levels of autonomy are defined based on the integration and advancement of the components that comprise an autonomous researcher (Fig. 1.1).

Fig. 1.1: Five levels of an accelerated experimental workflow, with each successive level integrating more components to eventually reach full autonomy. The red dot marks the threshold for crossover from experimental automation to experimental autonomy.

• Level 0 – base level. A typical human-centric experimental process and/or current-day laboratory. The majority of experimentation is carried out by humans. Some pieces of equipment are used and provide signals in digital formats, but there is essentially no exchange of information other than what is measured and processed by a human. As an example, a device might collect data and provide some peak positions, but the output format is usually an image which ends up in a scientific publication, with no outside data transfer.
• Level 1 – human assistance. This sees the implementation of some type of assistance to the human tasks, whether physical or digital. Assistance is minimal and confined to only a single task, such as the use of pipetting machine and/or some form of automated data collection. At this stage, one can expect the acceleration factor of experimentation to increase and the experimentation protocol to be more efficient, but still the workflow is entirely dictated by the decisions of the human.
• Level 2 – partial automation. At this stage, the experimental workflow is becoming much more automated, with various experimental procedures capable of operating over extended periods of time. This could be considered a “high-throughput” stage, as the workflow pace greatly increases. While the workflow is enhanced through the incorporation of either physical or digital means, the human is still in control of the experiments being run, with their own analysis and decisions controlling the workflow of the platform.
• Level 3 – conditional automation. This stage sees the application of AI/ML to create optimization loops which are coupled to the platform automation. Through the incorporation of statistical and/or deep learning algorithms into the workflow, the system will be capable of making its own decisions based on the empirical data that is produced during a single campaign. This can be considered the base level for a self-driving laboratory. While direct human involvement in the process is reduced compared to previous levels, domain knowledge is still key to successfully drive the experiments effectively. This includes the human defining the experimental features, and scientific intuition to develop the proper workflow is still necessary.
• Level 4 – partial autonomy. At this stage, the system starts to truly become autonomous, as it begins to infer chemical properties and relations from outside sources (literature, past experiments, etc.) to truly plan and execute autonomous experimentation. Coupling of theory and empirical data allows the system to understand and infer chemical processes, using this knowledge to more efficiently search the material landscape. The human role in this platform is limited.
• Level 5 – full autonomy. This indicates a truly autonomous researcher where the human can effectively be removed from the picture. The platform is able to complete all tasks of discovery and reporting. The platform determines what experiments, simulation, and theoretical approaches need to be applied to optimize the active learning algorithm in a truly closed-loop design process. Furthermore, this stage represents a truly autonomous researcher, the platform is also capable of reading the literature to update its own knowledge, while also reporting findings in a human readable format (natural language processing).

As technology advances and we can scale the “autonomous material discovery” ladder, we become closer to realizing the concept of a truly autonomous researcher. This is the required paradigm shift to greatly accelerate the pace at which deep-technology experimentation is performed, while promising significant cost reduction and time savings for the material development process [5]. This paradigm shift has been recognized by others, in particular the role that big data and data-driven processes will move research into a new paradigm [6].

In what follows, we will discuss the role laboratory automation, machine learning, and data infrastructure will play in reaching truly autonomous experimentation, our view of where the field is moving in the near and the medium term, and where the opportunities are for new advancements in this area. The field of accelerated material discovery is currently only at its infancy and is an exciting advance in the history of scientific discovery – the beginning of a new era for material discovery, design, and control.

1.2 Laboratory automation to enable machine learning and artificial intelligence

Our discussion will begin with laboratory automation, as this is the key enabler of data acquisition and is the heart of any self-driving laboratory for accelerated material discovery. Laboratory automation refers to the coupling of physical experimentation via robotic control with software to perform repeated tasks. Much akin to a world-class researcher working in the lab, automated experimentation is key in ensuring a consistent, non-variant workflow that is theoretically able to run for days on end, collecting useful and meaningful data throughout the entire experimental campaign. Automated experimentation, has changed and become more widely adopted over the years as new concepts and tools are developed to advance research capabilities beyond traditional human-centric approaches. Regardless of all advances, the purpose has remained the same – decrease...

Table of contents

• Title Page
• Copyright
• Contents
• 1 An overview of accelerated materials discovery
• 2 Artificial intelligence for catalysis
• 3 Artificial intelligence for materials spectroscopy
• 4 Flexible automation for self-driving laboratories
• 5 Algorithms for materials discovery
• 6 Industrial materials informatics
• About the editor
• Index