Development of Standards to Support a Modular and Autonomous Laboratory Ecosystem

Summary

Autonomous experimentation (AE), or self-driving laboratories, combines AI and automation while leveraging human intuition and creativity to guide experiment campaigns. While the underlying technology for AI and automation exists in many domains, their application to materials research is still in its infancy. Currently, there is no standardized AE ecosystem for materials R&D, which is a major obstacle to industrial adoption. We are working to develop standards for the following aspects of modular labs: (1) sample management standards, (2) instrument control and communication standards, (3) data and knowledge management standards, and (4) algorithm and model integration standards.

Description

Credit: Zachary Trautt

Autonomous experimentation (AE), or self-driving laboratories, combines AI and automation while leveraging human intuition and creativity to guide experiment campaigns. AE can reduce the time and resources needed to discover new materials by orders of magnitude and reduce the time-to-market for critical technologies based on them. Further, AE has the power to address problems that are otherwise intractably complex or ever-mutating.

While the underlying technology for AI and automation exists in many domains, their application to materials research is still in its infancy. Currently, there is no standardized AE ecosystem for materials R&D, which is a major obstacle to industrial adoption. Creating a standards based ecosystem will dramatically reduce the cost of engineering a platform, as well as reduce the risk of obsolescence by ensuring it is expandable and upgradeable. Relatedly, a defined standard will empower equipment and software vendors to design their products for autonomous integration, reducing the cost and liability associated with bespoke engineering and ensuring a customer base for their products.

Just as the internet revolution was the outcome of low-level communication standards, we aim to initiate a laboratory revolution that will be powered by a standards-based modular laboratory ecosystem. We are working to develop standards for the following aspects of modular labs:

Sample Management Standards

Compared to the many existing AE platforms in the chemistry and bio-pharmaceutical fields, a particular challenge in automated systems for materials science is the need to move around solid samples. Solid materials almost entirely fall into three categories: thin films on substrates, bulk samples, and powders. Each of these has unique sample-handling challenges and requirements. The challenge here would be to create a series of universal sample holders that could handle one or more of these sample form factors. Consider the USB-C standard; it defines a variety of aspects (e.g., physical form factor, data transfer protocols, and rate, power transfer rate, alternate use modes) that can be supported (or not supported) at a variety of levels. Similarly, the sample holder interchange standard can describe a variety of aspects, such as sample form factor, size, number of samples, temperature, and sample atmosphere. Each instrument and various sample holders can be designed to handle each of these aspects to a different degree. For example, one instrument may be designed to only support thin films while another can support thin films, bulk samples, and powders. Another example is that some sample holders may be designed to be heated and others can only be used at room temperature. Depending on the nature of a particular instrument, the sample holder may be used natively in that instrument (x-ray diffractometer being a likely instance), while in other cases the sample holder will not be conducive to the tool operation (a likely instance being a device to test a material as part of a device like a battery) and the sample will need to be manipulated by the instrument.

Instrument Control and Communication Standards

Digital connectivity between AI infrastructure, data infrastructure, and physical laboratory infrastructure is critical for the operation of all autonomous systems. For most of the past century, scientific instrumentation has been designed around human operators. The transition from operation via human intelligence to artificial intelligence has been mired by fragile hacks rather than robust interfaces. Efforts to address these issues have been made in Internet of Things (IoT) communication protocols (e.g., MQTT) and within some sectors of scientific instrumentation (e.g., SiLA, EPICS) as it pertains to controlling and monitoring instruments. This new body will evaluate existing protocols and identify paths forward that address the unique challenges faced by experimental materials hardware and related connected systems.

Data and Knowledge Management Standards

In autonomous systems machine-actionable and AI-Ready data are required to inform advanced algorithms and models. Significant progress has been made in elevating the FAIRness of computational data through the establishment of OPTIMADE and related community efforts. Unfortunately, experimental equivalents have lagged behind. This new body will place an increased emphasis on building consensus on the data interchange formats including priority instrument types and knowledge graphs.

Algorithm and Model Integration Standards

Currently, there is a lack of portability that would enable an algorithm to engage multiple autonomous systems. Some of this challenge is addressed with instrument communication standards, but an additional layer of abstraction may provide significant benefits for future portability. The power of uniform high-level abstractions has been clearly demonstrated by scientific modeling software that enables generalized scientific problem statements to be addressed by diverse underlying computational codes. One example is the Atomic Simulation Environment, which provides a generic language for describing problem statements that can be seamlessly solved with a variety of different atomistic codes and methodologies. This new body will develop a comparable high-level interface for experimental materials science, such as an autonomous experimentation environment, which builds upon precursors such as BlueSky, ChemOS, Hermes, HELAO, and others. Once an open-source environment is established, outside vendors may produce their own proprietary innovation connected via the open standards.

Furthermore, such an autonomous experimentation environment allows for a more rigorous comparison of AI/ML algorithms for scientific tasks, which, in turn, fosters innovation. Some of the most impactful developments in generalist AI/ML have come from open-source libraries, such as PyTorch, Keras, TensorFlow, scikit-learn, and many others. Despite their success, they are not specialized in the specific needs of materials and manufacturing, nor are they informed by known physics. Unlike the unstructured data from social networks that the generalist AI/ML tools were developed to handle, scientific data is highly contextualized and imbued with specific physical meaning. Because of this, generalist AI/ML will fail to make meaningful insights on many scientific or engineering problems. Physics-aware scientific AI/ML is a nascent research topic that seeks to leverage the established physical knowledge and structure of scientific data. However, the nascent scientific AI field lacks any standardization in terms of how it is implemented, documented, or evaluated. This effort will develop an ecosystem of open-source software packages for scientific AI. These packages will be empowered with known physics and have the capability to discover new physics. These packages, together with the autonomous experimentation environment, will provide broad benefits as well as stimulate the development of proprietary AI within US industries.

Major Accomplishments

We are currently developing precursor standards within our robotic platform for autonomous synthesis of metal-organic frameworks.