AI Driving Autonomous Research at PNNL for Discovery

Most of us have seen the chatbots that pop up on many websites, ready to answer our questions and provide around-the-clock, automated customer service. 

This illustration of how artificial intelligence can improve quality, efficiency and productivity is not limited to our daily business transactions. 

It also can be used to solve extremely complex scientific problems much more quickly.

Researchers at the Department of Energy’s Pacific Northwest National Laboratory are drawing on their expertise in data science to develop and apply AI innovations that will significantly reduce the time and cost of scientific discovery. 

By driving efforts to enable autonomous experimentation, they are helping to revolutionize research in biology, chemistry and materials science.

Simply put, these researchers are using elements of AI to assemble and combine resources and reasoning that make it faster and easier for scientists to do their jobs. They are developing approaches that orchestrate the trustworthy use of AI across the entire scientific process, from the earliest stages of ideation and exploration to developing and testing hypotheses in the laboratory.

Much of this work involves what are known as AI agents. While chatbots match keywords, work from scripts and can complete only simple tasks, AI agents integrate multiple inputs and outputs to complete a series of complex activities. 

Researchers at PNNL are building and testing a proof-in-concept agent-based platform that incorporates different commercially available AI models and several tools to accelerate innovation in an area of chemistry known as catalysis. 

Whereas some AI systems can generate what are known as hallucinations—incorrect or misleading results—the team’s robust approach to create specialized agents that can access specific databases and scientific literature is designed to ensure accurate and relevant information. 

The PNNL-developed tool allows researchers to ask questions and gather information about the molecular and physical properties of different chemicals. It not only coalesces that information, but it does so in the context of catalysis, shedding light on substances that act as catalysts to speed chemical reactions. 

Researchers can use this “co-scientist” to access visualizations of 3D molecular structures, run simulations and automatically generate computer code for data analysis. 

Other agents built into the system can interpret the analyses, explaining the steps and outcomes, as well as pointing out the importance of certain variables or results. The power of AI does not stop there. Via the same interface, users can ask the AI agent to generate new hypotheses along with the associated rationale and challenges. Scientists can then refine and evolve their own hypotheses.

When ready, scientists can proceed to the next step, asking for a plan to test a given hypothesis. The tool determines the ingredients needed to synthesize the catalysts, and it also provides step-by-step instructions for conducting an experiment.

To close the loop between ideation and experimentation, scientists can use the AI agents to translate the validation protocols into instructions that direct the robotic equipment in the laboratory. Of course, a human would still manage the experiment, ensure safety and learn from results. When all is said and done, this novel co-scientist approach might make the research endeavor 100 times faster.

Researchers at PNNL are exploring how their platform could help perform end-to-end research in the real world, starting with specific tools and data sources, along with instruments on our campus. 

Researchers in different fields also are working to develop the hardware, software and instrumental knowledge required to expand this platform for other disciplines. 

A wise person might say, “I don’t need to know everything; I just need to know where to find it when I need it.” 

At PNNL, researchers are expanding on this concept by advancing the use of AI. In so doing, they are paving the way for scientists to more quickly and easily explore and analyze data, automate reasoning and facilitate experiments—ultimately accelerating scientific discovery.

Steven Ashby, director of Pacific Northwest National Laboratory, writes this column monthly. To read previous Director's Columns, please visit our Director's Column Archive.