A faster way to screen supply chains for harmful chemicals
IBM’s new Safer Materials Advisor can speed up the review of thousands of chemical products for PFAS by more than four times while reducing the chance of errors.
More than 4,500 chemical products are cleared for use within IBM Research headquarters in Yorktown, N.Y., and documentation for just about each one has crossed Angela Hutchinson’s desk at some point in the last two decades.
As the Yorktown chemical coordinator, it’s Hutchinson’s job to review all chemical requests tied to research happening at the lab, from semiconductors to quantum computing. As IBM’s business has evolved, so have the requests.
The one thing that hasn’t changed is the tedious nature of reviewing each chemical product’s “safety data sheet” (or SDS) for banned or restricted ingredients.
“I have to read through each of these SDSs manually and then cross-check each chemical on several spreadsheets,” Hutchinson said, glancing at a pile of papers on her desk marked up with highlighter and handwritten comments.
Hutchinson runs through each ingredient in each chemical product requested by researchers to make sure none are restricted or banned internally, or are subject to local, state, or federal regulations. If a chemical ingredient has restrictions, she may approve the product under certain conditions or search for available substitutes. It’s a trying and time-consuming process, even for straightforward requests.
When it comes to PFAS chemicals, nothing is straightforward. Per- and polyfluoroalkyl substances (PFAS), come in thousands of varieties and are covered by a patchwork of still-evolving regulations. The U.S. Environmental Protection Agency (EPA), among others, has enacted rules to limit or ban the manufacturing and use of many kinds of PFAS.
IBM, in collaboration with the Semiconductor Industry Association and other trade groups, has been working to steadily identify and eliminate PFAS from its operations. But in some instances, viable alternatives have yet to be found for critical processes.
To speed the transition away from PFAS, IBM researchers recently developed the Safer Materials Advisor, a tool that uses AI to screen products for PFAS and suggest alternatives.
When the opportunity arose to try out the tool as its first potential user, Hutchinson jumped on it. Over the last several weeks, the team that built the tool has tracked her progress, estimating that it improved her productivity by about 430%.
On its own, the tool can’t guarantee that a product is PFAS-free. But it can give employees in Hutchinson’s shoes greater confidence that each SDS review was done thoroughly. “With so many lists and spreadsheets to check, you worry about missing something,” she said.
The Safer Materials Advisor features a parser for converting product documentation to a machine-readable format, and several AI models designed to extract insights from the data.
One of the tool’s main jobs is to identify mystery or missing ingredients. Manufacturers sometimes provide only vague descriptions of ingredients in their products to avoid revealing any proprietary chemicals. Sometimes, also, the volume of ingredients also fails to add up to 100%. These kinds of omissions add to the challenge of screening for PFAS and other chemicals of concern.
When the ingredients on an SDS don’t add up, the advisor sends the full list to an AI agent that uses several methods to estimate the odds that PFAS is present. The tool has also been designed to recognize them by their molecular name, chemical registry number, or molecular structure.
If the tool flags a chemical as PFAS, it calls a second AI model to run a regulatory analysis to check which rules may apply. Reporting requirements vary based on where the PFAS are made or used as well as their various applications. The tool checks if the PFAS variant appears on any internal corporate list of banned or restricted chemicals or would be subject to local, state, or national regulations.
Finally, the tool can call a third AI model to assess the environmental and health risks associated with a given chemical, factoring in its toxicity, and how long it sticks around in the environment. The hazard analysis, powered by IBM’s chemistry foundation model, cites its sources for quick verification, and tells users how confident it is in its rating.
The analysis is meant to help employees in Hutchinson’s position find alternatives to not only PFAS but other chemicals of concern. The Safer Materials Advisor can also be adapted to new chemicals of concern as needed.
Introduced commercially in the 1940s, PFAS substances have unique properties that have led to their widespread use in everything from nonstick cookware and stain-resistant fabric to semiconductor manufacturing.
PFAS today exists in thousands, if not millions, of varieties that share some combination of carbon and fluorine. The carbon-fluorine bond is one of the strongest known to chemistry and it gives PFAS molecules their ability to repel water and grease, and withstand harsh chemicals, radiation, withering heat, and high pressure.
It's also why PFAS are so good at sticking around. They’ve been detected in the blood of most Americans and in the rainwater of most places on Earth at levels above the EPA’s safe drinking water standard.
Amid rising concerns about PFAS’s health and environmental impacts, governments worldwide have moved to regulate their use. Today, some forms of PFAS are limited or banned in the U.S., European Union, and Japan.
The EPA last year set a limit on how much PFAS could be present in drinking water but recently gave water utilities additional time to comply with the standard. In the interim, several states have established their own PFAS reporting rules. The EPA has also required companies to retroactively report how much PFAS they’ve used since 2011, creating more paperwork for employees like Hutchinson that lead environmental health and safety reviews.
IBM Research has, for several years now, been applying AI and advanced computing to make computing more sustainable, in part, through the design of safer materials. The focus on PFAS accelerated with the launch of “PFACTS,” an ongoing project funded by the National Science Foundation and led by IBM ChemForward, Cornell University, Digital Science, Numat, and the University of Pittsburgh. The PFACTS team recently released pfasID, an open-source screening tool hosted by ChemForward.
IBM researchers built on the technologies underlying pfasID to create the Safer Materials Advisor, to help IBM and its clients move faster to identify and address potentially harmful chemicals. The tool can work on piles of different kinds of documents, and it can be extended to other chemicals of concern. Enterprise users can also add new PFAS definitions and regulations as they evolve.
Other chemical coordinators at IBM have started using the Safer Materials Advisor at Hutchinson’s recommendation. The tool is generalizable and can be extended to help find safer alternatives for not only PFAS but other regulated chemicals.
Hutchinson provides detailed reporting on all chemicals of concern used at the Yorktown lab, both internally, to IBM’s Chief Sustainability Office, and externally, to several federal agencies. In addition, she meets once a year with town officials in Yorktown, New York, to go over the amount of chemicals used on-site by various researchers and how they were ultimately disposed of.
Automating some of this bookkeeping would not only free up time for other tasks, she said, but it might add another layer of oversight to reduce errors. Just as software simplified the chore of filing tax returns, it could do the same for environmental reporting.
“We have to know what materials of concern we’re using and measure how much we are using to progress toward cleaner, safer processes and systems,” said Jed Pitera, a materials researcher at IBM. “Beyond PFAS, we also as a society want to reduce water use and carbon emissions as well as toxic, persistent pollutants. We hope this tool can be a good start.”
Related posts
- Technical noteDmitry Krotov and Leo Kozachkov
A 360 review of AI agent benchmarks
ResearchKim MartineauThe simplest protocol for AI agents to work together
ResearchKim MartineauAn invisible watermark to keep tabs on tabular data
ResearchKim Martineau