Cleaning up contaminated soil, sediment and water is a huge but vital task globally. This can be a particular challenge on properties impacted by per- and polyfluoroalkyl substances (PFAS). These are manufactured chemicals used in many products -- firefighting foam, breathable outerwear, fast-food packaging and others. They’re sometimes called “forever chemicals” because they don’t break down easily. As time goes on, the toxicity of these chemicals becomes better understood and the need to remediate and reduce landowners’ liabilities becomes more and more important.
Now, current developments in information technology are helping unwrap the PFAS puzzle to drive more effective remediation of this threat to human and environmental health.
There are several uses of PFAS and thousands of different formulations of this class of chemicals, making it hard to determine the origin of the PFAS found on a site.
A major concern around PFAS contamination comes from its molecules’ durability. Some PFAS can survive high temperatures, making them a key ingredient in foam that can smother a fire. Aqueous Film-Forming Foam (AFFF, or “A-triple-F”) is a highly effective firefighting agent, especially useful for flammable liquid fires. Many airports, military bases and municipal fire departments have fire training areas where AFFF was used, some of which has leaked into the ground.
Because of the wide variety of potential sources, types and uses of PFAS, it is important to get as full a picture as possible, when there is a need to remediate PFAS impacts on a property. An effective site investigation needs to consider whether:
-
PFAS may be migrating via a plume of impacted groundwater -- a threat to the safety of the water wells of nearby residents or other receptors.
-
PFAS may have spread from the soil and groundwater into surface streams and water bodies, a possible threat to aquatic species and environments.
-
More problematic varieties: Some types of PFAS are attracting greater regulatory and stakeholder concern.
Information on these points can help:
-
Set priorities for the remediation work, so that areas of the site with greatest PFAS risks can be treated first -- for example, locations generating the greatest PFAS mass flux or affecting the most sensitive receptors.
-
Set budgets, so that spending goes to meet the greatest need.
-
Determining responsibility: Property owners can find out which PFAS impacts they are responsible for, and which contaminants may have been brought onto the site from another adjacent property, perhaps through groundwater flow, or via atmospheric deposition.
To understand a site with PFAS impacts better, the typical procedure is to take samples of soil, water or other environmental media from the property. Each sample will have a unique chemical signature of PFAS, which will be used to build a picture of the concentration, types of PFAS, and probable origin. This process is called “fingerprinting.”
This helps develop a Conceptual Site Model that shows the sources, contaminant transport, exposure pathways, and receptors that must be protected. This helps drive the creation of an effective remediation plan.
How machine learning supports PFAS fingerprinting
One problem comes from the large volumes of data coming from many PFAS testing programs. A site may yield hundreds of samples with dozens of PFAS analyzed per sample, and the task of assigning each sample to a “bucket” of PFAS types, is huge. What researchers need is a way to reduce that complexity, so they can understand the big picture.
Machine learning is fast becoming part of the solution. This is an aspect of artificial intelligence that’s concerned with statistical algorithms that can learn from data and generalize to unseen data, and thus perform tasks without explicit instructions. Machine learning can help make sense of vast streams of data, and “learn” so that over time, the results produced are more accurate. Due to leveraging powerful data-processing software like R or Python, application of these tools reduces the time spent on number-crunching otherwise done manually in a spreadsheet, as well as the time to repeat the analysis if updates are required. This can shorten the time to do the analysis by a few days or a couple of weeks, depending on the scope and other factors.
Unsupervised learning, a subset of machine learning, constitutes the bread and butter of multivariate chemical fingerprinting. Matrix factorization methods, like Principal Components Analysis (PCA) collapses large chemistry datasets with many chemicals down into a few pictures that are easier to understand. This is called dimensionality reduction.
When the data are presented in this smaller, simpler way, patterns begin to appear, including chemical trends and groups. If trends are present, matrix factorization automatically extracts the most unique fingerprints that bookend those trends. Clustering algorithms automatically partition the data into groups, and can be useful for separating conceptually different areas of a site. Even more advanced methods exist that can begin to reconcile multiple datasets – environmental features, fingerprints, PFAS precursors, etc.
Some possible benefits of machine learning for PFAS fate and transport assessment, and remediation are:
Determining cause and effect
Machine learning helps determine cause and effect. For example, we might determine the signature of PFAS found at a site, and track its migration through a plume of impacted groundwater by taking into account changes to the PFAS composition during transport. Then, PFAS with a related signature might be found in the wells used for drinking water on properties down gradient from the PFAS source, indicating the source of the well-water contaminants and help distinguish it from other PFAS sources. Machine learning can help assess whether and how the “fingerprints” of the source(s) and in the well water are related, and help support fate and transport evaluations.
Understanding background contamination
Machine learning can also help deal with the fact that, because these chemicals are so widespread, there are levels of PFAS that can be considered ambient or “background” – carried there by wind or precipitation. It’s important to understand what these ambient levels are and their characteristics, because it may well be that those ambient levels are higher than the prescribed cleanup levels. Any efforts to clean up to below those ambient levels will be effectively wasted, because those PFAS impacts will just keep coming back.
Refining the historical review analysis
Many site investigations start with an historical review, which is generally a desktop study relying on existing records such as historic records and aerial images to identify potential sources of contaminants such as PFAS. Machine learning analysis of test samples may turn up areas of potential concern that were missed in the historical review or provide better characterization of the areas identified, leading to a greater understanding of the site and where the risks may lie.
More understandable data and assistance with stakeholder communication
Machine learning can be used to generate visual outputs in the form of charts and graphs that are easily understood by stakeholders, as well as maps of the site that are color-coded to indicate types and severity of PFAS impacts as well as linkages between sources, plumes and receptors.
Going a step further towards planning mitigation
Machine learning can also be applied towards understanding the broader picture. Lessons learned and case studies from other sites are viable data that can also be mined for both qualitative and quantitative insights. For example, a desktop study of an entire site portfolio can assist with tracing waste streams and even help to prioritize which site to focus on.
Machine learning applied to PFAS site investigations is a good example of the use of this technology. It helps deliver results more quickly than traditional human analysts, and it helps increase confidence in the results. It helps project owners to determine which are the areas of a site most in need of remediation, to plan the work and set budgets appropriately. The result is better management decisions and a better outcome for all stakeholders.
