PFAS Prediction Using

Fingerprinting and
Machine Learning
Sarabeth George

Groundwater Resources Association of California’s PFAS Virtual Conference

April 28, 2020
A

(Paleo)climatology

B C

2
A

(Paleo)climatology

B C

3
1st quarter Public
Water System Wells:
570 wells tested;
23 > 70 ppt
Raw Data
Structure, Clean, and Scale Data
 Visualize Correlation

Correlation = 0.89

7
 Visualize Correlations

Correlation = 0.89

8
 Train Model and Predict [PFAS]

(true positive rate)

g
sin
es
gu
m
do
n
Ra

(false positive rate) 9

 (true positive rate)

Ra
n do
m
gu
es
sin
g
(false positive rate)
10
 With Model Without Model

447 exceedances detected

500 wells tested

(true positive rate)

g
sin
es
gu
m
do
n
Ra
(false positive rate)
11
 With Model Without Model

447 exceedances detected 20 exceedances detected

500 wells tested 500 wells tested

(true positive rate)

g
sin
es
gu
m
do
n
Ra
Faster response to high-risk wells (false positive rate)
12
 With Model Without Model

4% > EPA HAL

Detecting 100%
…. After testing every well in the state

(true positive rate)

90% > EPA HAL

g
sin
Detecting 90%

es
gu
m
do
n
Ra
Faster response to high-risk wells (false positive rate)
13
 With Model Without Model

100% > EPA HAL 4% > EPA HAL

Detecting 50% Detecting 100%
…. After testing every well in the state

(true positive rate)

90% > EPA HAL

g
sin
Detecting 90%

es
gu
m
do
n
Ra
Faster response to high-risk wells (false positive rate)
14
 With Model Without Model

100% > EPA HAL 4% > EPA HAL

Detecting 50% Detecting 100%
…. After testing every well in the state

(true positive rate)

90% > EPA HAL

g
sin
Detecting 90%

es
gu
m
do
60 % > EPA HAL

n
Ra
Detecting 98%

Faster response to high-risk wells (false positive rate)

15
 With Model Without Model

High Risk

100% > EPA HAL 4% > EPA HAL

Detecting 50% Detecting 100%
prioritization

…. After testing every well in the state

(true positive rate)

90% > EPA HAL

g
sin
Detecting 90%

es
gu
m
do
60 % > EPA HAL

n
Ra
Detecting 98%

Low Risk

Faster response to high-risk wells (false positive rate)

16
 udies indicate that in the ‘caret’ package.
erations and simpli-
Typical
mprove the Geoscience
predic- Modeling Approach
t al., 2009; Mfumu
effect of five differ-
of each statistical
ting area of a given
ent radii (1500 m,
ume a homogenous
ng site (Fig. 2). Fur-
5° and 90° oriented
fect of the dominat-
direction is derived
r surface for Hesse
tial data is derived

rate concentration,
,754 grid cells) and
predictors. The grid
2–S5 (SM).

17
ds, we evaluate four Knoll et.
Fig. 2. Five different buffer designs for the allocation of spatial al, 2019
data (https://doi.org/10.1016/j.scitotenv.2019.03.045)
to the groundwater
Novel Geoscience Modeling Approach

Cross Section
Map View W1 W2 W3

A’ A’
A
Water Bearing Zone 2

A
Water Bearing Zone 1 Fault

18
Application: ‘Fingerprint’ PFAS (and other!) chemical profiles,
identify source areas, and expedite the cleanup process

Industry A ‘Fingerprint’

19
Application: ‘Fingerprint’ PFAS (and other!) chemical profiles,
identify source areas, and expedite the cleanup process

Industry A ‘Fingerprint’

Unknown Sample

20
Application: ‘Fingerprint’ PFAS (and other!) chemical profiles,
identify source areas, and expedite the cleanup process

Industry A ‘Fingerprint’

Unknown Sample
Source = Likely Industry A

21
 Application: Guidance for monitoring programs where
groundwater is used as a source of drinking water

• Input well data Well ID Latitude Longitude Well Depth

• Output risk of HAL exceedance, other chemicals of concern, percent

identity to fingerprint(s)
Well ID Latitude Longitude Risk PFOS+PFOA Chemical 1 Chemical 2 Fingerprint
> 70ppt
66 % Source A
Fictional Well A 37.8x -121.9x 70% PCE AS 21% Source B
13% Source C

22
Questions & Discussion

PFAS Prediction Using Fingerprinting

and Machine Learning
Sarabeth George

23