SUPPORTING INFORMATION

ASSESSMENT OF WOODCHIP BIOREACTOR CHARACTERISTICS AND

THEIR INFLUENCES ON JOINT NITRATE AND PESTICIDE REMOVAL

AUTHOR NAMES Olivia M. Wrightwood1, Madison E. Hattaway1, Thomas M. Young1, Heather N.

Bischel1*
1Department of Civil and Environmental Engineering, University of California, Davis

*CORRESPONDING AUTHOR ADDRESS

Heather N. Bischel
Department of Civil and Environmental Engineering
2001 Ghausi Hall
University of California
1 Shields Avenue
Davis, CA 95616
hbischel@ucdavis.edu

CONTENTS

SI 1 – Diagrams of two bench-scale bioreactor designs

SI 2—Dissolved oxygen, water level, and nitrate levels across sequencing batch reactor cycles

SI 3—LC-MS operational information

SI 4 – Monitored field woodchip bioreactor

SI 5—Small-scale (40mL) batch test nitrate removal

SI 6—Hydraulic and denitrification model construction, and bioreactor scaling considerations

SI 7 – Composition of synthetic tile drainage

SI 8—Denitrification results from air exposure time (AET) experiments

SI 9—Imidacloprid isotherm and woodchip extraction results

SI 10 – Modelling and statistical analysis of small-scale kinetic batch tests

SI 11 – Description of workflow and experimental conditions

1
SI 1 – Diagrams of two bench-scale bioreactor designs

Figure SI 1.1. Continuous-flow bench-scale bioreactor design. The reactors were constructed from 3/8”
clear cast acrylic sheets, with bulk volume dimensions 53.3cm x 11.7cm x 14.0cm (LxWxH). Because the
dimension ratios differed from the field bioreactors, four vertical baffles were placed along the length of
the bench-scale reactors to limit short circuiting and to achieve similar dispersion coefficient to the
monitored field bioreactors.

Figure SI 1.2. Sequencing-batch bench-scale bioreactor design. Fed with the same synthetic tile drainage
as described in the Figure 1 caption, this design is gravity fed and gravity drained. Filling and draining are
controlled by an Arduino program logic controller (PLC) that received signals from a float switch to
indicate water level and programmed hydraulic residence and air exposed times.

2
SI 2—Dissolved oxygen, water level, and nitrate levels across sequencing batch reactor cycles

Figure SI 2. Profile of dissolved oxygen (DO) concentration throughout seven fill/refill cycles in the
sequencing batch style reactor. Presented plot shows dynamics of operating at a 12h HRT and 0h AET;
other HRT:AET ratios exhibited similar behavior in terms of introducing DO into the treatment cycle.

SI 3—LC-MS operational information

LC-QTOF-MS (Agilent 1260 Infinity HPLC coupled to an Agilent 6530 QTOF-MS with a Zorbax
Eclipse Plus C18 column; 100 mm, 2.5 mm, 1.8 μm, Agilent Technologies, Inc.) analysis was as described
in Moschet et al. (2016). 20 μL of extract was injected with the following mobile phases used in a 23 min
run at a flow rate of 0.35 mL/min in electrospray ionization (ESI) positive mode: (A) deionized water plus
0.1% formic acid, (B) acetonitrile plus 0.1% formic acid. All-Ions fragmentation mode was used for
acquisition, meaning all ions with m/z 50−1050 were fragmented in the collision cell with collision
energies (CE) of 0, 10, 20, and 40 eV. CE = 0 means no fragmentation and is equivalent to a full MS scan.

Imidacloprid and diuron were quantified using Agilent MassHunter Quantitative Analysis
software (B.09). The [M+H]+ ion was used as the quantifier (exact mass window ±10 ppm) with the two
most abundant MS/MS fragments in the higher collision energy scans used as qualifiers. Imidacloprid-d4
and diuron-d6 were added to extracts as internal standards for quantification. Method blanks (ultrapure
water extracted with the same method as sample water), prespiked (known concentration of pesticide
standard mix added to water samples before extraction) samples, and postspiked (known concentration

3
of pesticide standard mix added to water samples after extraction) samples were used for quality
control purposes in quantification.

Woodchip extractions were analyzed “semi-quantitatively” as pre- and postspikes were not
performed across distinct sampling events. Pre- and postspiking of woodchip samples was performed
for the woodchips used in the October 2020 experiments, yielding 131% and 155% recovery of
imidacloprid, respectively. The excess recovery of the pre- and postspiked imidacloprid was likely due to
varying levels of imidacloprid still adsorbed to the woodchips, and differences in these imidacloprid
levels between grab samples. Therefore, a “semi-quantitative” approach was used for woodchip sample
extractions, where pre- and postspike corrections were not performed after pesticide analysis and
quantification.

SI 4 – Monitored field woodchip bioreactor

Figure SI 4.1. Photograph of PG&E Multichannel Bioreactor, Castroville, CA. There are three woodchip
channels at this site, with dimensions 22.6m x 1.68m x 0.762m (LxWxH).

SI 4.2—Field monitoring results: pH, temperature, and electrical conductivity (EC)

4
Figure SI 4.2.1. Field monitoring results of pH readings along the length of the Multichannel Bioreactor
channels. Channel readings were averaged for each sampling event.

Figure SI 4.2.2. Field monitoring results of temperature readings along the length of the PG&E
Multichannel Bioreactor channels. Channel readings were averaged for each sampling event.

Figure SI 4.2.3. Field monitoring results of conductivity (EC) readings along the length of the PG&E
Multichannel Bioreactor channels. Channel readings were averaged for each sampling event.

5
SI 4.3—Field measurements of dissolved organic carbon (DOC) and sulfate
Dissolved Organic Carbon (mg-C/L)
25

20

15

10

0
9/24/2018 1/24/2019 5/24/2019
Field Sampling Date

Influent Effluent (Avg., n = 3)

Figure SI 4.3.1. Dissolved organic carbon (DOC) concentrations at the influent and effluent of field
bioreactors. Grab samples taken on listed sampling dates and analyzed at UC Davis. While seasonal
differences may be attributed to irrigation patterns and storm events, there appeared to be no large
difference between the influent and effluent concentrations across all sampling events. This suggests
that DOC is not a limiting factor in the field bioreactors as an electron donor for microbial processes
such as denitrification.

500
Sulfate concentration (mg/L)

450
400
350
300
250
200
150
100
50
0
9/24/2018 1/24/2019 5/24/2019
Field Sampling Date

Influent Effluent (Avg., n=3)

6
Figure SI 4.3.2. Sulfate concentrations at the influent and effluent of field bioreactors. Grab samples
taken on listed sampling dates and analyzed at UC Davis. Effluent series error bars indicate standard
deviation (n = 3). Apart from the February 2019 sampling event (during a storm event), concentrations
remained relatively constant, showing no appreciable reduction or gain across the woodchip
bioreactors.

SI 5—Small-scale (40mL) batch test nitrate removal

Figure SI 5.1. Nitrate concentration by treatment during the imidacloprid kinetic batch test. The
microbial suppression of the microbial control (MC) treatment adequately suppressed denitrification, as
the MC curve greatly resembles the no woodchip control curve. Nitrate was completely removed in the
treatment (microbially active) batch reactors by day 8 (192hrs). Error bars represent the standard
deviation of treatment triplicates.

7
Figure SI 5.2. Nitrate concentration by treatment during the diuron kinetic batch test. The microbial
suppression of the microbial control (MC) treatment adequately suppressed denitrification, as the MC
curve greatly resembles the no woodchip control curve. Most of the nitrate was removed in the
treatment (microbially active) batch reactors by day 2 (48hrs).

SI 6—Hydraulic and denitrification model construction, and bioreactor scaling considerations

SI 6.1: Modeling potassium bromide tracer data

A dispersed-flow model (DFM) or plug flow with dispersion reactor (PFDR) model was
constructed by two methods (Crittenden et al., 2012). Here, the tracer data was used to fit an effluent
tracer concentration model as a function of time, C(t), using Péclet and mean detention time of the
residence time distribution as fitting parameters. The model for C(t) is shown in Equation SI 1 as a
function of E(t), defined as the residence time distribution function.

Equation SI 1
2

𝑒𝑥𝑝
[
(1 ― 𝑡 𝑡)
4( 𝑡)( 𝑃𝑒)
𝑡 1 ]
𝐶(𝑡) = (𝑀𝑄)𝐸(𝑡) = (𝑀𝑄) 2𝑡 𝜋(𝑡 )(1 )
𝑡 𝑃𝑒

Fitting parameters, the dimensionless Péclet number (𝑃𝑒) and mean residence time (𝑡) were
determined using two methods: the integral/trapezoidal method, and a non-linear regression. The
integral method employed Equations SI 2 and SI 3 to calculate (𝑡) and (𝑃𝑒), respectively. The integrals of
Equations SI 2 and SI 3 were approximated by summations across the tracer measurement time
intervals.
∞
∫0 𝐶𝑡 𝑑𝑡
𝑡= ∞ Equation SI 2
∫0 𝐶 𝑑𝑡

2𝑡2
𝑃𝑒 ≈ ∞ Equation SI 3
∫0 (𝑡 ― 𝑡)2𝐶 𝑑𝑡 ∞
∫0 𝐶 𝑑𝑡

The non-linear regression was carried out with the nls() function in R (The R Project for
Statistical Computing). Note: C(t) was modeled en lieu of E(t) due to the need to also fit the added
bromide tracer mass (M). This was because the bromide concentration of the added tracer solution was
not obtained due to sample loss and was instead fitted as a third fitting parameter within bounds of
expected values of added bromide mass.

Table SI 6.1. Dispersed Flow Model (DFM) comparison for field tracer data

8
 𝑡 Pe 𝑡 95% CI Pe 95% CI
Integral Estimation/Trapezoidal Method (Ch. 5) 48.55 12.01 - -
Nonlinear Regression Method (Ch. 1) 16.50 14.32 [16.2,16.8] [13.2,15.4]
Nonlinear Regression Method (Ch. 5) 43.20 22.13 [42.3,44.1] [18.8,25.4]
SI 6.2: Denitrification kinetic modeling methods

Three methods were used and compared for the determination of rate law parameters (e.g.,
order and rate constant) for denitrification at bench scale. The experimental data used to fit the three
below-described models was obtained by running a series of batch-style experiments in the constructed
bench-scale reactors at different hydraulic residence times (HRTs). The influent and effluent nitrate
concentrations were measured over a minimum of four treatment volumes for each HRT. Three HRTs
were observed: 6 hours, 12 hours and 24 hours. The resulting dataset was used to calculate
denitrification rate (reactant disappearance rate) as well as average concentration over the time
intervals, in this case between HRTs.

The first method to fit the obtained data to a rate law was the differential method. The
differential method graphically fits (linear regression) kinetic rate data to a linearized generic rate law
(Eq. SI 4) by plotting log10(-dC/dt) vs. log10(𝐶), where -dC/dt is the rate and 𝐶 is the average nitrate
concentration over an experimental time interval. The y-intercept of the resulting line is log10(k) and the
slope is n.
𝑑𝐶
𝑑𝑡 = ―𝑘𝐶𝑛 Equation SI 4

The second method used was the integral method. Like the differential method, the integral method is
also a graphical method in which a rate order is assumed, and the kinetic data is linearized and fit to a
rate law via linear regression. The linearization is plotted as log10(C/C0) vs. time, with the plot forced
through the origin. Of the resulting line, the slope is equal to -k.

The third method used to model the denitrification rate was a nonlinear regression to fit the
kinetic data to the Michaelis-Menten model (Eq. SI 6). Like with the previous two models, the
experimentally observed rate and average nitrate concentration per time step were used to fit the
model, where the fitted parameters were 𝑉𝑚𝑎𝑥 and 𝐾𝑚.

The parameterization results of fitting the denitrification data to three kinetic models are
displayed in Table SI 6.2. The R2 and p-values were obtained from the linearizations of all three models;
the Lineweaver-Burk linearization was used for the Michaelis-Menten model. While R2 values are useful
for determining the overall fit of the linearization, in this case they create a misleading comparison
between models, as high R2 values may be indicative of model overfitting. Therefore, the mean squared
error (MSE) and the residual standard error (RSE) were calculated for each ‘non-linearized’ model in rate
form.

9
 Table SI 6.2. Bench-scale woodchip bioreactor denitrification kinetics.
linearized fit fit to rate data model parameters
Kmg Vmaxh
Mult. R2a Pb MSEc RSEd k (h-1)e nf
(mg/L) (mg/h)
Differential Method 0.895 0.2100 0.210 0.793 0.187 0.863

Integral Method 1st order 0.975 0.0017 1.515 1.507 0.223 1

Michaelis-Menten 0.991 0.0622 0.054 0.401 4.774 2.193

∑(𝑦𝑖 ― 𝑦𝑖)2
a Multiple R-squared (coefficient of determination): 𝑅2 = 1 ― ∑(𝑦𝑖 ― 𝑦)2
b p-value of regression analysis F-statistic; evaluation of difference of all regression coefficients = 0.
1 2
c Mean squared error: 𝑀𝑆𝐸 = 𝑛∑(𝑦𝑖 ― 𝑦𝑖)
1
d Relative squared error: 𝑅𝑆𝐸 = ∑
𝑛 ― 2 (𝑦𝑖 ― 𝑦𝑖)2
𝑑[𝐴]
e Rate constant, k, for generalized rate equation: for 𝑟 = 𝑘[𝐴]𝑥[𝐵]𝑦 𝑓𝑜𝑟 𝑎𝐴 + 𝑏𝐵 →𝑐𝐶; First order: 𝑟 = ― 𝑑𝑡 = 𝑘[𝐴]
𝑑[𝐴] 𝑛
f Reaction order, n, in generalized rate law equation: ― = 𝑘[𝐴]
𝑑𝑡
𝑘𝑜𝑓𝑓 + 𝑘𝑐𝑎𝑡 𝑘𝑜𝑛/𝑘𝑜𝑓𝑓 𝑘𝑐𝑎𝑡
g Michaelis constant, Km, given by: 𝐾𝑚 = 𝑓𝑜𝑟 𝑒𝑛𝑧𝑦𝑚𝑎𝑡𝑖𝑐 𝑟𝑒𝑎𝑐𝑡𝑖𝑜𝑛 [𝐸] + [𝑆] [𝐸𝑆] [𝐸] + [𝑃]
𝑘𝑜𝑛
h Maximum reaction rate, Vmax: 𝑉𝑚𝑎𝑥 = 𝑘𝑐𝑎𝑡[𝐸]𝑇 𝑤ℎ𝑒𝑟𝑒 [𝐸]𝑇 = [𝐸] + [𝐸𝑆]

Upon comparing the MSE and RSE values of each model, it is apparent that the Michaelis-
Menten model has the lowest error indices, followed by the differentially obtained model then the first-
order kinetic model. Obtaining the lowest error indices with the Michaelis-Menten fit makes sense
theoretically, as the model is derived from an enzymatic reaction network; the Vmax parameter is a
function of enzyme density. Because determining enzyme concentration in a woodchip bioreactor
system would be experimentally difficult, it may be assumed to be proportional to the woodchip surface
area, as the enzyme-producing microbial biomass attaches to the woodchip surfaces. Though there are
several enzyme-catalyzed elementary reactions comprised in the overall denitrification reaction, the
model may be an average of each step.

The differential and integral fitting methods yielded the second lowest and greatest MSE and
RSE values, respectively. This is unsurprising, as the differential method fits two parameters, obtaining a
more precise model than the integral method.

10
Figure SI 6.2. Plotted Michaelis-Menten model from this study alongside obtained zero-order rate
constant from Krone et al.26 adjusted to 22°C using Krone et al.’s obtained Arrhenius constant, 1.12.

SI 6.3: Field- to Bench- scale bioreactor scaling considerations

Hydraulics between the field- and bench- scale woodchip bioreactors with the bromide tracer
data from each. The non-linear regression fitted parameters, namely the Péclet number (Eq. SI 5) were
compared. To achieve a similar Péclet number to the field bioreactors, diffusive mass transfer at the
bench-scale was limited by sieving woodchips to a smaller sieve size (9mm) and by placing vertical
baffles along the reactor length (Figure SI 1.1). Like in the field, a conservative bromide tracer was used
in the bench-scale reactors to characterize hydraulic behavior. The Péclet numbers of both the field- and
bench-scale reactors were calculated from the bromide tracer data to confirm similar dispersion
behavior.
𝐿𝑢
𝑃𝑒 = 𝐷 Equation SI 5

Where: L = characteristic length

u = local flow velocity
D = mass diffusion coefficient

Table SI 6.3: Fitted Péclet and HRT values from tracer studiesa
Bench Scale Field - Ch1 Field - Ch5
Pe 6.422 0.9228 14.3161 0.5288 22.1344 1.5809
HRT (tbar) 43.43 1.47 16.4982 0.132 43.2037 0.4272
MDI 3.83 2.54 2.13
aFitted value with respective standard error values

11
 It is evident that even between channels of the same field-scale bioreactor, the hydraulics can
vary significantly. Both Channel 1 and Channel 5 were originally determined to have a theoretical HRT of
24h based on original porosity and volumetric flow rate to the channels. The disparity in hydraulic
character between the two channels is likely due to different levels of sedimentation (primarily loading
from storm events) that caused clogs and channelization and early breakthrough in the field-scale Channel
1 bioreactor during the winter of 2019, prior to sampling. Because the modelled hydraulic retention times
between the bench-scale and Channel 5 field-scale reactors were most similar, as presented in Figure SI
6.3.1.

Figure SI 6.3.1. The results of fitting plug flow with diffusion models to the bromide tracer results from
the field-scale bioreactor, Channel 5 and Channel 1 (a) and the bench-scale continuous-flow reactor (b).
Fitted average HRT value from Channel 5 most closely resembled that of the bench-scale reactor.
Bromide data from Krone et al.26 was used to fit each of the field curves.

12
 Microbial activity was also considered when scaling bench-scale reactors. Denitrification efficacy
was used to confirm biochemical similarity between scales. Denitrification depends on substrate
(nitrate) concentration as well as enzyme concentration and is often modeled with Michaelis-Menten
kinetics (Eq. SI 6) in soils,53 wastewater54 as well as woodchip bioreactors.44
𝑑[𝑁] 𝑘𝑐𝑎𝑡[𝐸]𝑇[𝑁]
𝑑𝑡 =― 𝐾𝑚 + [𝑁] Equation SI 6

𝑑[𝑁]
Where: 𝑑𝑡 = rate of nitrate consumption, M·s-1

𝑘𝑐𝑎𝑡= catalytic rate constant, s-1

[𝑁] = Nitrate concentration, M

[𝐸]𝑇 = Enzyme concentration, M
𝑑[𝑁]
𝐾𝑚 = [𝑁] when: 𝑑𝑡 = 𝑘𝑐𝑎𝑡[𝐸]𝑇 (half-saturation constant), M

Instead of direct enzyme quantification, enzyme concentration within the field bioreactors was
approximated by porosity, a parameter that captures the relative volume of woodchips (and thus
presence of enzymes found in the attached biofilm) used for treatment. Because of this, we assumed
the Damköhler numbers to be similar between the field and lab-scale bioreactors, as the chemical
kinetic parameters (enzyme concentration/porosity, influent concentration, and HRT) remained
comparable. Denitrification efficiency of three separate bench-scale operation intervals are displayed in
Figure SI 6.3.3.

Figure SI 6.3.3. The denitrification efficiency of the continuous-flow reactors across three separate
bench-scale operations.

SI 7 – Composition of synthetic tile drainage

The final salt concentrations were as follows: 3.7mmol MgSO4, 1.7mmol KNO3 (Ward’s Science,
Rochester NY), 7.5mmol NaCl and 5.7mmol NaHCO3 (BioWORLD Molecular Tools for Life Science, Dublin
OH). Imidacloprid and diuron were chosen as the pesticides of interest. These three compounds
represent three different classes of pesticides, exhibiting a range of hydrophobicities, molecular size,
and functional groups. The commercial formulations used are Altriset Termiticide by Syngenta Crop

13
Protection, LLC (chlorantraniliprole; 200.1 kg/L active ingredient), Diuron 4L by Drexel (diuron; 479.3
kg/L active ingredient), and Dominion 2L by Control Solutions Inc. (imidacloprid; 239.7 kg/L active
ingredient).

SI 8—Denitrification results from air exposure time (AET) experiments

Nitrate Removal with Increasing Air-Exposed Time

100
Nitrate Removal (%)

80

60

40

20

0
0h 2h 6h 12h
Air-Exposed Time (AET; hrs)

Figure SI 8.1: Results from testing the effect of air-exposed time (AET) on denitrification. Hydraulic
residence time (HRT) was kept at 12hrs for all trials. This set of experiments suggests that the air-
exposed time does not substantially compromise degree of denitrification.

SI 9—Imidacloprid isotherm and woodchip extraction results

Woodchip Imidacloprid Concentration

Percent Imidacloprid removed

90 250
80
70 200
(ng/g dwt.)

60
150
50
40
100
30
20 50
10
0 0
Feb_2020 Sep_2020 Oct_2020
Experiment Date

Imidacloprid Removal, Continuous Flow Woodchip Imidacloprid Concentration

Figure SI 9.1: Inverse relationship between percent imidacloprid removal and woodchip imidacloprid
concentration (ng/g dry weight). Imidacloprid extraction from each batch of woodchips was collected
prior to lab experiments to obtain initial levels of imidacloprid residue. Error bars indicate standard
deviations.

14
 900

800

700

600
x/m (ng/g)

500

400

300

200

100

0
0 10000 20000 30000 40000 50000 60000
Ce (ng/L)

300

250
x/m (ng/g)

200
x/m (Freundlich)
150 Isotherm Data
February 2020 Continuous-Flow
100
August 2020 Continuous-Flow

50 October 2020 Continuous Flow

0
0 2000 4000 6000 8000 10000 12000
Ce (ng/L)

Figure SI 9.2: Imidacloprid isotherm (line) with isotherm data and Ce and x/m (qe) values from the bench
scale continuous-flow reactors during the three side-by-side bench-scale experiments (i.e., February,
August, and October 2020). Panel (a) displays the complete isotherm and panel (b) displays a detailed
view of the lower range of the isotherm to visualize experimental data.

15
SI 10 – Modelling and statistical analysis of small-scale kinetic batch tests
The performance of the treatment batch (TB) and the microbial control (MC) small-scale batch
reactors were compared for both the imidacloprid and diuron kinetic batch tests. The kinetic time series
data was first fit to pseudo-first (Eq. SI 10.1) and pseudo-second order functions (Eq. SI 10.2), as
commonly performed with adsorption data.36,37 For each of the initial C0 values corresponding to the TB
and MC treatments for both pesticides, the average (n=3) aqueous concentration of the first time point
for each respective pesticide was used.
𝐶𝑡
𝐶0 = 𝑒 ― 𝑘1𝑡 Equation SI 10.1

𝐶𝑡 𝑘2𝐶0𝑡
𝐶0 = 1 ― 1+ 𝑘2𝐶0𝑡 Equation SI 10.2

Data from treatment batch and microbial control reactors were fit separately to each of the
pseudo-kinetic models using the nls() function in R (The R Project for Statistical Computing), making a
total of eight model fittings between the diuron and imidacloprid experiments. The best fitting model as
determined by mean squared error (MSE) and root mean squared error (RSE) was used to compare
fitted parameters between the TB and MC models for each pesticide (smaller values report lower errors
of the model fit). For all pesticide/treatment combinations, the pseudo-second order model fit better
than the pseudo-first order model, as indicated by the error terms in Table SI 10.1.

Table SI 10.1: Pseudo-first and -second model fits of small-scale batch tests.

First order model Second order model

Pesticide Treatment MSE RSE k1 MSE RSE k2

TB 2.28E-02 2.28E-02 1.68E-02 1.12E-02 1.09E-01 4.19E-05
Imidacloprid
MC 2.43E-02 1.60E-01 1.47E-02 1.17E-02 1.11E-01 3.36E-05

TB 5.15E-02 2.34E-01 1.12E-02 3.31E-02 1.87E-01 8.26E-07

Diuron
MC 3.03E-02 1.79E-01 5.84E-03 1.98E-02 1.45E-01 4.65E-07

Next, the pseudo-second order model was used to assess whether the model fits of each
treatment differed significantly between treatments, or if the two treatments may be modelled
together without worsening the model fit error. Physically, if the model performs worse when the two
treatments are modelled together, this indicates that the pesticide behavior differs between the two
treatments, showing that microbial activity may contribute to pesticide removal within the reactors.

This assessment was conducted by the addition of an interaction term into the pseudo-second order
model. Complete data (i.e., both TB and MC data) for each pesticide was fit to a pseudo-second order
model with an interaction term (Eq. 4), again displayed below:

16
 𝐶𝑡 (𝑘2 + 𝑏𝐺)𝐶0𝑡
𝐶0 = 1 ― 1+ (𝑘2 + 𝑏𝐺)𝐶0𝑡 Equation SI_10.3 (Eq. 2 in main text)
Where:
𝐶𝑡: aqueous concentration of pesticide at time 𝑡 (ng/L)
𝐶0: aqueous concentration of pesticide at initial time point (ng/L)
𝑘2: second order rate constant (h-1)
𝐺: binary treatment group code; TB = 1 and MC = 0
𝑏: interaction coefficient

For the initial C0 values in the “combined models”, the average (n=6) aqueous concentration of
the first time point (both treatments) for each pesticide was used. The b term ranges from 0 to 1 to
provide a scale of influence of the “group” term G: if the fitted interaction term (b) is statistically
different from 0, we conclude that G improves the model fit, and the two treatment groups (TB and MC)
are best modelled separately. If b is not statistically different from 0, we conclude that the two
treatment groups are best modelled together and are therefore statistically similar.

Table SI 10.2: Model results with interaction term and combined (TB & MC) pseudo-second order.

Pesticide MSE RSE k22 b b p-value

with int. term 2.67E-02 1.68E-01 4.19E-07 4.88E-07 1.89E-02
Imidacloprid
combined model 3.16E-02 1.80E-01 6.28E-07
with int. term 1.17E-02 1.11E-01 3.75E-05 0.00E+00 1.00E+00
Diuron
combined model 1.17E-02 1.10E-01 3.75E-05

Comparing the combined models (i.e., second order model fitted using both TB and MC data)
with the individual TB and MC models, we see that the imidacloprid treatments are best modelled
separately, and for the diuron treatments are better modelled together. This is illustrated by the
increase in error when the treatments were modelled together (RSETB < RSEMC < RSEcombined) and the
decrease in error when the treatments were modelled together (RSEcombined < RSEMC < RSETB) for
imidacloprid and diuron, respectively. Moreover, when modelled with an interaction term b, the
imidacloprid model fit improves with a value of b significantly different from zero (p < 0.05). This
indicates that the addition of a treatment identity term (i.e., TB or MC) significantly influences the
imidacloprid model fit when compared to a model that does not take treatment identity into account.
Because the imidacloprid TB curve achieves a greater level of imidacloprid removal (Figure SI 10.1),
microbial activity may provide some benefit in removing imidacloprid from water under denitrifying
conditions. However, even though statistically significant, the impact is relatively small as indicated by a
b interaction term very close to zero (b = 4.88E-7), meaning that adsorption is the dominant removal
mechanism compared to microbial degradation.

17
Figure SI 10.1: The separate and combined pseudo-second order models of imidacloprid adsorption.

Figure SI 10.2: The separate and combined pseudo-second order models of diuron adsorption.

18
SI 11 – Description of workflow and experimental conditions

19