Wintertime phytoplankton bloom in the Subarctic Pacific supported by continental margin iron
Phoebe J. Lam1,2, James K.B. Bishop1, Cara C. Henning3, Matthew A. Marcus4, Glenn A. Waychunas1 & Inez Y. Fung3
APPENDIX

A1. SAMPLE COLLECTION AND SELECTION
	We collected size-fractionated particulate samples using the Multiple Unit Large Volume in-situ Filtration System (MULVFS)[Bishop et al., 1985].  The MULVFS separates particles into 1-53 um and >53 um size fractions on microquartz and polyester mesh filters, respectively, during the filtration of up to 10,000 L of seawater; up to 12 samples are collected to depths of 1000 m. MULVFS samples preserve the morphology of fragile aggregates[Bishop et al., 1985], and do not induce particle aggregation [Bishop et al., 1977; Bishop, 1988].  All MULVFS samples were collected and processed in trace metal-clean conditions and stored dried in class 100 plastic bags.
	MULVFS samples were collected along "Line P" in the northeast Subarctic Pacific as part of the Canadian JGOFS program [Boyd et al., 1999].  We focus on samples collected in February and May 1996 from Ocean Station Papa (OSP, 50¡N 145¡W) in the high nutrient, low chlorophyll (HNLC) open ocean, and samples from Stations P16 (49.97¡N 134.67¡'W) and P4 (48.65¡N 126.67¡W) in the transition and coastal zones[Whitney et al., 1998], respectively, collected in February 1997 (cf. Fig. 9 for station locations).  MULVFS samples were also collected in January/February 2002 from the HNLC Subantarctic Pacific (55¡S, 170¡W) during the Southern Ocean Iron Experiment (SOFeX)[Bishop et al., 2004; Coale et al., 2004]. Southern Ocean samples used a 51 um instead of a 53 um polyester mesh filter. Here we analyze samples collected before iron addition. 

A2. ICP-MS ANALYSES 
	Bulk chemical properties of both size fractions of the MULVFS filters were analyzed by inductively coupled plasma mass spectrometry (ICP-MS).  Subsamples of the 1-53 um microquartz and >53 um polyester MULVFS filters (~1/40 of the 506.7 cm2 MULVFS filter area) were leached overnight in 0.6 N HCl (pH=0.22) at 60¡C. The leachate was filtered and analyzed on a Finnigan Element II Instrument at Lawrence Berkeley National Lab or on a VG PlasmaQuad PQ XS at the University of Victoria.  This procedure is quantitative for Ca[Bishop et al., 1985] and Mn oxides[Bishop and Fleisher, 1987], and is believed to recover hydrous iron-oxides, but is unlikely to allow detection of silicate-bound Fe. 

A3. XRF DATA COLLECTION
	We mapped the spatial distribution of Fe and other elements in marine particles at micron-scale resolution using the synchrotron x-ray microprobe at beamline 10.3.2 [Marcus et al., 2004]. The x-ray beam excites elements with electron binding energies below the incident beam energy; the resultant fluorescent x-ray emissions are detected by a solid state detector.  Since samples are not in a vacuum, fluorescent x-rays of elements lighter than S are attenuated by air and are not detectable.  
	For all samples, we used an incident beam energy of 8500 eV, a beam size of 7x7 um, and a step (pixel) size of 5 um.  We used a dwell time of 0.1 sec/pixel to map marine aggregates from the >53 um size fraction during the beamtime we had in July 2003 (hereafter referred to as BT1), December 2003 (BT2), and June 2004 (BT3).  We used a dwell time of 0.3 sec/pixel to map 1-53 um samples in September 2004 (BT4).  Typical XRF maps (1 mm2 for >53 um samples; 0.25 mm2 for 1-53 um samples) required approximately an hour to collect using these parameters.  A 1 mm2 map represented ~1/50,000 of the entire MULVFS filter area. The intensity of the incident beam, I0, was recorded to correct for the two-fold decay in beam intensity over each eight-hour storage ring fill cycle. Fluorescent x-ray counts were detected using a Canberra 7-element Ge detector and data were collected in 6 channels: the CaKa (3510-3850 eV), TiKa (4390-4620 eV), FeKa (6150-6690 eV), MnKa (5560-6150 eV), CrKa (5260-5660 eV), and total (all energy bins; 0-20470 eV) regions of interest. All XRF counts were normalized by dwell time and by I0.
	All XRF data were collected with a protective Be window in front of the detector except in BT2.  For BT2, XRF data were collected with no window, a Be window, a 175 um Kapton window, or a 25 um Al foil window in front of the detector. 
	To quantify XRF counts, samples mapped in BT2, BT3, and BT4 had accompanying XRF membrane standards (Micromatter Co.) for Fe (48.20 ug/cm2 as Fe metal) and Ca (47.5 ug/cm2 as CaF2).  A Ti (44.5 ug/cm2 as Ti metal) standard was also run with 1-53 um samples in BT4.
	The >53 um MULVFS samples consist of aggregates and the polyester mesh upon which they were collected. XRF maps of unused polyester mesh filters (Sefar/Tetko PeCap(c) polyester mesh) provided the basis for blank correction.  Separate maps were collected for mesh filters used in the Subarctic Pacific and during SOFeX since different mesh lots were used in each location.  The 1-53 um MULVFS samples were on microquartz fiber filters.  XRF maps of unused microquartz fiber filters were collected for blank correction.

A4. XRF DATA ANALYSIS: >53 um samples 
	We developed a data processing algorithm to quantify the >53 um XRF Fe data.  The major steps involved were as follows: we scaled XRF counts (section A4.1), identified and removed artifactual Fe counts not originating from the sample (section A4.2), determined a detection limit (section A4.3), and converted map XRF Fe into equivalent Fe concentrations in the water column (section A4.4).  

A4.1 Scaling of Counts
	Different sample-to-detector distances and detector windows used across the four beamtime runs caused variable attenuation of XRF counts. To make the data comparable across the different runs, XRF counts for all samples were scaled to a single reference level. 
	The detector configuration of the SOFeX mesh blank map from BT2 was chosen as a reference, and all samples from BT2 and BT3 were scaled to this reference with the aid of the Fe and Ca standards: Fe and Ca scaling factors were determined for each sample as the ratio of Fe or Ca counts/um2 for that sample's standard to Fe or Ca counts/um2 of the reference standard. The scaling factors ranged from 0.1 to 1 (Table A1). Fe and Ca counts for all samples were quantitatively scaled by dividing by the Fe and Ca scaling factors, respectively. Ti counts were scaled by dividing by the Ca scaling factor.  Cr, Mn, and total counts were scaled by dividing by the Fe scaling factor. 
	The amount of x-ray attenuation in air and through different detector windows varies as a function of x-ray energy. The x-ray emission energies of Fe, Mn, and Cr are sufficiently similar and also high enough (FeKa1=6404 eV; MnKa1=5899 eV; CrKa1=5414 eV) that attenuation by air is negligible and by a thin Be window is comparable for the three elements: a 80 um Be window in front of the detector transmits 95.3%, 96.3%, and 97.1% at 5414, 5900, and 6400 eV, respectively (http://www.cxro.lbl.gov/optical_constants/filter2.html).  Ti counts scaled using the Ca scaling factor will be slightly overestimated since the x-ray emission energy for Ti (4511 eV) penetrates further than for Ca (3692 eV): a 80 um Be window transmits 92% of Ti x-rays compared to 85% of Ca x-rays.
	Samples mapped in BT1 did not have accompanying XRF standards. Mn and Ti in the polyester mesh yielded sufficiently high count rates (median mesh-Mn=528 cps; median mesh-Ti=4966 cps) to allow the mesh itself to be used as our standard since the same mesh was standardized in BT2.  Mn and Ti scaling factors were determined as the ratios of summed Mn and Ti counts from a 0.0625 mm2 area of the BT1 maps to those from a standardized mesh from BT2.  Mn scaling factors ranged from 0.7 to 1; Ti scaling factors ranged from 0.6-1.  Mn and Ti counts from BT1 were quantitatively scaled by dividing by the Mn and Ti scaling factors, respectively; Ca counts were scaled using the Ti scaling factor; Cr, Fe and total counts were scaled using the Mn scaling factor.

A4.2 Correction of Artifactual Counts
	Artifactual Fe counts result from count-rate dependent electronic noise in the detector, background Fe sources in the hutch and detector, Fe in the polyester mesh filter on which our sample was collected and mounted, and overlap with other fluorescent emission lines (notably MnKb from Mn in the mesh).  We used a multi-element approach to determine the amount of "Fe" from each of these sources. The polyester mesh contained high levels of Ti and Mn, so the TiKa and MnKa channels were used to locate pixels containing the mesh.  Since we do not expect Cr in our samples, we used the CrKa channel as a proxy for electronic noise.
	XRF maps of unused polyester mesh filters were used to define parameters used to determine the artifactual Fe counts arising from background Fe sources and the mesh. Electronic noise as a function of total XRF counts was determined empirically using the sample with the highest total count rate. The sequence of steps for the correction of artifactual counts was the following:  all counts were corrected for detector dead time (0.46us for the aggregate of 7 detector elements) using beamline software; we subtracted Fe counts from electronic noise first, then background Fe counts, and then Fe counts from the mesh.  The resultant Fe ("corrected-Fe") was assumed to originate entirely from the sample. Fig.A1 shows a FeKa map through each successive data analysis step. The details for how each step was performed are described below.

A4.2.1 Determination of background and mesh parameters. 
	Since each of the detector windows used in BT2 has a different x-ray attenuation spectrum, a unique set of background and mesh parameters used in our algorithm was determined for each mesh and detector window combination (SOFeX mesh: no window; Subarctic Pacific mesh: no window, Be, Kapton, Al).  In the text, we report mesh parameters for the unused Subarctic Pacific mesh with no detector window as examples; the complete list of mesh parameters for all mesh and detector window combinations are in Table A2.
	The detection of mesh at each pixel was evaluated by comparing Ti counts to a Ti threshold, which was defined as the 10th percentile of all Ti counts (Ti threshold=1583 cps; Table A2).  Pixels that had Ti counts below the Ti threshold were "holes" in the mesh (Fig A2). 
	Any counts coming from "hole" regions in the mesh blank were assumed to be from background sources in the experimental hutch (shielded x-ray chamber) or detector.  Background counts for Cr (37 cps), Mn (151 cps), Ca (342 cps), and total (17,020 cps) were determined as the mean value in all "hole" regions (Table A2), and removed from all pixels. Background Fe (116 cps) was also determined this way, but was calculated after electronic noise-Fe had been removed (see below).  Background Ti levels (1112 cps) were determined as the 5th percentile of all Ti counts (Table A2).

A4.2.2 Electronic noise.  
	The detector experiences a rise in its baseline at high count rates. We correct for this electronic noise by assuming there is a reproducible relationship between total counts and noise in each ROI.  Since there should be no Cr in our samples, any Cr counts remaining after removing background Cr could only be due to Cr in the mesh and from electronic noise in the detector.  The median background-corrected Cr:Mn ratio (0.1212) of all mesh pixels was determined on each mesh blank (Table A2).  For all samples, mesh-Cr was determined by multiplying the Mn at each mesh pixel (diagnosed by Ti) by the appropriate Cr:Mn ratio, and was subtracted from the background-corrected Cr map (Fig. A3).  All noise calculations were done on raw data that was not normalized to I0.  Remaining Cr counts after mesh was removed were assumed to be entirely due to electronic noise ("noise-Cr"). 
	We derived an empirical relationship between noise-Cr and total counts using the sample with the highest total counts ("96006p", OSP, 71m; 2x105 counts). Raw total counts were binned into 12 16,000-count wide bins and the mean noise-Cr in each bin was determined (Fig. A4a).  Noise-Cr values were linearly interpolated between bins to determine a Cr noise function (Fig. A4a), which was best fit to the raw FeKa data for this sample using a constant scaling factor of 2.5.  The resultant empirical Fe noise function (Fig. A4b) was applied equally to all samples to remove Fe counts due to electronic noise.  After noise removal (Fig. A1b), the background-Fe was removed from all pixels (Fig. A1c).

A4.2.3 Mesh-Fe removal.  
	After correcting for noise-Fe and background-Fe, remaining Fe counts were from the mesh ("mesh-Fe") and from the sample. Mesh-Fe was determined similarly to mesh-Cr: for each mesh filter, we plotted the noise- and background-corrected Fe against background corrected MnKa for all mesh pixels, and determined the slope of the best-fit line through the data forced through 0 (Mn:Fe=4.511; Table A2).  For each sample, the appropriate Mn:Fe slope was multiplied by background corrected MnKa at each mesh pixel to determine mesh-Fe (median mesh-Fe=117cps; Table A2). Since the sample is expected to have Mn, we defined a maximum Mn threshold (Max Mn Threshold=1300 cps) by rounding up from the maximum value of mesh-Mn (max Mn=1300 cps; Table A2) to ensure that we used Mn from the mesh and not the sample to remove mesh-Fe.  For mesh pixels with MnKa above the Mn threshold, we used the median mesh-Mn value (528 cps; Table A2) to calculate mesh-Fe.  Samples from the coastal P4 station had significant Mn levels (7-10 pmol Mn/mm2) compared to the mesh (30 pmol Mn/mm2), which precluded the use of Mn as a mesh-Fe determinant. For these samples, we repeated the above procedure using Ti instead of Mn to determine mesh-Fe (Table A2).
	Mesh-Fe counts were due to both Fe and Mn in the polyester mesh fibres.  This is because FeKa (6,397 eV) and MnKb (6,490 eV) counts are detected in the same ROI. The MnKb signal is generally 17% as intense as the MnKa1 signal, so given a median MnKa of 528 cps, we estimate that MnKb contributed 90 cps or 76% of the mesh-Fe signal.  Actual Fe in the mesh is thus 1.6 pmol Fe/mm2 when averaged over the filter, compared to 2550 pmol Ti/mm2 and 30 pmol Mn/mm2 in the mesh.

A4.2.4 Corrected-Fe.  
	After correcting for all artifactual counts, remaining Fe counts were assumed to come from Fe in the sample (Fig. A1d). The success of our algorithm was judged by applying it to a mesh blank.  A perfect algorithm would result in a median corrected-Fe of zero counts.  Our algorithm resulted in corrected-Fe medians of <1-4 counts/pixel, which was ~1% of our detection limit (Table A2).  We subtracted the median corrected-Fe of the mesh blank from each pixel.

A4.3. >53 um XRF Detectable Fe 
	The detection limit of XRF Fe per pixel (167 cps) was determined as 3 times the standard deviation of the corrected-Fe pixels of each mesh blank (Table A2), and represented the threshold above which we were confident the pixel had XRF detectable Fe coming from the sample.  "Hotspots" are defined to be Fe above the hotspot threshold, chosen to be 10 times the detection limit (1670 cps), which amounts to 2.7x10-9 umol Fe/pixel or 6.8x10-8 umol Fe/um2.  This threshold is significantly above the detection limit, and visually separates the Fe hotspots as distinct features from the diffuse low-level Fe in the sample (Fig. A5b). 
	We defined "hotspot Fe" and "XRF-detectable Fe" as the sum over an XRF map of all pixels with corrected-Fe counts greater than the hotspot threshold and detection limit, respectively. We defined "total XRF Fe" as the sum of corrected-Fe counts over all pixels of an XRF map.  Mesh blanks were subject to the same analysis procedure. Total XRF Fe, XRF-detectable Fe, and hotspot Fe from mesh blanks were subtracted from samples.
	We examined the sensitivity of the concentration of Fe hotspots to hotspot threshold definitions.  Lowering the threshold to 5 times the detection limit increased the hotspot Fe concentrations by ~50% and starts to integrate the abundant low level Fe that is distributed throughout the map (Fig. A5a).  Increasing the threshold to 15 times the detection limit (Fig. A5c) decreases the hotspot Fe concentrations by ~25%.  Although the absolute concentration of Fe hotspots is sensitive to the definition of the hotspot threshold, the samples all respond similarly to shifts in the threshold, so comparisons between samples are valid.

A4.4. Concentration of >53 um XRF Fe: bias correction factor
	To convert the Fe detected in a 1 mm2 XRF map to a Fe concentration in the water column, we first corrected for sample heterogeneity on the filter using Ca concentrations determined by both XRF and ICP-MS. ICP-MS values are a better representation of the bulk properties, since ~1/40 of the 506.7 cm2 MULVFS filter is analyzed, whereas only ~1/50,000 of the MULVFS filter is mapped by XRF. We determined a "bias correction factor", which was the ratio between the Ca/um2 determined by ICP-MS and Ca/um2 detected in an XRF map (Table A1). The Ca-based bias correction works for the February 1996 OSP samples and January/February 2002 SOFeX samples from 55¡S, since both sample sets had high particulate Ca as CaCO3 coccoliths spread throughout the aggregates.  The May 1996 OSP samples had low particulate Ca, so that there is greater uncertainty in the bias correction. 
	XRF Fe counts were summed over a map, multiplied by the bias correction factor (typically ~0.45; Table A1), converted to umol Fe using the Fe standard, and divided by the volume filtered through the equivalent filter area of the map (~100 mL; Table A1) to obtain the concentration of XRF determined Fe in the water column (Table 1).

A4.5. Estimates of Error for >53 um XRF Hotspot Fe Concentrations
	XRF analysis is non-destructive.  Duplicate maps were made of two OSP samples from February 1996 and provide a preliminary sense of uncertainty in our analysis.  The same aggregate in "96008p" (OSP, 144 m) was mapped in both BT1 and BT2.  Corrected Fe hotspot concentration for the two maps differed by 18%.  Different areas of "96006p" (OSP, 71m) were mapped in BT1 and BT2.  Here, hotspot concentration differed by 14%.  These errors reflect the sum of the uncertainties in our data analysis algorithm.  Unlike February OSP samples, Fe hotspots in the SOFeX samples were much more rare, so Fe hotspot concentrations in SOFeX samples have higher uncertainty due to hotspot heterogeneity on the filter.  Different areas of "2002002p" (SOFeX, 15m), "2002005p" (SOFeX, 89m), and "2002014p" (SOFeX, 904m) differed by 33%, 53%, and 79%, respectively.

A5. XRF DATA ANALYSIS: 1-53 um Samples 
A5.1 Correction of Artifactual Counts
	All 1-53 um samples were mapped with the same detector configuration during BT4, so the scaling steps required for the >53 um samples were unnecessary.  In addition, since the microquartz filter is to first order a smooth surface compared to the >53 um polyester mesh, the calculation of corrected Fe counts was much more straightforward.  Counts were converted to umol using Fe and Ti standards.  All XRF counts were corrected for detector dead time, and were normalized by I0 and dwell time. We determined background and filter contributions to XRF counts as the median count rate over the XRF maps of unused filter blanks from the Subarctic Pacific series (Fe blank=1099 cps; Ti blank=73 cps) and from SOFeX (Fe blank=1161 cps; Ti blank=83 cps). These Fe and Ti values were subtracted from the XRF Fe and Ti maps to give corrected Fe and Ti counts.  The Fe detection limit, which was three times the standard deviation of the background-subtracted blank, was 457 cps for Subarctic Pacific samples, and 746 cps for SOFeX samples.  We used hotspot thresholds defined on the >53 um size fraction (10-9 umol Fe/um2) (section A4.3); this was equivalent to 2500 cps.  

A5.2 Concentration of 1-53 um XRF Fe: attenuation factor
	Sample heterogeneity was not a problem for XRF mapping of the 1-53 um size fraction, since biomass is spread evenly throughout the filter. XRF counts seem to experience more attenuation in this size fraction, however, as evidenced by the 3-fold difference in ICP-MS determined Ca compared to XRF determined Ca on a dipped blank.  The attenuation is likely due to the fact that the microquartz filter captures material throughout the thickness of the filter rather than solely on the surface, so outgoing fluorescent x-rays may have to travel through quartz fibers before reaching the detector.  To correct for attenuation of XRF Fe counts due to the quartz filter, we first determined the level of attenuation for Ca XRF counts, and then corrected for the differences in energy between Ca (CaKa=3690 eV) and Fe (FeKa=6397 eV) x-ray emission lines.  The level of attenuation for Ca XRF counts was determined as the ratio of ICP-MS Ca to XRF Ca for each sample ("Ca attenuation factor" in Table A3).  The transmittance, T, or fraction of emitted x-rays that reach the detector, can be described by the following equation:

		T=XRFdetector/XRFemitted=exp(-(4*Pi*b(l)/l)*x)

where b(l) is the wavelength- and substrate-dependent extinction coefficient, or the imaginary part of the refractive index, l is the wavelength of the x-ray, and x is the path length [Attwood, 1999].  The Ca attenuation factor is just the inverse of the transmittance of Ca.  However, Ca x-rays are at lower energy than Fe x-rays and are thus more strongly attenuated, so the Ca attenuation factor overcorrects for the attenuation of Fe x-rays.  The Fe attenuation factor can be derived from the Ca attenuation factor according to:

  Fe_attenuation_factor=1/exp(-(lCa/lFe)*(bCa/bFe)*ln(Ca_attenuation_factor))

where lCa and lFe are the wavelengths of CaKa and FeKa x-rays, and bCa and bFe are the extinction coefficients experienced by CaKa and FeKa x-rays.  bCa and bFe for CaKa and FeKa x-rays going through quartz of density=0.188 g/cm3, the approximate density of a Whatman QM-A filter, are bCa=1.63x10-7 and bFe=1.95x10-8 (http://www.cxro.lbl.gov/optical_constants/getdb2.html).
	XRF Fe counts were summed over a map, multiplied by the Fe attenuation factor (Table A3), converted to umol Fe using the Fe standard, and divided by the volume filtered through the equivalent filter area of the map (~40 mL; Table A3) to obtain the concentration of XRF determined Fe in the water column (Table 1).

A5.3 Ti:Fe Ratios of Hotspots
	Unlike the >53 um polyester mesh prefilter, microquartz filters used for the 1-53 um size fraction have low background Ti.  We identified individual hotspots as any contiguous pixels with corrected Fe counts above the hotspot threshold.  We determined the average molar Ti:Fe ratio for each hotspot identified in the 1-53 um samples using background-subtracted Ti and Fe values (counts converted to moles using Ti and Fe standards).  

A6. XAS DATA COLLECTION AND ANALYSIS
	We collected Fe K-edge x-ray absorption spectra (XAS) of Fe hotspots for Extended X-ray Absorption Fine Structure (EXAFS) analysis to determine Fe speciation. The >53 um and 1-53 um XRF maps of sample "96004" (OSP, 46m) were used to identify Fe hotspots for EXAFS.  We collected six EXAFS spectra for hotspots in the >53 um size fraction, and three for the 1-53 um size fraction.  We used a 16x7 um incident beam spot size on Fe hotspots. We also collected Fe K-edge EXAFS of three Fe oxide minerals (ferrihydrite, goethite (FeOOH), and hematite (Fe2O3)), an Fe silicate mineral (fayalite (Fe2SiO4)), and an organically bound molecule (Fe(III) oxalate (Fe2(C2O4)3.6H2O)) as reference compounds. Depending on the Fe count rate, 4-9 scans of 40 minutes each were collected and averaged for each hotspot or reference compound. Energy was calibrated using an iron metal foil.
	We used two approaches for preliminary EXAFS data analysis: linear least-squares fitting[Manceau et al., 2002] and shell-by-shell fits using theoretically calculated phase shift and amplitude functions[Brown et al., 1988].  
	We used EXAFS data analysis software developed at beamline 10.3.2[Marcus et al., 2004] for linear least squares fitting of the Fe hotspot using spectra from our reference compounds. We also used the program EXAFSpak[George and Pickering, 2000] for shell-by-shell fits using theoretical phase and amplitude functions after Fourier transforming and filtering for the hotspots [Manceau et al., 2002].   We generated phase and amplitude functions for O, Fe, and Si with Fe as the central absorber atom using the program FEFF[Rehr et al., 1991].  We used an energy threshold of 7125 eV.  We fit the first two electron backscattering shells for all hotspots, and fit the third shell for hotspots that had high enough data quality. Distance (R), coordination number (CN), the mean-square relative displacement (sigma2), and Delta E0 for each absorber-backscatterer pair were allowed to vary to optimize the fit for the first shell.  For higher shells, Delta E0 was fixed to the value determined from the first shell fit, and the other parameters allowed to vary. For simplicity, we used the fewest number of components that could adequately fit each shell.  For "Decspot 1", which was well-fit by a linear combination of iron hydroxide model compounds (goethite and ferrihydrite), CNs for the 2nd shell were fixed and distances were initialized to values reported in the literature for goethite [O'Day et al., 2004]. 

A7. SCANNING TRANSMISSION X-RAY MICROSCOPE (STXM)
	The STXM provides very high-resolution images (spatial resolution=40nm) in addition to providing chemical information.  Because STXM relies on transmission of x-rays for imaging, samples must be very thin (microns).  Our aggregates were generally too thick to be imaged, so we searched for Fe hotspots at the thin edges of aggregates.  Images were taken with an incident x-ray energy below (704 eV) and above (710 eV) the Fe L3-edge using a 2 ms dwell time.  Fe hotspots were located by subtracting the two images.

A8. OCEAN GENERAL CIRCULATION MODEL
A8.1 Model Description
	We ran the Community Climate System Model version 2.0.1 (CCSM2, www.ccsm.ucar.edu/models/ccsm2.0.1) with an active version of the ocean model (the Parallel Ocean Program, POP), and passive "data" versions of the atmosphere, land, and ice components. In this configuration, the ocean model is forced using monthly-averaged climatological windstresses, heat fluxes, and freshwater fluxes from a previous integration of the fully active coupled CCSM2, where the freshwater fluxes include both precipitation from the atmosphere model, river runoff from the land model, as well as melt water from ice.
	The ocean model is run at relatively coarse resolution (~1o in the North Pacific), but includes parameterizations for eddy-induced mixing. It also parameterizes the mixed layer using the KPP scheme [Large et al., 1994], which solves for the mixed layer depth according to the amount of turbulent mixing near the surface (giving a depth close to 40 m in February in our model, which is somewhat of an underestimate). 

A8.2 Tracer release and advection from the Continental Margin
	Tracer was released from grid boxes shallower than 200 m.  To implement this source on the model grid, we first found the depth of the ocean bottom at all coastal grid points next to the solid continent. The model grid gives a stair-step representation of a flat shelf at a single model level, but in reality the shelf slopes smoothly down from the surface.  Instead of having a unit tracer source flux originate solely from the flat bottom of the model grid cell, we allowed all vertical grid layers from the surface to the ocean bottom (down to 200 m) to be tracer sources as well.  To determine the contributions of each vertical grid layer above the ocean bottom, we approximated the fractional area of shelf at each vertical level as the thickness of the vertical grid cell normalized by the total depth to the bottom. We then multiplied this fraction by the horizontal area of grid cells above 200 m to approximate the area of shelf contribution at these shallower grid levels, and allowed these to be tracer sources with a unit flux of 1 m-2d-1.
	To advect the tracer, we used the centered advection scheme, which conserves the first moment of the tracer (mass), but gives some negative tracer concentrations at the leading edge of the tracer signal.

REFERENCES
	Attwood, D. T., Soft X-Rays and Extreme Ultraviolet Radiation: Principles and Applications, Cambridge University Press, Cambridge, 1999.
	Bishop, J. K. B., The Barite-Opal-Organic Carbon Association in Oceanic Particulate Matter, Nature, 332 (6162), 341-343,1988.
	Bishop, J. K. B., J. M. Edmond, D. R. Ketten, M. P. Bacon, and W. B. Silker, Chemistry, Biology, and Vertical Flux of Particulate Matter from Upper 400 M of Equatorial Atlantic Ocean, Deep-Sea Research, 24 (6), 511-&,1977.
	Bishop, J. K. B., and M. Q. Fleisher, Particulate Manganese Dynamics in Gulf-Stream Warm-Core Rings and Surrounding Waters of the Nw Atlantic, Geochimica Et Cosmochimica Acta, 51 (10), 2807-2825,1987.
	Bishop, J. K. B., D. Schupack, R. M. Sherrell, and M. Conte, A Multiple-Unit Large-Volume In-situ Filtration System for Sampling Oceanic Particulate Matter in Mesoscale Environments, Advances in Chemistry Series (209), 155-175,1985.
	Bishop, J. K. B., T. J. Wood, R. E. Davis, and J. T. Sherman, Robotic observations of enhanced carbon biomass and export at 55 degrees S during SOFeX, Science, 304 (5669), 417-420,2004.
	Boyd, P. W., P. J. Harrison, and B. D. Johnson, The Joint Global Ocean Flux Study (Canada) in the NE subarctic Pacific, Deep-Sea Research Part Ii-Topical Studies in Oceanography, 46 (11-12), 2345-2350,1999.
	Brown, G. E., G. Calas, G. A. Waychunas, and J. Petiau, X-ray absorption spectroscopy and its applications in mineralogy and geochemistry, in Spectroscopic Methods in Mineralogy and Geology, edited by F.C. Hawthorne, pp. 431-512, Mineralogical Society of America, Washington, D.C., 1988.
	Coale, K. H., et al., Southern ocean iron enrichment experiment: Carbon cycling in high- and low-Si waters, Science, 304 (5669), 408-414,2004.
	George, G. N., and I. J. Pickering, EXAFSPAK: A suite of computer programs for analysis of X-ray absorption spectra., Stanford Synchrotron Radiation Laboratory, Menlo Park, CA, 2000.
	Large, W. G., J. C. McWilliams, and S. C. Doney, Oceanic vertical mixing: a review and a model with a nonlocal boundary layer parameterization, Reviews of Geophysics, 32 (4), 363-404,1994.
	Manceau, A., M. A. Marcus, and N. Tamura, Quantitative speciation of heavy metals in soils and sediments by synchrotron X-ray techniques, in Applications of Synchrotron Radiation in Low-Temperature Geochemistry and Environmental Sciences, pp. 341-428, 2002.
	Marcus, M. A., A. A. MacDowell, R. Celestre, A. Manceau, T. Miller, H. A. Padmore, and R. E. Sublett, Beamline 10.3.2 at ALS: a hard X-ray microprobe for environmental and materials sciences, Journal of Synchrotron Radiation, 11, 239-247,2004.
	O'Day, P. A., N. Rivera, R. Root, and S. A. Carroll, X-ray absorption spectroscopic study of Fe reference compounds for the analysis of natural sediments, American Mineralogist, 89 (4), 572-585,2004.
Rehr, J. J., J. M. Deleon, S. I. Zabinsky, and R. C. Albers, Theoretical X-Ray Absorption Fine-Structure Standards, Journal of the American Chemical Society, 113 (14), 5135-5140,1991.
	Whitney, F. A., C. S. Wong, and P. W. Boyd, Interannual variability in nitrate supply to surface waters of the Northeast Pacific Ocean, Marine Ecology-Progress Series, 170, 15-23,1998.