the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
ROMSPath v1.0: offline particle tracking for the Regional Ocean Modeling System (ROMS)
Elias J. Hunter
Heidi L. Fuchs
John L. Wilkin
Gregory P. Gerbi
Robert J. Chant
Jessica C. Garwood
Download
- Final revised paper (published on 02 Jun 2022)
- Supplement to the final revised paper
- Preprint (discussion started on 20 Dec 2021)
Interactive discussion
Status: closed
-
RC1: 'Comment on gmd-2021-400', Morane Clavel-Henry, 06 Jan 2022
General Comment
In the following paragraphs, the editor(s) and the authors will find an assessment of the Manuscript of Geoscientific Model Development named “ROMSPath v1.0: Offline Particle Tracking for the Regional Ocean Modeling System (ROMS)”.
To my general understanding, the manuscript presented an updated and upgraded version of an Offline Particle Tracking model and described the impact of the main code modifications on the simulations.
The Model Development presented in the manuscript has several consequences, with mainly an improvement of the OPT model precision. Overall, I have no general remarks on the contents and core of the manuscript. The scientific approach and applied methods are valid. I think the manuscript is well balanced and has a clear structure. Nonetheless, I need some quantitative explanations when comparing the simulations of different scenarios. Indeed, while the results of the manuscript mainly relied on visual explanations (maps of particle distributions), some metrics could be more informative and clear (see Specific comments). Additionally, in the Discussion, I expected that the study’s results would be compared with existing publications. As for the manuscript format, I found it well written except for a few confusing sentences (See Technical comments). I would like the authors and the proofreading service to specifically care about the spacing mistakes (lack of space or double spaces) that were recurrent in the manuscript.
Specific comments
L.93-94: In that statement, I am curious about one thing: what about ROMS models that have a small spatial extent and, somehow, have a less pronounced curvilinearity; thus, potentially small errors in the coordinate interpolation? Would the performance of ROMSPath be still better than LTRANS? That is something I would have liked to see discussed as it has significant consequences for the choice of the software.
Figure 1: this figure should be put in a supplementary file. It is not a graph showing novelty and can be easily found on the website of ROMS.
L.114: Your hydrodynamic refinement ratio is 7:1. It is stated that a ratio higher than 5:1 can degrade the model performance (e.g., doi: 10.1016/j.pocean.2004.07.017 and within references). Was the hydrodynamic model verified on that point?
Table 1: 1) I need a rationale on why “2”, “30”, and “90” days transport duration and the particle number of “3285”, “6000”, and “32000” have been selected. 2) For the vertical experience (i.e., Vert. LTRANS and Vert. ROMSPath), I got confused. Please, indicate the depth range and also indicate that the release is made of evenly distributed points along a segment instead of “Line” (For example: Evenly distributed points between X and Y depths). As for “Point”, please, indicate the coordinates instead of “point”.
Section 3.3: I think that you should add in each section if you used both the parent and child hydrodynamic models (i.e., DOPPIO and SnailDel) to track particles or just one of the hydrodynamic models. See below)
- In line 286, you said you used the DOPPIO model for online tracking of particles (i.e., ROMSFloat). Did you also only use DOPPIO fields for particle tracking with LTRANS and ROMSPath?
- In sections 3.3.2 and 3.3.4, did you use DOPPIO and SnailDel, or just DOPPIO?
Sections 4.1 and 4.2: the main result (or global outcome) from the tests should be put on the first line of the paragraphs. In these two sections, I had an introduction of the figures instead of the main findings.
Section 4.1: the results from ROMSPath being closed to the online simulation ROMSfloat should be a valorised outcome of the manuscript. I expected a few comparisons with peer-reviewed studies that could have compared online and offline particle tracking simulations. Consider also my first comment (for L.93-94).
L.338-339: Is it relevant to write about a result when neither the methods nor supporting graphs are shown? It confused me because I am not sure what you refer to by this statement. I suggest removing these two sentences or to provide an annex with methods, results, and discussion.
L.352-354: Please, note that this is a non shown result that took half the paragraph of section 4.2. I think this result is interesting to have at least a supplementary figure and a short explanation of the method in 3.3.2.
Section 4.3.: 1) Considering the results relied only on visuals, I would have appreciated, in complementary, to have quantitative information such as a spatial aggregation index or the surface that contained 95% of particles at day X and per scenario. It would quantify the idea of “more horizontal dispersion”(L.366) and at least put some contrast between figures 6e and 6g. 2) Regarding the particles advected in the estuary with ‘Nest/No Turb’ but not with ‘No Nest/No Turb’, a small discussion would be welcomed. I don’t know the surface of the Delaware Bay but I easily guess that the resolution of the Doppio is too coarse for capturing the water circulation as the SnailDel can do. Hence the importance to do particle tracking simulation using the parent and child grid of hydrodynamic models in intertidal zones.
Figure 6: Please, be considerate of colour-blinded people and avoid having green and red on the same graph.
Section 4.4: Here too, I would appreciate some elements of discussion including comparison with peer-reviewed studies. This is an interesting result, which, beyond including it as a Model development, can have consequences for particle modelling in shallow marine areas in the future.
Technical comments
L.16: “now”, it looked like ROMSPath was already introduced before and this manuscript is going to present upgrades of the OPT. Please, remove.
L.18: “ROMSPath enables simulated particles to pass between nested ROMS grids, which are an increasingly popular tool to simulate” here, I think “tool” is not adequate for presenting the use of two ROMS grids. Maybe a “scheme” or a “module”.
L.30: a reference about the use of OPT in search and rescue is needed (e.g., doi: 10.1016/j.oceaneng.2021.110098)
L.35: I am confused. You “limited” your approach to 4-dimensional hydrodynamic model output. It can suggest that more than 4 dimensions exist. Did you mean you “focused” your discussion on 4D instead of using fewer dimensions?
L.36: Please, add references.
L.37: I think it should be clearer that you are talking about online particle tracking. “at model run time” can be slightly confusing because both “hydrodynamics” fields and “particle trajectories” are from a model. In other words, at first read, it is not evident which type of model “at model run time” points to.
L.88: “Necessary”. Please, remove. In the paragraph, it is not clear which “necessary” parameters are needed.
L.154: The formatting of Hunter and Visser’s citations should be improved: “Hunter et al. (1993) and Visser (1997)”.
L.157: “unrealistic particle clustering”. Please, include a reference for your statement.
L.194: “as is the case ‘here’”, do you mean “in your simulation”?
L.260: “following equation 3.3.5 in Phillips (1966)”.
L.291 and L.292: “two similar runs” and “at the same location”. Please, remind your reader that your simulation parametrisations are similar to the previous test (section 3.3.1).
L.311-312: “at 1 meter below the sea surface” why did you not indicate that precise statement at l. 286 too?
L.312: What does the “NJ” abbreviation stand for, and why do you use an abbreviation just once?
L.314: “released particles”
L.363: Please, remove “fairly”.
L.364-365: “The addition of the nested SnailDel grid in ‘Nest/No Turb’ is revealing. This sentence sounds empty. Please modify or delete.
L.365: “similar” To what it was similar?
L.368-369: “The time-averaged, near-bottom currents at the mouth of Delaware Bay are into the Bay[…]”. Please, reformulate.
L.390: Please, remove: “also”
Citation: https://doi.org/10.5194/gmd-2021-400-RC1 -
AC1: 'Reply on RC1', Elias Hunter, 01 Mar 2022
Dr. Clavel-Henry,
Thank you for your comments on our manuscript titled “ROMSPath v1.0: Offline Particle Tracking for the Regional Ocean Modeling System (ROMS)”. While the overall impression from these comments is positive, you note some areas for improvement. Specifically there is room for quantitative results related the test cases presented and discussion from relevant literature to give context to the results. We endeavored to respond to these concerns and address these issue point by point below. Note that the order of these comments have been rearranged somewhat. Your comments are in bold.
“L.93-94: In that statement, I am curious about one thing: what about ROMS models that have a small spatial extent and, somehow, have a less pronounced curvilinearity; thus, potentially small errors in the coordinate interpolation? Would the performance of ROMSPath be still better than LTRANS? That is something I would have liked to see discussed as it has significant consequences for the choice of the software.”
Yes, a configuration for LTRANS is possible where ROMSPath and LTRANS show similar results. This, however, requires careful consideration of domain size, location, study objectives, and choice of geographic reference. We will note this in the revised manuscript. However, a systematic examination of the conditions when LTRANS output matches ROMSPath output is outside the scope of this study.
“L.338-339: Is it relevant to write about a result when neither the methods nor supporting graphs are shown? It confused me because I am not sure what you refer to by this statement. I suggest removing these two sentences or to provide an annex with methods, results, and discussion. “
It is relevant in that it addresses the questions posed about L.93-94 above regarding the sensitivity of LTRANS results to basic configuration concerns. An LTRANS simulation with a reference coordinate horizontally distant from the grid showed results extremely divergent from the run with a reference coordinate selected using the recommended criterion. We thought this an unfair comparison as the LTRANS manual warns against it. We will remove the reference to the simulation and replace it with a comment on the careful consideration of LTRANS model parameters.
“Figure 1: this figure should be put in a supplementary file. It is not a graph showing novelty and can be easily found on the website of ROMS.
We will add it as a supplementary figure. Even though it available elsewhere, it is also represents a major change to the LTRANS code. “
“L.114: Your hydrodynamic refinement ratio is 7:1. It is stated that a ratio higher than 5:1 can degrade the model performance (e.g., doi: 10.1016/j.pocean.2004.07.017 and within references). Was the hydrodynamic model verified on that point?”
While the Barth et al (2005) recommends a refinement of 3:1 or 5:1, the Spall and Holland (1991) reference suggests 7:1 is acceptable. And a 7:1 ratio was successful in previous work, such as Warner et al (2017). We will note this in the manuscript. While a skill assessment of the ROMS output is outside the scope of this study, aspects of the estuarine circulation (tides, exchange flow, temperature, etc) were evaluated and found satisfactory. This will be the subject of upcoming work.
Barth, A., et al. (2005). "Two-way nested model of mesoscale circulation features in the Ligurian Sea." Progress in Oceanography 66(2-4): 171-189.
Spall, M. A. and W. R. Holland (1991). "A Nested Primitive Equation Model for Oceanic Applications." Journal of Physical Oceanography 21(2): 205-220.Warner, J. C., et al. (2017). "Inner-shelf ocean dynamics and seafloor morphologic changes during Hurricane Sandy." Continental Shelf Research 138: 1-18.
“Table 1: 1) I need a rationale on why “2”, “30”, and “90” days transport duration and the particle number of “3285”, “6000”, and “32000” have been selected. “
There are three sets off conditions listed in table 1. A) A 2-day run with 3285 particles. B) A 30-day run with 6000 Particles. C) A 90 day run with 32000 particles.
Case A is used to illustrate the vertical clustering issue with LTRANS relative to ROMSPath. 3285 is the number of particles evenly distributed throughout the water depth of the initialization point. While the simulation was run longer, 2 days is sufficient to illustrate the clustering issues. While the clustering is easily identifiable after 6 hours, the added time simply highlights the point.
Case B is used for a number of cases, in sections 3.3.3 and 3.3.4. Most related to nesting. 30 days was chosen as it was sufficient time to particles to enter and leave the snaildel domain. 6000 particles was used as a starting point for these runs, to minimize computational time while maintaining a coherent patch for analysis. With more particles added if needed. 6000 particles proved sufficient for coherent particle patches after 30 days.
The run time for Case C was increased to 90 days to allow for the possibility of particles traversing leaving the shelf. The number of particles in Case C was increased to 32000 due to the 90 day run time, as a decrease in coherence as the footprint of the particles is expected as time increase. Previous work suggested particle numbers in the 20k-30k range would suffice. 32000 was chosen for computational reasons unique to our computing cluster.
We will add this information to the manuscript for clarity.
“2) For the vertical experience (i.e., Vert. LTRANS and Vert. ROMSPath), I got confused. Please, indicate the depth range and also indicate that the release is made of evenly distributed points along a segment instead of “Line” (For example: Evenly distributed points between X and Y depths). As for “Point”, please, indicate the coordinates instead of “point.”
We will clarify these in the table and in the text. And see below.
Section 3.3: I think that you should add in each section if you used both the parent and child hydrodynamic models (i.e., DOPPIO and SnailDel) to track particles or just one of the hydrodynamic models. See below)
We attempted to make this clear in the table, under the nested column. Nested=yes means both Doppio and Snaildel output is used. Nested = no means Doppio only.
- In line 286, you said you used the DOPPIO model for online tracking of particles (i.e., ROMSFloat). Did you also only use DOPPIO fields for particle tracking with LTRANS and ROMSPath?
Yes.
2. In sections 3.3.2 and 3.3.4, did you use DOPPIO and SnailDel, or just DOPPIO?
Both.
Overall the clarity regarding the different model configurations is lacking. We will add more detail to the table, and to the text in the revised manuscript.
Sections 4.1 and 4.2: the main result (or global outcome) from the tests should be put on the first line of the paragraphs. In these two sections, I had an introduction of the figures instead of the main findings.
We will restructure these in the revised manyscript.
Section 4.1: the results from ROMSPath being closed to the online simulation ROMSfloat should be a valorised outcome of the manuscript. I expected a few comparisons with peer-reviewed studies that could have compared online and offline particle tracking simulations. Consider also my first comment (for L.93-94).
I am not entirely clear about what is meant by the term “valorised”. But we can comment on comparison with other peer reviewed studies.
While there is extensive literature on Lagrangian analysis of ocean, there are few studies directly comparing online vs offline particle tracking for the same model run. e.g. Wagner et al. (2019) and Cassiani et al (2016). However, these compare offline particles dispersion or online tracer dispersion. Typically the choice of offline vs. online is due to the practical considerations of computing time and storage space.
Cassiani, M., et al. (2016). "The offline Lagrangian particle model FLEXPART–NorESM/CAM (v1): model description and comparisons with the online NorESM transport scheme and with the reference FLEXPART model." Geoscientific Model Development 9(11): 4029-4048.
Wagner, P., et al. (2019). "Can Lagrangian Tracking Simulate Tracer Spreading in a High-Resolution Ocean General Circulation Model?" Journal of Physical Oceanography 49(5): 1141-1157.
L.352-354: Please, note that this is a nonshown result that took half the paragraph of section 4.2. I think this result is interesting to have at least a supplementary figure and a short explanation of the method in 3.3.2.
We will add a figure similar to Figure 5 as a supplement. The method is no different than the other figures, simply a configuration change.
“Section 4.3.: 1) Considering the results relied only on visuals, I would have appreciated, in complementary, to have quantitative information such as a spatial aggregation index or the surface that contained 95% of particles at day X and per scenario. It would quantify the idea of “more horizontal dispersion”(L.366) and at least put some contrast between figures 6e and 6g.”
We will add a dispersion coefficient to the panels in Figure 6. Calculated based on Lacase (2008). See new figure 6. The Dispersion coefficient increased consistent in the text. See attached
LaCasce, J. H. (2008). "Statistics from Lagrangian observations." Progress in Oceanography 77(1): 1-29.
“2) Regarding the particles advected in the estuary with ‘Nest/No Turb’ but not with ‘No Nest/No Turb’, a small discussion would be welcomed. I don’t know the surface of the Delaware Bay but I easily guess that the resolution of the Doppio is too coarse for capturing the water circulation as the SnailDel can do. Hence the importance to do particle tracking simulation using the parent and child grid of hydrodynamic models in intertidal zones.
Yes, the Doppio resolution is too coarse (7km). We need at least 1km to resolve the relevant Delaware Bay circulation. We will add a few sentences to this effect.
“Figure 6: Please, be considerate of colour-blinded people and avoid having green and red on the same graph.”
Of course, thank you for pointing it out.See attached
Section 4.4: Here too, I would appreciate some elements of discussion including comparison with peer-reviewed studies. This is an interesting result, which, beyond including it as a Model development, can have consequences for particle modelling in shallow marine areas in the future.
It is long known that stokes transport has an impact on coastal/estuarine circulation. We will add some discussion.
(Monismith and Fong(2004)) and in particular Delaware Bay(Pareja Roman et al. 2019)
Monismith, S. G. and D. A. Fong (2004). "A note on the potential transport of scalars and organisms by surface waves." Limnology and Oceanography 49(4): 1214-1217.
Pareja‐Roman, L. F., et al. (2019). "Effects of Locally Generated Wind Waves on the Momentum Budget and Subtidal Exchange in a Coastal Plain Estuary." Journal of Geophysical Research: Oceans 124(2): 1005-1028.
Feng, M., et al. (2011). "Ocean circulation, Stokes drift, and connectivity of western rock lobster (Panulirus cygnus) population." Canadian Journal of Fisheries and Aquatic Sciences 68(7): 1182-1196.
Kumar, N. and F. Feddersen (2017). "The Effect of Stokes Drift and Transient Rip Currents on the Inner Shelf. Part I: No Stratification." Journal of Physical Oceanography 47(1): 227-241.
van den Bremer, T. S. and O. Breivik (2018). "Stokes drift." Philos Trans A Math Phys Eng Sci 376(2111).
Kumar, N. and F. Feddersen (2017). "The Effect of Stokes Drift and Transient Rip Currents on the Inner Shelf. Part II: With Stratification." Journal of Physical Oceanography 47(1): 243-260.
-
RC2: 'Comment on gmd-2021-400', Anonymous Referee #2, 26 Jan 2022
Summary:
The manuscript ROMSPath v1.0: Offline Particle Tracking for the Regional Ocean Modeling System (ROMS) describes the details and benefits of a new offline particle tracking (OPT) tool for analysis of hydrodynamic model output. The new OPT, ROMSPath, is assessed by comparing particle tracking results to an existing OPT code, the LTRANS Lagrangian transport model, and online particle tracking by enabling ROMS floats. The comparisons were conducted for various configurations of the code in order to assess and demonstrate the utility of a range of new features available to ROMSPath.
My general assessment is that this paper presents a substantial advance in modelling science. The approach and methods are valid and the approach and assumptions underlying the particle tracking framework were clearly structured and explained. However, in the specific comments below I have outlined some areas where the results section in particular could be improved in order to support the interpretations and conclusions, in addition to some minor formatting comments.
Specific Comments:
Line 101: Include a comparative analysis of the number of particles that run aground in order to support this claim.
Lines 284-286, 291, 312: How was this specific method chosen? Was there sensitivity to using starting positions in different parts of the domain, for example? Were these based on dynamics or patterns observed in previous analyses?
Line 329: What does it mean that “The ROMSPath OTP output is always closest to the ROMS floats output”? Is this at each time, on average, or also for each particle trajectory? Please specify and quantify this distinction.
Line 340, Figure 4: Include additional quantitative support to summarize this comparison, such as dispersion, offshore transport, and trajectory of the center of mass.
Lines 377-379: Comment on the implications, and how this compares across particle tracking models (note, this relates to the previous comment re: Line 340). For example, is there evidence to suggest that the increased dispersion apparent with including small scale hydrodynamics through nesting and turbulent parameterizations in ROMSPath simulations improves the accuracy of this model compared to other formulations?
Line 393-394: It is hard to see from this figure that the particles tended to be closer to shore. Is there a statistic you can use for comparison, such as the mean distance from shore between the two, or the mean water depth of particles to test the significance of this observation?
Line 405: Where is the improvement in “efficiency” with ROMSPath relative to LTRANS demonstrated in the results section? Also, is the improved “accuracy” in relation to the native ROMS result? These two points should be clarified in the text of this summary.
Summary (i.e. Line 403): Can the authors comment on the results and features of ROMSPath in the context of other OPT applications (e.g. Lines 39-41), and in the context of previous works?
Formatting Comments:
Line 41: remove extra space after the parentheses
Line 63: “we use hydrodynamic model output generated for the larval transport study mentioned above” — this is not specific, please specify what study is referenced
Line 313: add a space after the period
Line 328: Capitalize “table”
Line 391: “thetwo” should be the two (space needed)
Citation: https://doi.org/10.5194/gmd-2021-400-RC2 -
AC2: 'Reply on RC2', Elias Hunter, 01 Mar 2022
Referee 2,
Thank you for your comments on our manuscript titled “ROMSPath v1.0: Offline Particle Tracking for the Regional Ocean Modeling System (ROMS)”. We appreciate the positive response to our work and the useful criticism below. We will address these point-by-point. Reviewer comments are in bold.
Line 101: Include a comparative analysis of the number of particles that run aground in order to support this claim.
Results will be in section 4.1. For the cases of the LTRANS OTP vs ROMSPath OTP. 34% of the LTRANS particles are identified as passing through a “Land” grid cell at least once. . As opposed to ROMSPath which ls <0.01%
Lines 284-286, 291, 312: How was this specific method chosen? Was there sensitivity to using starting positions in different parts of the domain, for example? Were these based on dynamics or patterns observed in previous analyses?
This is copied from a reply to Dr. Clavel-Henry. Referee #1
There are three sets of conditions listed in table 1. A) A 2-day run with 3285 particles. B) A 30-day run with 6000 Particles. C) A 90 day run with 32000 particles.
Case A is used to illustrate the vertical clustering issue with LTRANS relative to ROMSPath. 3285 is the number of particles evenly distributed throughout the water depth of the initialization point. While the simulation was run longer, 2 days is sufficient to illustrate the clustering issues. While the clustering is easily identifiable after 6 hours, the added time simply highlights the point.
Case B is used for a number of cases, in sections 3.3.3 and 3.3.4. Most related to nesting. 30 days was chosen as it was sufficient time to particles to enter and leave the snaildel domain. 6000 particles was used as a starting point for these runs, to minimize computational time while maintaining a coherent patch for analysis. With more particles added if needed. 6000 particles proved sufficient for coherent particle patches after 30 days.
The run time for Case C was increased to 90 days to allow for the possibility of particles traversing leaving the shelf. The number of particles in Case C was increased to 32000 due to the 90 day run time, as a decrease in coherence as the footprint of the particles is expected as time increase. Previous work suggested particle numbers in the 20k-30k range would suffice. 32000 was chosen for computational reasons unique to our computing cluster.
This information will be added to the revised manuscript
Line 329: What does it mean that “The ROMSPath OTP output is always closest to the ROMS floats output”? Is this at each time, on average, or also for each particle trajectory? Please specify and quantify this distinction.
We will clarify this in the text. See Below
Line 340, Figure 4: Include additional quantitative support to summarize this comparison, such as dispersion, offshore transport, and trajectory of the center of mass.
Center of mass is shown in panels 4a and 4b, An additional figure quantitatively summarizing the primary result will be added and is attached. Following Simons et al, 2012. We calculated particle density distributions (PDD) for each model over time. Then the correlation coefficient between PDD’s. i.e. ROMSfloats to ROMSpath and ROMSFloats to LTRANS. The LTRANS correlation coefficient drops below .7 in 10 days. The ROMSPath stays above .9 for 25 days and stays above or around .7 for the remainder.
Simons, R. D., et al. (2013). "Model sensitivity and robustness in the estimation of larval transport: A study of particle tracking parameters." Journal of Marine Systems 119-120: 19-29.
Line 393-394: It is hard to see from this figure that the particles tended to be closer to shore. Is there a statistic you can use for comparison, such as the mean distance from shore between the two, or the mean water depth of particles to test the significance of this observation?
The center of mass of the simulation with stokes drift is approximating 9 km to the northwest (320 degrees) of the simulation without stokes drift. Additionally, in the Stokes case, 57% of the particles were in depths less than 50m, compared to 38% of the simulation without stokes drift.
Line 405: Where is the improvement in “efficiency” with ROMSPath relative to LTRANS demonstrated in the results section? Also, is the improved “accuracy” in relation to the native ROMS result? These two points should be clarified in the text of this summary.
Although measurements of efficiency depend on a number of factors, ROMSPath compute time for the simulations shown in Figure 4 were at least 20% faster than LTRANS. Under certain configurations ROMSPath was 400% faster than LTRANS
There are 2 comments I will respond to separately.
Thank you,
Eli Hunter
-
AC3: 'Reply on RC2 Contiutation', Elias Hunter, 06 Mar 2022
Continuation,
Lines 377-379: Comment on the implications, and how this compares across particle tracking models (note, this relates to the previous comment re: Line 340). For example, is there evidence to suggest that the increased dispersion apparent with including small scale hydrodynamics through nesting and turbulent parameterizations in ROMSPath simulations improves the accuracy of this model compared to other formulations?
A detailed skill assessment of ROMSPath is outside the scope of this study. Hence we were unable to measure the accuracy of the dispersion compared to observations. It is common practice for OTP models to include parameterizations for horizontal and vertical turbulence, (Van Sebille et al. (2018), North et al (2008)) and not novel to ROMSPath. Support for nested grids, however, is novel. The addition of the refinement grid increased dispersion in our simulations by 30% in simulations with no turbulence parameterized. And by ~10% for simulations with turbulence parameterized. Thus including the refinement grid impacts horizontal dispersion in a significant way. And should be considered in cases when smaller scale features are common, such as near the mouth of an estuary. We note this in the manuscript revision.
van Sebille, E., et al. (2018). "Lagrangian ocean analysis: Fundamentals and practices." Ocean Modelling 121: 49-75.
North, E., et al. (2008). "Vertical swimming behavior influences the dispersal of simulated oyster larvae in a coupled particle-tracking and hydrodynamic model of Chesapeake Bay." Marine Ecology Progress Series 359: 99-115.Summary (i.e. Line 403): Can the authors comment on the results and features of ROMSPath in the context of other OPT applications (e.g. Lines 39-41), and in the context of previous works?
Given the context of the manuscript (a description improvements/new features, to an existing OTP) the summary focuses on highlighting those differences i.e. LTRANS vs ROMSPath. We will include a few notes on the features ROMSPath provides as compared to OTP models more broadly. Such as nested grid support, advection on the eta/xi grid, and the inclusion of Stokes drift.
We will also comment on a few studies directly comparing online vs offline particle tracking for the same model run. e.g. Wagner et al. (2019) and Cassiani et al (2016). And comments on the importance of stokes drift. (Monismith and Fong(2004)) and (Pareja Roman et al. 2019)
Cassiani, M., et al. (2016). "The offline Lagrangian particle model FLEXPART–NorESM/CAM (v1): model description and comparisons with the online NorESM transport scheme and with the reference FLEXPART model." Geoscientific Model Development 9(11): 4029-4048.
Wagner, P., et al. (2019). "Can Lagrangian Tracking Simulate Tracer Spreading in a High-Resolution Ocean General Circulation Model?" Journal of Physical Oceanography 49(5): 1141-1157.
Monismith, S. G. and D. A. Fong (2004). "A note on the potential transport of scalars and organisms by surface waves." Limnology and Oceanography 49(4): 1214-1217.
Pareja‐Roman, L. F., et al. (2019). "Effects of Locally Generated Wind Waves on the Momentum Budget and Subtidal Exchange in a Coastal Plain Estuary." Journal of Geophysical Research: Oceans 124(2): 1005-1028.
Thank you,
Eli Hunter
Citation: https://doi.org/10.5194/gmd-2021-400-AC3
-
AC2: 'Reply on RC2', Elias Hunter, 01 Mar 2022
-
RC3: 'Comment on gmd-2021-400', Anonymous Referee #3, 04 Feb 2022
This manuscript describes the ROMSpath particle tracking code, which results from modifications to the LTRANS particle tracking code to improve the functionality and efficiency of the model. The main changes from LTRANS are to run directly on the ROMS circulation model grid, to allow particles to move between nested grids, to separate the time stepping for advection and turbulence, and to correct an error in the coding of the turbulence random walk. LTRANS has been used widely with ROMS for years, and the changes made in creating ROMSpath will only increase the utility and accessibility of the particle tracking code.
The changes to the code that are reported here are all sensible and clearly explained. The effects of the changes on the model output are illustrated with examples from a study of larval transport along the east coast of the U.S. and Delaware Bay. While this project motivated the changes that were made to the code, it’s not necessarily the most effective means for demonstrating the improvements. It would have been more informative to illustrate the results with simpler, idealized examples that are more easily diagnosed and transferrable to other applications. For example, the improvements in performance from splitting the advective and turbulence time steps will clearly provide model speed-up, but it’d be helpful to provide guidance on how much speedup users can expect for typical simulation parameters. Similarly, it is not surprising that using higher resolution grids will increase the resolved dispersion of particles, but it’d be useful to provide more context on how the increase in dispersion with nested grids compares with theoretical expectations. The examples with initially vertically uniform particle distribution illustrate how particle dispersion depends on having the random walk algorithm coded correctly, but I am left wondering whether the clustering of particles near the pycnocline in the ROMSpath case has a physical basis due to flow characteristics or is instead some residual error (that is nevertheless a big improvement on the LTRANS result). Among the stated aims are to “improve the model’s efficiency, accuracy, and generality” [47-48], so that end, it would enhance the presentation to provide more generalizable examples of how these code updates improve the model.
Some additional, specific comments are provided with line numbers in brackets.
[11] is “OPT” a commonly used acronym? It’s unfamiliar, and quick search did not turn up other instances of it. The added confusion to readers with creating a new acronym does not seem to be worth the savings in keystrokes or ink.
[22] Perhaps note in the abstract that the manuscript provides examples of the how the improvements affect the performance of the code?
[41] “that calculate particle trajectories for a variety of applications” can be deleted.
[45] “It is not uncommon for users to modify OPT models to add novel processes for individual studies. Here, we describe alterations and additions to an existing OPT code, the Lagrangian TRANSport model (LTRANS), to add specific larval behaviour and improve the model’s efficiency, accuracy and generality.” These statements seem contradictory. If most users add their own processes and you are adding your own specific behavior, how does that improve generality? Please clarify.
[195] is the Stokes drift necessarily output at the same times as the ROMS fields?
[232] Do the details of the Doppio implementation on data assimilation and nudging matter for ROMSpath? If not, suggest removing for clarity.
[245] Similar to the previous comment, it’s not clear if the details on the time stepping are important for ROMSpath (e.g., recommended output interval) or specific to the goals of this science project. For this manuscript the focus should be on the former, and the latter would be more appropriate for a manuscript reporting on the scientific results.
[331] As with OPT, “CM” is unnecessary, and is more a source of confusion than clarity.
[334] “LTRANS OTP fails to reproduce the off-shelf transport” Why is that? What aspect of the code modifications led to this improvement?
[Fig 4] It's confusing to have the center of mass line on all 3 plots since the rest of the info in each panel is just a snapshot in time, whereas the line represents the trajectory over time. It’s also hard to distinguish the center of mass lines from the dots. Suggest removing the center of mass lines since, as noted in the text, it is not a particularly good metric as the particles get strained out.
[353] Why does decreasing the advective time step mitigate the clustering problem in LTRANS?
[354] “numeric[al] efficiency…tens of thousands of particles” It'd be worth quantifying the speedup in efficiency gained by splitting the turbulence and advection steps, assuming an appropriate ratio for them. Presumably it depends on how computationally expensive the advection and turbulence calculations are? Does it depend on the number of particles, or just become more noticeable with increasing numbers of particles?
[Fig 5] As mentioned above, ROMSPath also appears to have clustering near the turbulence minimum, but much less so. If LTRANS were run with the correction to the sign error in the code, would it give a result similar to ROMSPath, or are there other factors contributing to the difference?
[393] “wave swell was onshore during this time period” Isn't swell usually onshore, and increasingly so as it approaches the coast? Perhaps the idea is that the wave direction was aligned with main axis of the estuary?
Citation: https://doi.org/10.5194/gmd-2021-400-RC3 -
AC4: 'Reply on RC3', Elias Hunter, 08 Mar 2022
Referee 3,
Thank you for your comments on our manuscript titled “ROMSPath v1.0: Offline Particle Tracking for the Regional Ocean Modeling System (ROMS)”. We appreciate the thoughtful and useful comments and will address them point by point below.
“While this project motivated the changes that were made to the code, it’s not necessarily the most effective means for demonstrating the improvements. It would have been more informative to illustrate the results with simpler, idealized examples that are more easily diagnosed and transferrable to other applications. For example, the improvements in performance from splitting the advective and turbulence time steps will clearly provide model speed-up, but it’d be helpful to provide guidance on how much speedup users can expect for typical simulation parameters. Similarly, it is not surprising that using higher resolution grids will increase the resolved dispersion of particles, but it’d be useful to provide more context on how the increase in dispersion with nested grids compares with theoretical expectations.”
“Among the stated aims are to “improve the model’s efficiency, accuracy, and generality” [47-48], so that end, it would enhance the presentation to provide more generalizable examples of how these code updates improve the model. “
We agree that a broader analysis of the parameter space associated with the computational speed and model skill of ROMSPath, including a series of runs on idealized model grids for each model, would be valuable. However, as referee 3 notes, this work was motivated by a single project with specific scientific objectives. A larger study of type referee 3 recommends would be an independent investigation and is not realistic given the available resources and time. The improvements illustrated in the manuscript are significant enough that we believe publication is warranted.
“The examples with initially vertically uniform particle distribution illustrate how particle dispersion depends on having the random walk algorithm coded correctly, but I am left wondering whether the clustering of particles near the pycnocline in the ROMSpath case has a physical basis due to flow characteristics or is instead some residual error (that is nevertheless a big improvement on the LTRANS result). “
Given the 4-d nature of these simulations, it is unlikely the minor increase in particle density seen in figure 5 are due to residual error. As the particles spread in space the vertical distribution of particles at any given horizontal point is less uniform. Further, even in the canonical case in fig. 3 of Visser et al (1997), there are some random increases in particle density at the diffusivity minimum.
Visser, A. (1997). "Using random walk models to simulate the vertical distribution of particles in a turbulent water column." Marine Ecology Progress Series 158: 275-281.
[11] is “OPT” a commonly used acronym? It’s unfamiliar, and quick search did not turn up other instances of it. The added confusion to readers with creating a new acronym does not seem to be worth the savings in keystrokes or ink.
This acronym does not seem confusing to us. Offline and online particle tracking are terms used in existing literature (e.g. van Sebille et al, 2018) and it is natural to use an acronym for a phrase used repeatedly, such as “offline particle tracking”. However, to minimize any confusion we will emphasize the acronym in the introduction, switching (line 38) “referred to as offline particle tracking (OPT)” to “referred to as offline particle tracking (hereafter designated OPT for readability)”
van Sebille, E., et al. (2018). "Lagrangian ocean analysis: Fundamentals and practices." Ocean Modelling 121: 49-75.
[22] Perhaps note in the abstract that the manuscript provides examples of the how the improvements affect the performance of the code?
We will add a comment.
[41] “that calculate particle trajectories for a variety of applications” can be deleted.
Yes, we agree.
[45] “It is not uncommon for users to modify OPT models to add novel processes for individual studies. Here, we describe alterations and additions to an existing OPT code, the Lagrangian TRANSport model (LTRANS), to add specific larval behaviour and improve the model’s efficiency, accuracy and generality.” These statements seem contradictory. If most users add their own processes and you are adding your own specific behavior, how does that improve generality? Please clarify.
It is not that the larval behavior itself adds generality, but as part of our project we added specific larval behavior while also improving generality, for example by adding functionality for nested grids and stokes drift and wet/dry cells. We will clarify this in the text.
[195] is the Stokes drift necessarily output at the same times as the ROMS fields?
Yes, and the same spatial grid. It requires front end processing of stokes velocities into the correct format.
[232] Do the details of the Doppio implementation on data assimilation and nudging matter for ROMSpath? If not, suggest removing for clarity.
In as far as the hydrodynamics are being used as ROMSPath input, yes. We are using Lopez et al. (2020) as a primary reference for the DOPPIO hydrodynamic model setup. Lopez et al. (2020) did not use nudging and nesting, so we need to describe these differences.
López, A. G., et al. (2020). "Doppio – a ROMS (v3.6)-based circulation model for the Mid-Atlantic Bight and Gulf of Maine: configuration and comparison to integrated coastal observing network observations." Geoscientific Model Development 13(8): 3709-3729.
[245] Similar to the previous comment, it’s not clear if the details on the time stepping are important for ROMSpath (e.g., recommended output interval) or specific to the goals of this science project. For this manuscript the focus should be on the former, and the latter would be more appropriate for a manuscript reporting on the scientific results.
Time stepping details are very important for reproducibility. Typical ROMS output is saved hourly or 3 hourly, due to disk space constraints. We saved hydrodynamic data every 12 minutes and used that as input to ROMSPath.
[331] As with OPT, “CM” is unnecessary, and is more a source of confusion than clarity.
CM is not used often so does not warrant an acronym. We will update that.
[334] “LTRANS OTP fails to reproduce the off-shelf transport” Why is that? What aspect of the code modifications led to this improvement?
This is explained in section 4.1 (Coordinate system), describing the results comparing LTRANS, ROMSPath and ROMS floats. Changing ROMSPath to the ROMS eta/xi coordinate system reproduces off-shelf transport by ROMS floats. Whereas the coordinate system used by LTRANS does not, because of the error introduced in the LTRANS grid transformation.
[Fig 4] It's confusing to have the center of mass line on all 3 plots since the rest of the info in each panel is just a snapshot in time, whereas the line represents the trajectory over time. It’s also hard to distinguish the center of mass lines from the dots. Suggest removing the center of mass lines since, as noted in the text, it is not a particularly good metric as the particles get strained out.
While the center of mass is not the best metric, it is still useful to see where the center of mass paths diverge. It is informative to see these paths in figures 4a and 4b, although they are unnecessary in other panels.
[353] Why does decreasing the advective time step mitigate the clustering problem in LTRANS?
Good question, most likely it’s that the error introduced in the turbulence parameterization scales with the time step. So a larger timestep= larger error. We plan to illustrate this in a supplemental figure.
[354] “numeric[al] efficiency…tens of thousands of particles” It'd be worth quantifying the speedup in efficiency gained by splitting the turbulence and advection steps, assuming an appropriate ratio for them. Presumably it depends on how computationally expensive the advection and turbulence calculations are? Does it depend on the number of particles, or just become more noticeable with increasing numbers of particles?
The speed change is difficult to quantify. Depending on the configuration we saw speed changes from 20% increase in speed to a 400% increase. It depends on the system I/O speed as well as turbulence calculations. Yes, more particles translate to longer compute times.
[Fig 5] As mentioned above, ROMSPath also appears to have clustering near the turbulence minimum, but much less so. If LTRANS were run with the correction to the sign error in the code, would it give a result similar to ROMSPath, or are there other factors contributing to the difference?
If the sign correction was changed in LTRANS, split time-stepping is implemented to mitigate the clustering.
[393] “wave swell was onshore during this time period” Isn't swell usually onshore, and increasingly so as it approaches the coast? Perhaps the idea is that the wave direction was aligned with main axis of the estuary?
The wave field over time is variable over the width of the shelf. And not all of the domain of interest is in very shallow water. So it seem prudent to be specific about the direction of the waves for this test case.
Citation: https://doi.org/10.5194/gmd-2021-400-AC4
-
AC4: 'Reply on RC3', Elias Hunter, 08 Mar 2022