Lin et describes and evaluates a forecasting system for predicting 3-D thermal structure of Lake Erie.  The manuscript is an evaluation of a true forecasting system (i.e., it is evaluating a set of forecasts of the future rather than mimicking the forecasting process with historical data).  It uses forecasted meteorology from Environment Canada Global Deterministic Forecast System to drive the model.  The model used (AEM3D) had previously been applied to Lake Erie so the novelty of the paper is using the model with forecasted meteorology.  I really appreciated the discussion sections on how the forecasting system could potentially be used to anticipate critical events for decision makers, emergency managers, and users of the lake.  I visited the website for the project and it is indeed up-to-date.  The paper evaluates the performance of the forecasts for a three-month period of time, although the manuscript discusses forecasts from outside this period.

The authors highlight the automation provided by a Python script but this seems to lack novelty (thought it appears to get the job done).  The model is a standardly used model, downloading one meteorology source and converting to another is straightforward, and using a task manager to run a job is routine. The application to Lake Erie is new but the manuscript is introducing a named forecasting system (COASTLINES) with the only model development being a Python script used to execute the model and download observations (actually also Matlab scripts).  Further, the code for running COASTLINES is not provided. Overall, the novelty of COASTLINES beyond the application to Lake Erie needs to be better motivated in the introduction.

The manuscript only uses observations to evaluate the model.  It does not perform data assimilation as other forecasting system do.  The workflow highlights the automation of the evaluation using the observations but does not offer a way that the evaluation is used to improve the model or the forecasts.  Therefore, it isn’t clear why the automated evaluation is necessary.  It would be great to explore how a feedback between the evaluation and the forecasts can be developed. 

The forecasts lack a representation of uncertainty. Uncertainty in forecasts is increasingly the state-of-the-art.  There is reference to uncertainty in the figures but the manuscript does not describe how uncertainty is estimated.  At minimum, the discussion needs to address the lack of uncertainty in the forecast and explore how it might be included in the forecasts.

The argument for why data assimilation is not necessary could be stronger.  They argue that it would only potentially decrease the RMSE by ~half (0.7C) by citing another study that used data assimilation.  However, how does the reader know whether this is a meaningful magnitude?

Instead of putting the Python script in the Appendix, I recommend putting them in a repository like Zenado.  That would allow someone to use the scripts without having to cut and paste from the Appendix.  Furthermore, the manuscript highlights the Python code but also has a dependence on MatLab for foundational parts of the workflow (i.e., line 131-132 – the conversation of weather forecasts into the AEM3D input).  Therefore, it is a Python + Matlab + Windows Schedule workflow.  Additionally, the code that is provided is only for retrieving the observations.  The code to run the forecasting system is missing.  Overall, I do not think that the availability of code and model output meets GMD’s standards.

The term hindcast is used throughout the manuscript but is not defined.   It would be help to define exactly what a hindcast is.

Line 440 states “To facilitate further development of open-access predictive modelling systems”, which is a major oversell of COASTLINES as being in the group of open-access predictive modeling.  The hydrodynamic model requires a license and the code to run the forecasting system is not made available.    

Can the authors point to a manuscript that demonstrates that the re-start works or provide a figure that shows that a restarted run matches a run that was continuous (i.e., same length of simulation but without stopping and restarting)?  Some hydrodynamic models are designed to be run from a cold-start and have internal variables that are not saved – thus preventing a true restart.

The manuscript highlights the use of the Environment Canada Global Deterministic Forecast System (GDPS) product.  Could other freely available forecast model output be used?  What about NOAA’s Global Forecasting System or NOAA’s Global Ensemble Forecasting system?

Specific comments

Line 47:  Are there 1-D water temperature forecasting systems that can also provide context in the introduction and discussion?

Line 63: I recommend starting a new paragraph here

Line 95: Is there a specific version of the AEM3D that was used?  Without that the forecasting system is not reproducible.

Line 119: what is CFL =?

Line 156: What happens when there are run-time errors?

Line 184:  Is the restart file used to generate a 216-hr forecast since the first 24-hr have already been generated?

Line 189: The Windows Task Scheduler is another dependency of the forecasting system.  Can the forecasting system only run in a Windows environment?

Line 224: The sentence refers to the estimation of uncertainty in the model forecast but the manuscript lacks methods describing the uncertainty estimation process.

Line 260:  What forecast horizon do these numbers refer to?

Line 329:  Readily what?  (missing word)

Fig 2: What is Phenomena detection (Supervisor)?  It is not mentioned anywhere in the manuscript

Fig 2: The caption says “Daily Python workflow” but the text states that Matlab is also used.

General comments:

This study develops an operational forecast system with a three-dimensional lake hydrodynamic model and validated the performance of hindcast experiments in Lake Erie. To conduct a robust forecast, the study separated 24-h and 240-h forecast simulations, and update the restart file for 240-h forecast every day. The data retrieval, numerical simulation, and validation are automated. This study is a development of a forecast system rather than a new model, but it is still of importance from the perspective of research implementation to society. However, some information for long-term steady operation is lacking, and providing us with such information would be helpful to other models/system developers, which are listed in the following major comments.

Specific comments:

How did you set the initial value of the model? You mentioned that the model was ‘cold started’ on day 99 (line 116) and generates a restart file with a 24-h forecast simulation (line 185), but some other sentences imply that the model was initiated on another date (e.g., lines 288 and 299). I imagine that “initiated” means the model was just restarted in the consistent simulation (if so, only lead (forecast) time information is enough), but the authors are requested to clearly explain the system setup because such information is useful for other forecast systems without data assimilation.

What is the difference of the meteorological forcing data between 15km and 25km? If your system can work in time with 15-km data, I think you can focus on the results of 15-km data only because the difference of results seems to be negligible. Or is there any known problem with 15­-km data?

You showed one-year results, but I have an interest in the long-term stability of the system operation. Do you reset the initial condition every year?

2 Data and methods

2.2 Model description

Line 96: Which programming language is the AEM3D written in? You mentioned that the wrapper code is written in Python, but it has no advantage on computational efficiency. In addition, you pointed out that previous hydrodynamic models are difficult to apply to the hind- and forecast applications due to the computational cost in line 42, how did you solve the problem?

2.3 Model setup and meteorological forcing variables

Line133: I have three questions.

 (1) The mass balance of lakes is described in + Precipitation – Evaporation + Riverine inflow – Riverine outflow ± Groundwater infiltration/seepage. Did you consider the groundwater component? If not, explicitly describe that assumption.

(2) In addition, is Precipitation – Evaporation balanced in Lake Erie? This budget controls the seasonal variability in water level, but mass imbalance may cause a problem in long-term operation.

(3) Even if the riverine in/outflows are balanced, can you ignore the effect on the fluid velocity field near the inlets and outlets? If you can, add the reference.

3 Results

You showed the confidence shade in the figures of time series, how did you do conduct ensemble simulations? This question is related to the reliability of the system if it is operated as a warning system to society.

3.2.1 Lake surface temperature

Line 273: Those results are interesting; longer forecast time does not increase the error for surface temperature (thermodynamics) even with consistent bias according to Fig. 6 but does the error for water level (hydrodynamics) according to Fig. 4 and 5. Can you discuss the difference? Comparison between the model bias and forecast time would be helpful in this respect.

Line 280: Why could you conclude the underestimation is due to ignoring river inputs? The underestimation occurs on the east side of the inlet from the Detroit River. Can you have a consistent discussion between line 223 and here?

Fig. 8: What is the main reason for the consistent underestimation in lake surface temperature?

4 Discussion

4.1 Bias and uncertainty

If the system developed in this study focuses on the forecast of some critical events like coastal up-welling and storm surge as discussed in the Sects. 4.1 and 4.2, could you move this Sect. 4.1 after the Sect. 4.3? The current Sect. 4.1 discussed the mean RMSD and compared it with a previous study incorporating data assimilation, but the data assimilation corrects only the initial condition of state variables in a model; not boundary conditions including meteorological forcing data. On the other hand, some of the critical events are caused by extreme atmospheric conditions in my understanding. So, can you discuss the model uncertainty and further improvement separating into initial-value and boundary-value problem?

Technical corrections:

2.3 Model setup and meteorological forcing variables

Line 116: How did you set the initial condition for the water level?

Line 119: What is “which is CFL = (Hodges et al., 2000)?

Line 121: “and net longwave radiation” (the former one) -> “and downward longwave radiation”? Because net (downward -upward) longwave radiation is calculated within the model as you mentioned.

3.2.1 Lake surface temperature

Line 252: Lake “S”urface temperature -> Lake “s”urface temperature

Line 288: needs a punctuation after “day 251”.

4.2 Prediction of coastal up-welling for fishery and drinking water management

Fig. 11:

(1) Correct the caption of the colorbar (remains selected?), and the same problem happened in Fig. D1.

(2) Can you show the observation data?

4.3 Prediction of storm surge events for public safety

Fig. 12:

(1) Can you show the spatial distribution of the 24-h and 96-h forecast to compare with the time series in (d)?

(2) According to (a) to (c), the forecasted water level is highly dependent on forecast time, and can you show the relationship between water level and forecast time? (It does not seem to be saturated even if the latest data is used according to (a).