This article presents a review of the current practices for validating the terrestrial biogeochemistry in CMIP6 Earth system models. The authors use the literature to show that the terrestrial biogeochemistry is a major source of uncertainty in future climate projections, and that this uncertainty can be linked to model structure. They study how 11 modeling groups participating in the CMIP6 exercise validated the terrestrial biogeochemical cycles in their land surface models and their fully coupled Earth System Models. They analyze the different validations presented by the modeling groups in terms of number of variables validated, quantity represented (for instance GPP), spatial and temporal scales, reference dataset and statistical metrics used. They also present two community methods designed to validate land surface models (ILAMBv2.1) and Earth system models (ESMValTool2.0). They present a critique of the validation approaches and suggest ways forward: mainly developing a standard protocol for validation which could be based on the existing communal software packages ILAMBv2.1 and ESMValTool2.0

The article is very well written, very clear and very detailed. The authors read in detail the articles describing the validation of the CMIP6 models. They give a very detailed and informative description of the techniques used by each modelling group in Annex A. I have a few detailed comments (see below) and only two general comments:

• I agree that the use of communal validation software packages would be highly beneficial to the community. However, these should be used to help understand the behavior of the models and help improve them. But I would be reluctant to see them used as tools to select models allowed to participate in certain exercises. If this was the case it would contradict Lovenduski and Bonan’s recommendation to improve process understanding and observation accuracy instead of reducing model spread (line 378). The authors don’t suggest to use these validation softwares as selection tools but it might be implied by their remark about GCP on line 217
• The authors don’t mention at all the links between the terrestrial biogeochemical cycles and the hydrological cycle. They mention that the modelers should evaluate the response of the biogeochemical cycles to temperature or the effect of nutrient limitations but never mention the effect of moisture conditions. I think the evaluation of the coupling between water and biogeochemical cycles at the land surface is as much of a concern.

Detailed comments:

L94 I find the figure a bit difficult to understand. An example would help like GPP being the variable corresponding to the left-most bar (if I understand correctly).

Table 6: the authors choose to separate vegetation carbon and vegetation biomass although these two type of data are very much related. I guess this is due to a choice made by iLAMB and ESMValTool. But Saatchi et al for instance, although being a tropical biomass dataset also gives data in terms of aboveground and belowground vegetation carbon mass. I’d suggest adding a comment specifying that these 2 quantities are not independent.

L215: The problem with the quality criteria of the Global Carbon Project is that they neglect the uncertainties related to the atmospheric forcing. As shown by Lawrence et al, 2019 (and also shown in Figure 3 in this paper) these may have a pretty strong impact. GCP switched from the CRU-NCEP forcing to the GSWP3 forcing in 2019, and model results were very different. Modeling teams had to retune their models to fit the criteria. That doesn’t mean criteria shouldn’t be imposed, it just shows their limit.

L 249: I would argue that the soil moisture control on all the variables of the biogeochemical cycles is as crucial in a climate change perspective as temperature, if not more because more uncertain (for instance the future of peatland/wetland in the high latitudes)

L320-324: I totally agree that site-level evaluation is very important to really understand how a model behaves. This is partly because at site-level, there is usually much more information available: the type of vegetation, the type of soil, the presence of an aquifer, land-use practices (irrigation, multi-cropping) etc. The meteorological forcing is also much more precise. In a global simulation, this type of information comes from global dataset (for instance soil texture, depth to bedrock, vegetation type) or are calculated. And this adds a huge part of uncertainty not related to the model structure but to model dataset. I think it might be interesting for the readers to get that insight

L419: I probably miss something here but I don’t see the link between N and P monitoring and turnover time assessment

Technical comment

L23:  I am not a native speaker but I find this sentence strange: “growing season extension in cold-limited regions, enhanced heterotrophic respiration, and potentially others, as well as environmental”

L378: I believe it should be “precedence” instead of “president”

This paper does an excellent job of reviewing the land biogeochemistry subcomponent evaluation approaches conducted by different modeling groups for 16 CMIP6-generation land surface models, both in a coupled and uncoupled context. This is useful since finding all the individual land model evaluation papers can be challenging. While Table 1, reprinted from Arora et al. (2020), provides such references, it lists only 11 models, not the 16 models identified in Figure 1.  This paper would better serve the community with a table similar to Table 1 that includes all the models reviewed in the paper.

While the scope of this paper is to review only the land biogeochemistry subcomponents of land surface models, describing differences in  hydrology and radiation/energy subcomponents and summarizing the assessments of them would be beneficial. Even if the authors do not wish to expand the scope to include other interacting model subcomponents, I recommend adding more key model configuration characteristics to an updated Table 1 to provide additional context to the summary. For example, the table could identify which models employ a river transport scheme, which models have depth-resolved soil carbon, which have an explicit permafrost representation, which used a prognostic dynamic vegetation scheme, etc.

The authors discuss the use of community-developed model evaluation and benchmarking packages, ESMValTool and ILAMB, which are increasingly being adopted as a means of standardizing model evaluation metrics and observationally constrained reference datasets. The wide variety of variables, datasets, and metrics employed in assessing model performance by different modeling groups makes direct comparison across models challenging.

Section 4.1, Variable Choice, begins with, "A comprehensive validation of a process-based model should include all simulated interacting variables for which a reliable empirical reference is available." However, the discussion that follows includes only biogeochemistry variables and there is little acknowledgement of the important interactions with water and energy variables. It may be useful to add a sentence or two that explicitly mentions the interdependence of and need to co-evaluate carbon, water, and energy cycles.

The authors do not describe the difference between metrics useful for evaluating offline models versus those that can be used for fully coupled models. Fully coupled models should exhibit the same statistical variability over decadal time scales as indicated by observational data, but the observed timing of ENSO and other drivers of climate variability are not reproduced in fully coupled models. Thus, metrics that assess biases or RMSE of time series model output should not be used when evaluating fully coupled model output. This discussion of applicable approaches should likely be included in Section 4.3.

The citation to Lawrence and Bonan et al. (2019) in line 371 should likely be Bonan et al. (2019).

On line 378, "president" should be "presedence".

On line 426, "in junction" likely should be "in conjunction".

Developing a standard validation protocol for model intercomparison activities within CMIP would be useful, and it has been done to some extent for CMIP6 historical land and ocean model performance in comparison with corresponding CMIP5 models in the IPCC AR6 Working Group I report in Figure 5.22, currently accessible at https://www.ipcc.ch/report/ar6/wg1/downloads/report/IPCC_AR6_WGI_Chapter_05.pdf#page=214