CMIP6 biogeochemical models under-represent the deep chlorophyll maximum and diverge in how they express it: a regime-stratified BGC-Argo evaluation in the North Pacific

Abstract. Surface chlorophyll, the variable most directly linked to satellite ocean colour, is widely used to evaluate marine biogeochemical models, yet it cannot by itself reveal how chlorophyll is distributed through the water column. Using Biogeochemical-Argo (BGC-Argo) chlorophyll profiles over the North Pacific, we evaluate the vertical chlorophyll structure of nine Coupled Model Intercomparison Project Phase 6 (CMIP6) biogeochemical models, and we ask how faithfully the models reproduce that structure - including where modelled and observed surface chlorophyll already agree. The observations show that the deep chlorophyll maximum (DCM; median depth 71 m) is not a single feature but a family of vertical structures: grouped by the observed DCM depth, the observed maximum both deepens and becomes more sharply expressed, from shallow and weakly expressed to deep and strongly expressed. Relative to these observations, the models allocate a smaller fraction of chlorophyll to the subsurface and express a weaker vertical chlorophyll contrast (DCM prominence) - the subsurface-fraction bias is negative and the prominence ratio below unity in all nine models across the full comparison - and they place the chlorophyll maximum increasingly too shallow as the observed DCM deepens, with a median DCM-depth bias of -9, -22, and -37 m across the shallow, intermediate, and deep observed-DCM regimes. The models further separate into distinct types of DCM representation - flat, shallow, and deep - none of which reaches the observed combination of DCM depth and prominence, so they differ in kind and not only in degree. Because the models place their maxima at different depths, the ensemble median profile washes the maximum out and is a poor summary of its own members; ensemble-level agreement in DCM depth would therefore reflect cancellation among per-model errors of opposing sign rather than per-model fidelity. These vertical-allocation biases persist even where modelled and observed surface chlorophyll agree: within the surface-matched subset the median subsurface chlorophyll fraction bias is about -5 to -6 percentage points with an interquartile range entirely below zero, and the DCM prominence ratio remains below unity (0.81). Surface chlorophyll agreement is therefore not sufficient to diagnose biases in vertical chlorophyll allocation, and surface-focused evaluation can leave systematic subsurface structural biases undetected. Diagnosing and improving the vertical distribution of chlorophyll in biogeochemical models benefits from depth-resolved constraints such as those provided by BGC-Argo.