Repository navigation
Initialisation Concepts
Initialisation is not a minor technical step in OSMOSE. It is the process by which the model is given a starting ecological state from which all subsequent dynamics emerge. Because OSMOSE is stochastic, school-based, size-structured, and multispecies, the initial state cannot be reduced to a single biomass number per species. It must define, explicitly or implicitly, how biomass is distributed across ages, sizes, schools, and space, and how those elements enter the modelled life cycle from the first simulated time step.
This section introduces the main conceptual ideas needed to understand the implemented initialisation methods described later in the document. Its purpose is not to document any particular parameter, but to clarify what an initial condition means in OSMOSE, why it matters, and why practical initialisation methods are needed.
In any dynamic model, the initial condition is the state of the system at the beginning of the simulation. Once the model is started, all later states depend jointly on the model structure, the parameter values, the external forcing, and this initial state. Two runs with identical equations, parameters, and forcing can therefore produce different trajectories if they start from different initial conditions.
In principle, the initial condition should describe all state variables that are needed to start the simulation. In simple population models, this may mean specifying only abundance or biomass. In more structured models, it may also require the distribution of the population across ages, sizes, stages, or space. In ecosystem models, the problem becomes broader still, because several populations interact simultaneously and their initial states must be specified in a mutually coherent way.
For this reason, initialisation should be understood as part of the model definition rather than as a purely technical pre-processing step. A poor or inconsistent initial condition can generate unrealistic early transients, delay convergence towards a meaningful simulation regime, or even direct the system towards qualitatively different trajectories.
{#fig:initial-conditions-matter fig-align="center"}
The terminology used around model initialisation is not always fully consistent across fields. In ocean and climate modelling, some terms have long-established technical meanings [@ouranosNdInitializing; @thornton2005spinup], whereas in ecosystem modelling they are often used more loosely or in a more application-dependent way. In this document, the following working definitions are adopted and interpreted in the context of ecosystem models such as OSMOSE.
Initial condition refers to the state of the model at simulation time zero [@ouranosNdInitializing; @zhang2005enso]. In OSMOSE, this is not limited to total biomass by species. It includes, explicitly or implicitly, the demographic and ecological structure required to start the simulation, such as biomass distribution across age or size, the number of schools, trophic level, and, depending on the method, spatial position.
Initialisation refers to the process of choosing, constructing, or reconstructing that initial condition. Depending on the application, these starting values may come from observations, climatologies, previous model runs, simplified assumptions, or derived state estimates [@ouranosNdInitializing; @seferian2016cmip5]. In ecosystem models, this may also involve reconstructing a structured population state from partial biological and observational information rather than specifying every model entity directly.
Transient dynamics refer to the model behaviour that arises because the simulation begins from a particular initial condition that is not yet fully adjusted to the model structure, parameterisation, and forcing [@hewitt2016spinup; @seferian2016cmip5]. In ecosystem models, these transients may involve changes in abundance, demographic structure, trophic interactions, or spatial occupation as the system reorganises from its imposed starting state.
Spin-up refers to the adjustment phase during which the model evolves away from the imposed initial state and towards a regime that is dynamically consistent with its equations, parameter values, and forcing [@hewitt2016spinup; @ouranosNdInitializing; @thornton2005spinup]. In ecosystem models, spin-up often involves the establishment of a coherent age or size structure, trophic organisation, and reproductive regime. The duration of spin-up depends on the variables of interest and on how close the initial state is to the regime implied by the model.
Equilibrium should not be understood here as a perfectly constant state. In ecosystem models, it is more useful to define equilibrium as a regime in which the long-term tendencies of the variables relevant to the analysis have become sufficiently small for the intended purpose. Shorter-timescale variability may still remain [@thornton2005spinup; @seferian2016cmip5].
Statistical equilibrium is the appropriate concept when the model remains variable through time. In that case, what matters is not that every variable becomes constant, but that the mean behaviour, variance, or broader distribution of the simulated state becomes sufficiently stable on the timescales relevant to the analysis [@thornton2005spinup; @ouranosNdInitializing].
Quasi-equilibrium refers to a practical rather than absolute notion of equilibrium. A model can be considered to be in quasi-equilibrium when the remaining drift is small enough that it no longer compromises the scientific objective of the simulation, even if some slow adjustment is still present [@seferian2016cmip5; @lavingullon2023timeslices].
Drift refers to the residual systematic change that remains because the model has not fully adjusted, or because some part of the imposed starting state, forcing, or coupled structure is not fully consistent with the model’s own preferred regime [@hewitt2016spinup; @seferian2016cmip5]. In ecosystem models, drift may appear as gradual changes in biomass, age structure, trophic organisation, or other slow state variables.
Memory of the initial condition refers to the extent to which the simulated state still reflects the imposed starting configuration. This memory is variable-specific. Fast components may lose it quickly, whereas slow demographic, trophic, spatial, or biogeochemical structures may retain it for much longer [@ouranosNdInitializing; @seferian2016cmip5; @lavingullon2023timeslices].
Attractor is used here in a broad dynamical sense to denote the regime towards which the model tends under a given set of equations, parameters, and forcing. In practice, this may correspond to a stable equilibrium, a fluctuating statistical regime, or, in some cases, collapse. In this sense, saying that a model “forgets” its initial condition is only meaningful if different initial conditions converge towards the same attractor [@thornton2005spinup; @seferian2016cmip5].
Adequate spin-up is therefore variable-specific and question-specific. A simulation is adequately spun up when the remaining influence of the initial condition is small enough for the variables, timescales, and processes relevant to the problem being studied [@thornton2005spinup; @seferian2016cmip5; @lavingullon2023timeslices].
These definitions are especially useful for distinguishing steady-state and interannual applications. In a steady-state application, the initial condition mainly serves to place the model within reach of the regime of interest, and spin-up is a central part of the workflow. In an interannual simulation, by contrast, the early transient dynamics are often themselves part of the target of the analysis. In that case, a satisfactory initial condition is one that already captures the relevant demographic and ecological structure at the start of the simulation [@ouranosNdInitializing; @seferian2016cmip5].
The initialisation problem is especially important in OSMOSE because the model does not represent populations as continuous deterministic state variables. Instead, focal species are represented as collections of schools, and each school carries its own state information, including species identity, abundance, age, length, weight, trophic level, and position.
As a result, the initial state of a population in OSMOSE is not just a collection of biomass values for every species. It is a distribution of biomass among many discrete entities, the super-individuals or schools, each of which will undergo growth, mortality, movement, predation, reproduction, and fishing as the simulation unfolds. This means that the same total biomass can correspond to very different effective initial states, depending on how that biomass is allocated among sizes, ages, schools, and locations.
Stochasticity reinforces this point. In a stochastic model, the realised trajectory depends not only on expected quantities but also on the specific configuration of the simulated entities. In OSMOSE, the initial condition is therefore better understood as a structured representation of a population state, not as a single deterministic configuration. This is one of the main reasons why initialisation matters so strongly in practice.
The need for initialisation is closely related to, but not identical with, the idea of spin-up. In modelling practice, a spin-up period is a preliminary simulation phase during which the model is allowed to adjust from its starting state towards a more internally consistent regime under the chosen forcing and parameters [@hewitt2016spinup; @thornton2005spinup].
Spin-up is often useful because the initial condition supplied by the user is rarely a perfect representation of the dynamic state implied by the model. Even when biomass or demographic structure is approximately correct, the detailed balance between growth, mortality, reproduction, predation, and fishing may require some time to stabilise. Early simulation years can therefore reflect adjustment to the initial condition as much as the processes of real interest.
However, spin-up does not remove the importance of initialisation. The duration and behaviour of the spin-up depend on the starting state, and different initial conditions can lead to different transients and, in some cases, to different long-term regimes. Starting closer to a realistic internal state may shorten the adjustment period and improve the usefulness of the early simulation years. Starting from an inconsistent or poorly structured state may instead generate long or misleading transients.
In other words, spin-up is not a substitute for initialisation. It is a way to let the model adjust from the initial state that has been chosen.
{#fig:init-spinup-regime fig-align="center"}
The distinction between initialisation and spin-up is especially important when comparing simulation objectives. In a steady-state application, the main purpose of the initial condition may simply be to place the system within reach of the regime of interest, and spin-up is then a central part of the workflow. In an interannual simulation, by contrast, the early transient dynamics are often part of the target of the analysis. In that case, a satisfactory initial condition is one that already captures the relevant demographic and ecological structure at the beginning of the simulation, so that the early trajectory reflects the intended system dynamics rather than only adjustment from an artificial starting state.
A common intuition in modelling is that, after a sufficiently long spin-up, the model “forgets” its initial condition. In practice, this intuition is only partly true. Some aspects of the starting state may indeed fade with time, but the speed and extent of that forgetting depend on the structure of the model and on the nature of the processes involved [@seferian2016cmip5; @lavingullon2023timeslices].
In OSMOSE, the initial condition affects the early demographic structure, the first reproduction events, the onset of predation interactions, and the school composition of the system. These effects can influence the transient response of the model over several years. Even when the system approaches a statistical equilibrium or quasi-equilibrium, that approach may depend on the initial configuration of schools and cohorts.
Different initial conditions may alter not only the duration of the transient adjustment, but also the equilibrium or stabilised regime eventually reached by the simulation. If one initial condition leads to persistence and another to collapse, then the collapse cannot meaningfully be described as a case in which the model has forgotten the initial state. Rather, the initial condition has selected a qualitatively different long-term outcome.
{#fig:init-different-equilibria fig-align="center"}
This point is especially relevant when the model is calibrated or interpreted over a limited time window. If the early dynamics of the simulation are strongly shaped by the initial condition, then the apparent fit to observations or the apparent response to forcing may partly reflect initialisation choices rather than only the core parameterisation. A technically valid initial condition is therefore not automatically a satisfactory one. What matters is whether it produces a coherent ecological starting state relative to the objectives of the simulation.
At a conceptual level, initialisation in OSMOSE must provide enough information to define the starting state of each focal population in a school-based representation. This involves several distinct components.
First, each population must be given a total initial biomass or an equivalent amount of population material from which abundance can be inferred. Without this, there is no starting population to simulate.
Second, this biomass must be distributed across demographic structure. In practice, that means allocating the population across age and size classes, or directly across schools already characterised by age and size. This is essential because growth, mortality, maturity, predation vulnerability, and fishing selectivity all depend on age or size.
Third, the initial state must define the number of schools used to represent the population. Since OSMOSE operates on schools rather than on a continuous biomass density, the way biomass is partitioned among schools is part of the initial condition itself.
Fourth, trophic level must be assigned, at least in a provisional way, so that schools enter the simulation with a defined trophic state. Even if trophic level is not the main driver of demographic dynamics, it is part of the state variables tracked by the model.
Fifth, some representation of spatial distribution is needed. In direct school-based initialisation this is explicit, because each school is given spatial coordinates. In reconstructed population-level initialisation the spatial component is more indirect, but the initial state must still ultimately be compatible with the spatial structure of OSMOSE.
These components show why initialisation in OSMOSE cannot be reduced to “setting an initial biomass”. The model requires a structured initial population state.
Although the conceptual requirements of initialisation are broad, users rarely have direct empirical access to a complete school-level state at the beginning of a simulation. In most applications, they do not know the exact abundance, age, size, trophic level, and position of every school that should exist in the model domain at the initial time.
Instead, users typically have partial information: total biomass estimates, survey indices, yield series, catch-at-length data, growth and maturity parameters, or restart files from previous simulations. The purpose of the initialisation module is therefore to bridge the gap between the ecological state that the model requires and the information that users actually have.
Different methods achieve this in different ways. Direct school-based initialisation uses an explicit school file when such a file already exists. Reconstructed population-level initialisation uses species-level information to infer a plausible class-based starting state, which the OSMOSE core simulator later expands into schools. Egg seeding initialisation uses a reproduction-based mechanism to let the model generate early cohorts when no mature population is initially available.
All these methods respond to the same practical problem: the full initial state required by OSMOSE is rarely observed directly, so it must either be supplied from a previous model state, reconstructed from partial information, or allowed to emerge indirectly through model dynamics.
The conceptual initial state described above is not represented in a single unique format in OSMOSE 4.4. Instead, the model supports different implemented input structures corresponding to different initialisation strategies.
In direct school-based initialisation, the implemented input is a restart-like netCDF file containing one record per school. In reconstructed population-level initialisation, the implemented input is a set of population.initialization.* parameters that describe biomass and structure at class level, from which the OSMOSE core simulator creates schools at runtime. In egg seeding initialisation, the implemented input is neither a school table nor a class-based initial population block, but a set of seeding controls that allow the early population to emerge through reproduction.
This distinction between conceptual initial state and implemented input structure is important. Different methods may ultimately aim to represent the same underlying ecological idea—a starting population distributed across sizes, ages, schools, and space—while using different technical encodings to do so. The role of the following sections is therefore to explain how OSMOSE 4.4 moves from these conceptual requirements to the concrete methods and parameters that users must configure.
OSMOSE 4.4 includes several routes to define the initial state of focal species at the beginning of a simulation. These routes differ in the way the initial population is described, in the amount of information they require from the user, and in the degree to which the initial state is provided directly or reconstructed from other inputs.
At the broadest level, the initialisation module supports two main types of workflow. In one case, the initial state is provided directly as a school-by-school description. In the other, the initial state is reconstructed first at population level and only later translated into schools. In addition, OSMOSE retains an older seeding-based route in which the model is allowed to generate the initial population dynamically over an initial period.
These methods coexist in OSMOSE 4.4 because they reflect different historical stages in the development of the model and different practical needs. Some methods are more explicit and direct, while others are more practical when the user has standard biological parameterisation and a limited set of observational inputs but no full restart-like initial state.
For clarity, this document distinguishes between the following conceptual methods:
- Reconstructed population-level initialisation
- Direct school-level initialisation
- Egg seeding initialisation
In addition, the current codebase includes a conversion route that derives a population-level initialisation block from restart outputs. This route is best understood as a bridge between school-level and population-level workflows rather than as a fully separate conceptual method.
The three main initialisation families differ in what they require from the user, what they produce, and what type of question they are best suited to answer.
The reconstructed population-level initialisation requires a standard species parameterisation together with initialisation-relevant observational inputs, such as biomass, yield, and, when available, catch-at-length. It produces a population-level initialisation block, which is then expanded into schools by the core simulation engine. Its main advantage is practical usability: it allows a coherent initial state to be generated without requiring an explicit school file. Its main limitation is that the reconstruction is approximate and depends on the quality and consistency of the supplied biological and observational inputs.
Direct school-level initialisation requires a restart-like netCDF file describing the initial schools explicitly. It produces no intermediate reconstruction, because the file itself is already the direct initial state. Its main advantage is precision and transparency at school level. Its main limitation is that preparing such a file can be technically demanding, especially outside restart or coupling contexts.
Egg seeding initialisation requires only the configuration of a seeding duration and seeding biomasses by species. It does not define the full initial state explicitly, but instead lets the model generate it over time. Its main advantage is conceptual simplicity in some special cases. Its main limitation is that the starting state is highly indirect and may require longer stabilisation before the model reaches a satisfactory demographic structure, but this is not guaranteed.
The restart-derived conversion route is best regarded as a utility workflow. It uses an already available school-level state as its starting point and produces a population-level parameter block that can later be reused. It is therefore especially useful for moving between initialisation representations.
For OSMOSE 4.4, reconstructed population-level initialisation should generally be regarded as the preferred default method. This preference is based less on theoretical elegance than on practical balance.
Compared with direct school-level initialisation, the population-level route is easier to set up in ordinary modelling workflows because it does not require the user to prepare a complete school-by-school initial state file. At the same time, compared with egg seeding, it offers a much more explicit and controllable representation of the starting population. The user can inspect the generated population.initialization.* block, verify the reconstructed biomass and demographic structure, and diagnose inconsistencies before running the simulation.
This method also integrates naturally with the regular OSMOSE parameterisation. Most of the biological parameters it uses are already part of the standard model configuration, which reduces duplication and improves internal consistency. Although the current implementation includes patches added to extend the method to more data-poor situations, its overall logic remains well aligned with ordinary OSMOSE parameterisation practice.
For these reasons, reconstructed population-level initialisation provides the most useful compromise between realism, flexibility, and practical usability in the current version of the model.
A recurring difficulty in the use of the OSMOSE initialisation module is the tendency to mix up the representation level of the initial state.
In school-level initialisation, the initial condition is the explicit list of schools. Each school already exists in the input file with its own species, position, abundance, age, length, weight, and trophic level. The file therefore defines the initial state directly.
In population-level initialisation, the initial condition is not a created from a list of schools but a species-level description of the population distributed across classes. The population.initialization.* parameters do not identify individual schools in advance. Instead, they describe how much biomass should be placed in each class, what age and trophic level should be associated with that class, and how many schools should be created to represent it. The actual schools that represent the initial state are then generated from this aggregated description by the OSMOSE core simulator.
This distinction is important in its own right, but in OSMOSE 4.4 it is further complicated by the existence of preparatory procedures that generate the population-level block from data or from restart outputs. As a result, users often encounter three different objects in a single workflow:
- observational and biological inputs used to reconstruct a population-level initial state;
- the generated
population.initialization.*block; - the schools ultimately created by the simulator.
These are three different representations of the initial state at different stages of the workflow, but the latter are formally the initial conditions of the model. The rest of this document is structured precisely to keep them separate: Section 4 focuses on reconstructed population-level initialisation, Section 5 on the other implemented methods, and Section 6 on the user-facing configuration and parameter documentation.