Section: Scientific Foundations
Modelling and simulating microbial ecosystems
Microbial ecosystems naturally put into play phenomena at different scales, from the individual level at a microscopic scale to the population level at a macroscopic scale, with sometimes intermediate levels. The size of substrate molecules is a thousand time smaller than the size of microorganisms and usually diffuse much faster. The substrate consumption of one microorganism is negligible at the population level but the sum of the consumption of its neighbors can modify the local concentration of substrate, which itself modifies microorganism growth, acting as a feedback loop. For other variables that change slowly (pH, temperature...) cumulative effects create intermediate time scales, coupling individual and environment dynamics. The very large populations justify macroscopic modelling but for some ecosystems, spatial structures seen at intermediate scale need to be tackled. This is typically the case of biofilm ecosystems, for which the biofilm structure is responsible of characteristics of the overall ecosystem. Models that are purely individualbased or purely populational are rarely truly satisfactory to incorporate current knowledge on microbial ecosystems at various scales and to push ahead mathematical analysis or to derive operational rules.
Macroscopic models
The starting point is the knowledge of biologists that report a large number of mechanisms discovered or shown on laboratory experiments at a population level, such as competition for a growthlimiting substrate, predation interactions, obligate mutualism or communication between bacteria. If each elementary mechanism is today well understood and modelled at a macroscopic level, the consideration of several mechanisms together in a single model is still raising several questions of understanding and prediction. This is typically the case when there is more than one growthlimiting substrate in the chemostat model or when one couples species competition with a spatial structure (flocculation, niches...).

Nonspatial models.
Ordinary differential equations (ODE) are the common way to describe the evolution of the size or concentration of species populations and their functional contribution in resource transformation (such as substrate degradation) in homogeneous or perfectly mixed compartments (or ecological niches). The wellknown chemostat model used in microbiology for single strain:
$\begin{array}{ccc}\dot{s}\hfill & =\hfill & \frac{1}{y}\mu \left(s\right)b+D({s}_{in}s)\hfill \\ \dot{b}\hfill & =\hfill & \mu \left(s\right)bDb\hfill \end{array}$(where $s$ and $b$ stand respectively for the substrate and biomass concentrations), has to be extended to cope with the specificity of microbial ecosystems in the following directions.

very large number (hundreds or thousands) of species. This leads to characterization of their distribution during the transients, that is a way to study the functional redundancy of ecosystems.

environmental fluctuations (input flow rate, input concentration, temperature, pH...). This impacts the efficiency of a microbial ecosystem, when biological and environmental time scales are different. Singular perturbations is the technique we use to separate slow variables from fast ones, leading to approximations of the dynamics on slow manifolds to be determined and analyzed.

interactions due to several limited resources and trophic chains. Most of the literature on the chemostat considers models with single limited resource, while some work studied purely essential or substitutable resources.

several populations of bacteria (for each species) to describe the effects of certain spatial structures that are artificially created in bioreactors or naturally found in soils, like flocks, colonies or biofilms: the planktonic (or free) cells and the biofilm (or fixed) biomass (for telluric ecosystems, such a distinction is also relevant to represent the sticking/non sticking characteristics of soil). Considering simple models of aggregates (that are not spatialized) can provide a simplified model of the dynamics of the overall biomass.

active and dormant bacteria. This distinction is motivated by the observations made on ecosystems of sparse resources such as arid soils.


Spatial models.
In the spirit of lattice differential equations, representations in terms of networks of (abstract) interconnected bioreactors propose an intermediate level between models of average biomass (a single ODE) and a continuous representation of space (PDE). A model of interconnected bioreactors is a way to implicitly take into account spatial heterogeneity, without requiring a precise knowledge of it. It is similar to the island models used in ecology but coupled with the dynamics of abiotic resources and hydrodynamics laws (transport, percolation, diffusion) governing the transfers between patches. This approach appears to be relevant for telluric ecosystems, for which pedologists report that microbial activities in soil are usually concentrated in hotspots that could be seen as small bioreactors. Understanding the role of the topology of the interconnection network and how a spatial structure impacts the outputs is also relevant in biotechnology to improve the yield or stability of processes.
Microscopic models
In these models (birth and death processes, neutral models, individualbased models) the dynamic of the population is described in terms of discrete events: birth and death of individuals, or jumps in terms of biomass. These models can be gathered under the same framework that could be called Markov stochastic processes with discrete events. Most of the time they should be coupled with continuous components like the size of each individual or the dynamic of the resources (represented in terms of ODE or PDE).
The Markovian framework allows on the one hand sharp analyses and rescaling techniques; on the other hand it induces a simplification in the memory structure that is important in terms of simulation. Indeed, as the future state of the system depends from the past only through the present state, only the current state should be kept in memory for simulation.
We will consider three families of processes with discrete events, from simplest to most complex.

Birth and death processes.
These models are of first importance in small population size. They indeed allow investigation of neartoextinction situations in a more realistic ways than the classical ODE models: they permit the computation, analytically but most of the time numerically, the distribution of extinction time and the probability of extinction. Efforts should be made to developed efficient Monte Carlo simulation procedures and approximation techniques for extinction probability and time distribution evaluation. In larger population sizes, they are advantageously approximated by diffusion models (see next section).

The neutral models.
In neutral models sizes of different species evolve as birth and death processes with immigration: all individuals have the same characteristics and are not spatialized. Such hypotheses could be considered unrealistic from a purely biological perspective, but these models focus on some precise properties to be simulated and predicted (for instance the biodiversity).
Comparing the prediction of species abundance of these models to real observations provides a way to justify or invalidate the neutral hypothesis. Extensions of the neutral model, that was originally introduced for forest ecology, have to be developed in order to better suit the framework of microbial ecology, such as the non constant size of the populations and spatialized variations.

The individual based models.
IBM's appear to be well suited to describe colonies or biofilms: in addition to birth, death and movement events, one has to consider aggregation and detachment events. The mechanisms that lead to the emergence of spatial patterns of colonies, or the formation of biofilms, which adhere to surface via polymers generated by the bacteria under specific hydrodynamics conditions, are not well understood yet. Typically, one can consider that bacteria inside the aggregates are disadvantaged to access the nutrient.
IBM modelling is a convenient way to propose aggregation and detachment mechanisms at the individual level in terms of random events connected to the geometry of the neighborhood, and to compare generated images with microscopic observations (for instance the confocal microscopy).
One has to be aware that few methods are available to study systematically and rigorously the properties of IBM, contrary to models based on differential equations (ODE, PDE...).
Bridges between models
The “theory of a computational model”, that combine two kinds of models (typically ODE and IBM) that are different representations of the same objects, relies on two steps: the “program making” and the “theoretical study”, in the spirit of the double modelling approach (roughly speaking, it consists in grasping the complexity of a IBM by analyzing accurately the consequences of each hypothesis on the macroscopic behavior of the model, building an approximate model of its global dynamics). Two main tools can be considered.

Change of scale.
For IBM models (neutral or Markovian), we consider mean field and moments approximation techniques that provide information at the macroscopic (i.e. populational) level, to be compared with macroscopic models. From a birthanddeath process describing the individual level, a renormalisation can provide a stochastic differential equation at a mesoscale. The diffusion approximation technique can be understood as a numerical acceleration technique where the number of births and deaths follows a normal law. These stochastic models at mesoscale can provide additional information compared to deterministic models at a macroscale, such as parameter identifiability or finite time extinction. The price to pay is to give much more conceptual and numerical efforts, that become less relevant for very large populations.
For PDE models on spatial domains described with regular patterns (such as models of biofilm), the homogenization technique allows to obtain simpler PDE with constant parameters.

The multiscale modelling.
The spatial heterogeneity in microbial ecosystems require to consider simultaneously several scales:
 
a physical scale. In batch processes, nutrient diffusion can be modelled by adapting the heat equation with Dirichlet boundary conditions. In continuous reactors, a convectiondiffusion equation with Neumann boundary conditions is considered instead, the speed vector field being provided by the equations of fluid mechanics. The spatial scale used for the discretization is given by diffusion and flow parameters.
 
a biological scale, given by the size and mobility of bacteria. Usually, this scale is larger than the physical one (at least in the liquid phase).
 
an aggregation scale of colonies or biofilms, even larger, that provides the spatial patterns.
Theoretically, it is always possible to describe all the processes at the smaller common scale and then use particle methods for a global representation, but this leads to extremely long computation times. The challenge is to manage these overlapping scales together and guarantee the stability of the numerical schemes. This is the goal of the multiscale approaches. For microbial ecosystems, it consists in

proposing new representations of the various scales of aggregation of bacteria in a model, taking into account the attachmentdetachment processes determined by the local hydrodynamics conditions. Here, discussions with specialists of fluid mechanics are required.

coupling diversity models (e.g. models based on the neutral assumption) with spatial models (that reproduce the patterns observed on images of microscopy) to better understand the link biodiversity/structure.

introducing new control variables, considered as independent variables, each of them describing a proper scale. For this purpose, we investigate different techniques available to determine such variables: meanfield approximation, singular perturbations, unification by limiting layers or

renormalising, that aims at detecting invariants among models of different scales.
 