## Abstract

In this paper, we address the identifiability of constitutive parameters of passive or active micro-swimmers. We first present a general framework for describing fibres or micro-swimmers using a bead-model description. Using a kinematic constraint formulation to describe fibres, flagellum or cilia, we find explicit linear relationship between elastic constitutive parameters and generalized velocities from computing contact forces. This linear formulation then permits one to address explicitly identifiability conditions and solve for parameter identification. We show that both active forcing and passive parameters are both identifiable independently but not simultaneously. We also provide unbiased estimators for generalized elastic parameters in the presence of Langevin-like forcing with Gaussian noise using a Bayesian approach. These theoretical results are illustrated in various configurations showing the efficiency of the proposed approach for direct parameter identification. The convergence of the proposed estimators is successfully tested numerically.

## 1. Introduction

Using kinematics to reveal internal properties of deformable objects moving in fluids is a difficult and long-standing topic in mechanics, e.g. [1]. One can, for example, think about the motion of a deformable cell in shear flow where various dynamical modes are observed depending on elastic capillary number [2, 3], internal properties of red-blood cells [4] or capsules [5, 6] in a creeping flow. Inferring internal mechanical properties has indeed already being pursued for flexible capsules from their deformation within a flow in a number of contributions [7–9] where the combination of experimental measurements and direct simulations can lead to estimating some internal parameters of the membrane. In this case, the constitutive laws of the membrane deformation are generally nonlinear but they are supposed homogeneous. In this study, we restrict our interest to flexible filaments in flows at low Reynolds number, where the hydrodynamics is linear [10]. Passive and active fibres [11–17], flagellum or cilia [18–21] have also been analysed in order to infer the internal parameters of their mechanical modelling. Gadêlha *et al.* [19] have used flagellum buckling experiments [20] to infer passive properties, including the passive sliding resistance, of a flagellum. Riedel-Kruse *et al.* [21] have fitted beating shapes with model prediction using a linearized one-dimensional wave propagation model for the beating propagation along an active flagellum in order to obtain either passive or active parameters. In the cited contributions, the shape is obtained from numerically solving a coupled fluid–structure problem, with mechanical properties being given. Internal parameter are then obtained from a systematic exploration of parameter space so as to reach a satisfying shape compared with experiments.

Whether such empirical parameter identification might be sufficient to provide interesting and insightful estimation of fibre or membrane mechanical properties could be a matter of debate. In many cases, a very precise estimation might not be so crucial and the mechanical properties might be approximately known from other experimental observations (e.g. osmotic swelling [5]).

Nevertheless, one key question which ought to be known is the uniqueness of the shape–motion/parameter relation. More specifically, are the shape and motion of a flexible object unique when varying internal parameters? This is obviously a crucial question as a non-unicity scenario would possibly lead to wrong estimation for the internal parameters. To our knowledge, in the case of nonlinear constitutive law for a flexible object (having large deformation) in a Stokes flow, this is an open question. This question is known as the identifiability condition for internal parameters. Even in the case of *linear* constitutive laws, but arbitrary deformation, identifiability conditions for deformable membrane or fibres are not known. The goal of this contribution is to clarify internal parameter identification in the simplest case of elastic flexible fibres in Stokes flow. Our aim is (i) to provide clear conditions for which the internal elastic properties of a passive flexible fibre or an active micro-swimmer can be identified from the observation of their kinematics within the framework of a bead-model (BM) description, (ii) to propose an operational formulation for reliable identification with and without additional noise, and (iii) to validate this operational formulation in various configurations in order to illustrate its interest. As the elasto-hydrodynamic fluid–structure interaction problem associated with (passive or active) flexible objects is *a priori* nonlinear, there is not much use of Stokes linearity for parameter identification. However, when a discrete description of deformable filaments by a collection of rigid bodies is considered to describe complex flexible objects namely BMs in Stokes flows [11, 16, 22] some simplifications arise. When considering a collection of non-overlapping beads we show that the internal forces linearly depend on passive and active forcing, producing a linear elasto-hydrodynamic problem. However, the obtained results are general for extensible BMs with suitable linearizations, or with pre-averaging, for which these models can be recast as Gauss–Markov processes and would provide equally valid likelihoods.

A detailed description of BMs is presented in §2a while focusing on kinematic constraints associated with the fibre inextensibility. More precisely §2b(ii) describes how contact forces are found from using Lagrangian multipliers, so that internal parameters and kinematic observations are linearly related. This point opens many issues that we subsequently analyse and illustrate for parameter identification in §2b. First, we provide clear and simple identifiability conditions associated with active or passive internal parameters showing that identifying concomitantly the active and passive internal properties of micro-swimmers is not possible. By contrast, the identification of mechanical properties assuming the active part is known, or reciprocally, finding the active part knowing the passive properties is possible, and can be addressed with the direct inversion of a simple linear problem without the need for guess values for the parameters. Section 3 analyses identifiability conditions and parameter estimation either in the deterministic case in §3a or the stochastic one in §3b. Identification in the stochastic case is possible using a Bayesian approach (as in [23]) and provides unbiased estimators the convergence of which can be estimated. Finally, §4 provides the numerical test and validation of the approach in various configuration associated with passive in §4a,b, actuated in §4c or active fibres in §4d.

## 2. Models and parameters

### (a) Object discretization with the bead model

#### (i) Bead models.

To model the locomotion of a micro-swimmer from describing its structural, active and elastic properties, we hereby use a BM. This class of models is interesting as they permit describing a complex deformable object as a flexible assembly of simple rigid ones, such as

#### (ii) The gear model.

Recently, a linear formulation of kinematic constraints using Lagrange multiplier formalism has been proposed for BMs: the ‘gear bead model’ (GBM) [29]. In this model, there is neither gap nor repulsive force between the beads to allow bending and no need for numerical parameters to be tuned. Kinematic conditions are associated with rolling contact connections at each contact point between neighboured beads, providing constraints for the local velocity and angular momentum of each bead. For example, at the contact point *i* and *a* is the bead radius, *i*. Also, quantitative agreement with previous models has been obtained for both slender objects (passive fibre, actuated filaments) and non-slender swimmers (*Caenorhabditis elegans*), allowing its use in a wide variety of contexts [29].

### (b) Theoretical framework of parameter estimation

#### (i) Bead model formulation and context.

We consider the following quantities to describe flexible objects of complex shape; *b*’ for bending or ‘*a*’ for active) are also added when needed for the sake of clarity so as to distinguish those torques from tangent vectors. Generalized velocities *J* symbol), and might stand for a generic relation for linearized constraints. *t*, positions *contact* forces. These *contact* forces are given by
*λ* is the Lagrange multipliers vector. The entire framework presented in the paper is valid for both two- and three-dimensional dynamic configurations.

#### (ii) Contact forces and linearity of the kinematic constraints of bead model.

Hydrodynamic interactions are provided through the solution of the mobility problem which relates forces and torques to the velocity and angular velocity of the beads. This formalism is particularly suitable for low Reynolds number flows for which total forces and torques on each bead are zero (as the flow is inertia-less). The total generalized force, i.e. the sum of generalized hydrodynamic forces

Hence, the contact forces can be found explicitly and depend linearly on other forces *effective* mobility matrix taking into account the action of contact forces associated with kinematic constraints. Also, note that if those contact forces (as would be the case for extensible fibres) are zero and kinematic constraints are also zero

In §4, we illustrate the formulation in the case of GBM associated with kinematic constraints with sliding contacts for which in most cases (except in §4c) actuation vector

In the case (most often encountered) where the elastic forces have linear dependence with constitutive parameters

Relation (2.13) is a major result of this paper as it gives an explicit, algebraic formulation for the linearity between kinematics and internal parameters. As kinematics is supposed to be experimentally known, this formulation allows for a direct determination of the identifiability conditions since, in the linear case, this is a straightforward issue.

## 3. Identifiability conditions

### (a) Deterministic case

Let us suppose that generalized velocity *effective* mobility matrix

#### (i) Identification of passive elastic parameters.

Let us start with the first case, suppose

Equivalently, if one considers the kinematic observable, the minimization of

In both cases, a direct

#### (ii) Identification of active forces.

Identifying active forces knowing passive ones is trivial in force formulation, for which the functional minimization

#### (iii) Identifiability of both active force and passive elastic parameters.

Let us consider a first ‘naive’ approach for concomitantly identifying both active and passive forces over a single period of time. In this case, we can formulate parameter identification as the solution of the following minimization problem for the force formulation
*t* (a continuous formulation is also possible but adds unnecessary complexity), and where it is meaningful to suppose some periodicity in the active component, so that *i*, e.g. *d* a continuous

#### (iv) Identifiability of periodic active forcing.

However, if the active component is strictly periodic, instead of trying to identify at various discrete time step (phase) *T*-periodic. One can then take the difference between the first and the second line of (3.14) to eliminate the active force so as to identify passive properties using the ‘force formulation’ (3.3). To be explicit, denoting
*T*. Hence, even if this section has shown that both active and passive components are not identifiable at the same time, one can nevertheless find strategies to circumvent the indeterminacy in the special case of a periodic active force within a variable flow, not having the same period.

### (b) Stochastic case

In the stochastic framework, we keep with the kinematic formulation as the noise is added onto the kinematics. Hence, let us consider the Langevin-like equations for the bead motion
*k* of vector

In a stochastic framework,

In the following, as we found in the deterministic case that

#### (i) Identification of passive elastic parameters.

In this subsection, we consider (3.17) with *N* independent observations are
*N* increases.

#### (ii) Identification of active forces.

In this subsection, we suppose that passive elastic properties are known and estimated. We consider (3.17) with *t* with period *T*. As we are now searching for a time-periodic parameter vector, it is not wise anymore to consider *any* discrete set of time *t* is sampled *N* times along period *T*, at time *T*-periodic functions. In practice, *T* can be found from analysing the time variations of *T* is not critical as it will only affect the evaluation of *T* would be mandatory, which might depend on the number of period sampling of

In the following, we thus keep with a continuous time representation of

## 4. Validation tests

In this section, we illustrate through various examples the identification results obtained in the previous section. For conciseness, we focus our numerical tests to kinematic formulation. Nevertheless, in most cases we numerically evaluate both formulations without notable difference. Here, we consider a quiescent fluid (*s* of the fibre, the normal vector *i*th component of a bending rigidity elastic vector *i* and bead *i* is given by

Finally, it is important to mention that for these numerical illustrations, we use a Rotne–Prager mobility matrix (cf. [29] for more details) and a third-order Adam–Bashforth scheme for time integration.

### (a) Elastic relaxation of passive fibres without noise

In this example, we consider elastic relaxation towards equilibrium of a single filament discretized into

Our aim is to estimate the bending rigidity *i*, i.e. the vector *s* to show that our method works in the general case of bodies with heterogeneous properties. We found in §3(i) that the vector *t*,

### (b) Settling passive fibres without noise

Here we estimate the bending stiffness of a flexible filament settling under gravity force

### (c) Actuated passive filaments without noise

In order to test if the proposed formulation also permits identification for actuated filaments, we propagate a bending from one end along the passive elastic filament, similarly to [40–43]. To simulate the experiment described in [42, 43] (attached filament actuated by its base), we use actuated prescribed three-dimensional motion of the first two beads, so-called tethered base elements [29]. This actuation requires the addition of three vectorial kinematic constraints to the no-slip conditions, adding

Here, we just write the matrix formulation of the model and the parameter estimation. We note

### (d) Active filaments without noise

Locomotion of the nematode *C. elegans* is used here as in [44]. Here we are just interested in modelling the motion of the nematode using the framework of the BM. To do so, we use an oscillating driving torque *i* results from the difference in active bending moments across neighbouring links

Table 5 provides the results for the active amplitude wave at a given time step. The results are obtained from the direct inversion of relation (3.9), while solving

### (e) Elastic relaxation with noise

In this section, we test the estimator prediction in the stochastic case. We explore the elastic relaxation already investigated in §4a in the presence of noise. The main purpose of this section is to test the theoretical predictions (3.23) and test the convergence toward this prediction given by the deviation law (3.26). The deviation (3.26) provides a normal convergence with a rather complex covariance matrix which non-trivially depends on the number of independent observations *N*. Thus we compute the statistical quadratic error, when for any noisy configuration we estimate *N*. The relative error over *N* independent observations collected during the fibre relaxation have been averaged over 10 different statistical samples in order to compare results to the expected theoretical error (3.26). Both are in very good agreement, as shown in figure 4. In the presented case of large noise amplitude

This computation illustrates and provides consistent results with the prediction obtained in previous section for noisy measurements.

## 5. Conclusion and perspectives

We have achieved internal parameter identification of active or passive inextensible fibres experiencing elasto-hydrodynamic coupling with Stokes flows. The fibre dynamics is described by an assembly of beads. Using Stokes flow linearity and explicit computation of contact forces associated with kinematic constraints, we derived a linear relationship between kinematics and internal parameters. The relation involves an explicit combination of kinematic constraints (through the Jacobian matrix and/or the actuation vector), hydrodynamic coupling (through Mobility matrix), elastic behaviour (through internal constitutive matrix) and external flow. This linear relation has enabled us to provide identifiability conditions. We find that either active or passive parameter identification is possible but not both simultaneously. Nevertheless, in the case of a periodic active forcing, we propose two possible approaches to circumvent indeterminacy and provide successively both passive and active properties.

We also found that in each case (passive or active), the presence of noise does not lead to a ill-posed inversion, but a well-conditioned direct inversion problem. We derive theoretical convergent estimators of the parameters. We tested this convergence numerically for various configurations and found consistent results with theoretical predictions. In each case of passive, actuated or active fibre we provide a direct estimate for the internal parameters without need of an initial guess, at the very moderate cost of a small linear system inversion. For noisy data, we provide error estimations of the internal parameter versus evaluation number which are fully consistent with numerical test-cases.

These results have been derived in the case of linear dependence of constitutive laws with parameters, but in the general case of arbitrary deformations (not necessarily small) and possibly nonlinear dependence of constitutive laws with shape (e.g. elastic bending depends on local curvature which is a nonlinear function of shape). The hypothesis of constitutive laws linearity with parameters might be seen as a limitation of the proposed results. As a perspective, let us briefly touch on how part of the proposed approach might be useful to tackle nonlinear parameter estimation. Obviously, in the general case of nonlinear dependence of constitutive laws with parameters, most of the analytic explicit estimators will not apply anymore. Nevertheless, the BM approach might still be useful, if one is able to derive a discretization of these continuous constitutive laws. The main point is that part of the nonlinearity associated with the fluid–structure interaction issues can still be tackled by the contact force model. Indeed, the linear relation between general velocities and forces (2.9) is still valid for forces having nonlinear parameter dependence. From then, the inversion problems that we derived might be useful in various iterative schemes associated with nonlinear inversion problems (e.g. Newton methods). Hence, our derivation might also been seen as a first useful step to more complex parameter estimations.

To provide more comments on the physical and biological relevance of our result it is important to stress that we found that two independent evaluations of active and passive property might provide a complete inversion of the internal parameters of flagellum or cilia. Such independent evaluations are interesting because they can both be used under experimental conditions, provided a precise quantitative observation of the fibre kinematics. First, non-invasive, ‘on the fly’ observations of passive fibre deformations should permit to evaluate internal mechanical properties.

This approach should also appear fruitful to provide internal parameter estimation of active micro-swimmers in the future, if, by any mean, one is able to temporarily switch-off the active swimming forces (e.g. using ion-pumps inhibitors). Passive properties could first be evaluated while later on the active swimming force components could be estimated from switching-on active effects.

## Data accessibility

All parameters associated with the simulations are given in the tables.

## Author's contributions

F.P. has provided the theoretical background and writing of §§2 and 3 as well as appendix A. E.I.T. has performed the programming, testing and validation of §4 results, and also contributed to §3 and appendix A. B.D. and E.C. have helped in the programming of section 4 and careful reading, corrections, comments and modifications of the manuscript.

## Competing interests

We declare we have no competing interests.

## Funding

This work has been supported by the Agence Nationale pour la Recherche (ANR) under grant no. MOTIMO (ANR-11-MONU-009-01).

## Acknowledgments

The authors thanks Jérôme Fehrenbach, Alexandre Steyer and Florencio Balboa-Usabiaga for their enlightening discussions and comments.

## Appendix A

**(a) The internal constitutive matrix **

In the case of elastic relaxation, the bending moment reads
*i*, again, *i*, and *i*th component of bending rigidity elastic vector *i* with

Although the generalized elastic force, when the collection of the force

**(b) The internal constitutive matrix **

Now, in the case of swimming nematodes, the active bending moment components are written so an active, time-periodic curvature wave is prescribed on each bead

The active bending torque is given by

- Received June 27, 2016.
- Accepted December 12, 2016.

- © 2017 The Author(s)

Published by the Royal Society. All rights reserved.