Diffusion Coefficient Identification

Updated 9 July 2026

Diffusion coefficient identification is the process of determining transport parameters from observations such as PDE solutions, concentration profiles, and molecular trajectories.
It employs a range of methods, including inverse PDE theory, regularization techniques, and statistical analyses, to recover scalar and matrix diffusion coefficients.
The approach integrates experimental data with numerical reconstruction, offering practical guidelines for handling degeneracy, noise, and computational stability.

Diffusion coefficient identification is the determination of a diffusion parameter from observations of transport, broadening, fluctuations, or solutions of governing equations. In the literature represented here, the unknown may be a scalar coefficient in a parabolic or elliptic PDE, a matrix-valued tensor, a time-dependent coefficient in a time-fractional diffusion equation, an effective diffusivity inferred from concentration-step relaxation, an MSD slope extracted from trajectories, or a diffusion matrix coupling several conserved charges (Fragnelli et al., 2013, Mondal et al., 2020, Altybay, 26 May 2025, Nguyen et al., 2022, Bullerjahn et al., 2020, Dey et al., 2024). The associated observations range from final-time or boundary data and integral overdetermination conditions to fluorescence profiles, molecular trajectories, correlated tracer motion, and current-current correlators (Cannarsa et al., 2021, Ashurov et al., 5 Aug 2025, Zareh et al., 2011, Seki et al., 2014).

1. Problem classes and observables

A common organizing principle is the distinction between coefficient identification in governing equations and parameter estimation from transport observables. In PDE inverse problems, the coefficient enters the operator itself, as in

$\partial_t y - \partial_x\!\bigl(u(x)\partial_x y\bigr)=0, \qquad \partial_t u-\nabla\cdot(q(x)\nabla u)=f, \qquad u_t-\nabla\cdot(A(x)\nabla u)=f,$

and the task is to reconstruct $u(x)$ , $q(x)$ , or $A(x)$ from state data (Fragnelli et al., 2013, Cen et al., 2023, Mondal et al., 2020). In experimental and stochastic settings, the coefficient is identified from an observable relation such as an error-function concentration profile, an MSD law, a current-current Kubo formula, or a hydrodynamic mobility integral (Nguyen et al., 2022, Bullerjahn et al., 2020, Dey et al., 2024, Seki et al., 2014).

Setting	Unknown	Observation or identification relation
Degenerate or standard parabolic/elliptic PDE	$u(x)$ , $q(x)$ , $A(x)$ , $a(t,x)$ , $p(t)$	interior data, boundary flux, final-time observation, or integral overdetermination (Fragnelli et al., 2013, Cannarsa et al., 2021, Mondal et al., 2020, Ashurov et al., 5 Aug 2025, Altybay, 26 May 2025)
Microfluidic concentration-step relaxation	$D$	$u(x)$ 0 and $u(x)$ 1 vs $u(x)$ 2 (Nguyen et al., 2022)
Trajectory-based inference	$u(x)$ 3 or $u(x)$ 4	MSD slope, GLS fit, equilibrium fluctuation-response bound, or kernel regression (Bullerjahn et al., 2020, Dechant, 2018, Hammer et al., 2019)
Optical broadening of images or carrier clouds	$u(x)$ 5 or $u(x)$ 6	Gaussian width growth and revised MSD models (Zareh et al., 2011, deQuilettes et al., 2020)
Coupled transport	$u(x)$ 7, $u(x)$ 8, or $u(x)$ 9	hydrodynamic mobility integrals or Kubo formulas (Seki et al., 2014, Dey et al., 2024)

This variety is not merely terminological. In some problems the object to be recovered is a constitutive coefficient field constrained by ellipticity, box bounds, or degeneracy; in others it is an effective transport parameter that summarizes the net outcome of diffusion, screening, recombination, or collision correlations. A recurring theme is that identifiability depends as much on the observation operator as on the forward model.

2. Continuous inverse theory for diffusion coefficients

A substantial part of the literature treats diffusion coefficient identification as an inverse problem for PDEs. In the strongly degenerate parabolic equation

$q(x)$ 0

the coefficient satisfies $q(x)$ 1, $q(x)$ 2, $q(x)$ 3 on $q(x)$ 4, and $q(x)$ 5. The coefficient is sought in an admissible set $q(x)$ 6 under box constraints, a slope bound, and fixed endpoint values. Existence of an optimal coefficient follows by the direct method in the calculus of variations, and first-order necessary conditions are expressed through an adjoint equation and a variational inequality. For a final-time observation with homogeneous Dirichlet-Neumann boundary conditions, the unique minimizer admits the explicit piecewise linear-saturation law

$q(x)$ 7

with $q(x)$ 8 determined by the intersections with $q(x)$ 9 (Fragnelli et al., 2013).

Degenerate diffusion identification also appears in the form $A(x)$ 0 in the one-dimensional parabolic PDE

$A(x)$ 1

Here one recovers either a constant $A(x)$ 2 or the power $A(x)$ 3 from interior snapshot data, and for the general-coefficient case both $A(x)$ 4 and $A(x)$ 5 are recovered from the one-sided boundary flux $A(x)$ 6. The theoretical results include uniqueness and Lipschitz stability in the constant and power-like cases, and uniqueness for the general coefficient via global Carleman estimates for a hyperbolic problem combined with an inversion of an integral transform similar to the Laplace transform (Cannarsa et al., 2021).

For matrix-valued coefficients, the forward model is

$A(x)$ 7

with $A(x)$ 8 symmetric a.e. and uniformly elliptic. The inverse problem is nonlinear in $A(x)$ 9, but a “natural linearisation” around a reference $u(x)$ 0 rewrites it as a linear ill-posed operator equation $u(x)$ 1 for $u(x)$ 2, where $u(x)$ 3 and $u(x)$ 4 is defined through an auxiliary parabolic problem. Uniqueness is proved under a snapshot-determinant assumption $u(x)$ 5 a.e. in $u(x)$ 6 (Mondal et al., 2020).

Integral overdetermination produces a different identifiability mechanism. In

$u(x)$ 7

the coefficient $u(x)$ 8 is recovered from

$u(x)$ 9

The analysis combines the Fourier method with a priori estimates and yields local and global weak solutions, as well as local and global strong solutions under additional smoothness (Ashurov et al., 5 Aug 2025). A related time-fractional setting,

$q(x)$ 0

uses the spatial average

$q(x)$ 1

to reconstruct $q(x)$ 2, again under $q(x)$ 3 (Altybay, 26 May 2025).

These formulations clarify that diffusion coefficient identification is not a single inverse problem but a family of problems whose well-posedness depends on degeneracy structure, admissible sets, additional observations, and the analytic device used to expose the coefficient.

3. Regularization, discretization, and computational reconstruction

In elliptic and parabolic coefficient recovery, the standard discrete framework is regularized output least squares. For $q(x)$ 4, one minimizes

$q(x)$ 5

with Galerkin finite elements for the state and, in the parabolic case, backward Euler in time. The fully discrete functionals $q(x)$ 6 and $q(x)$ 7 lead to weighted $q(x)$ 8 estimates and, under the positivity condition

$q(x)$ 9

to standard $A(x)$ 0 error bounds. The analysis gives practical parameter guidance such as $A(x)$ 1, $A(x)$ 2, and $A(x)$ 3 (Jin et al., 2020).

A later hybrid formulation retains the same output least-squares structure but replaces the finite-element coefficient ansatz by a neural-network prior. The coefficient is represented as

$A(x)$ 4

where $A(x)$ 5 is a feed-forward network with tanh activations and $A(x)$ 6 enforces the physical box constraint. The state is discretized by standard $A(x)$ 7-FEM and backward Euler, gradients are computed by solving an adjoint PDE and back-propagating through the network, and a Riesz map is used to convert the $A(x)$ 8-dual gradient into an $A(x)$ 9-gradient. The resulting error bounds depend explicitly on the noise level, regularization parameter, mesh size, time step size, and network complexity. Numerically, the hybrid approach is reported as consistently at least as accurate as the pure-FEM inversion and often significantly more robust to large noise; for example, in 1D elliptic with 10% noise the relative $a(t,x)$ 0-error is $a(t,x)$ 1 for hybrid vs $a(t,x)$ 2 for FEM, and in 2D elliptic with 10% noise $a(t,x)$ 3 vs $a(t,x)$ 4 (Cen et al., 2023).

For matrix coefficients, Tikhonov regularization is applied to the linearized equation

$a(t,x)$ 5

with normal equations $a(t,x)$ 6. The adjoint admits the PDE representation

$a(t,x)$ 7

where $a(t,x)$ 8 solves a backward adjoint problem. Galerkin approximation is implemented by orthogonal projections acting entrywise on matrix fields, and an adaptive rule for the regularizing parameter yields an order optimal rate of convergence under a spectral source condition. When only relaxed noisy data are available, Clement interpolation is used to construct a smoothed version that satisfies the required gradient regularity (Mondal et al., 2020).

Time-fractional identification employs a fully implicit finite-difference scheme. The Caputo derivative is approximated by the L1 formula

$a(t,x)$ 9

and the resulting tridiagonal system is strictly diagonally dominant. The scheme is unconditionally stable in $p(t)$ 0, and the convergence theorem gives

$p(t)$ 1

The inversion step uses an explicit discrete version of the integral identity for $p(t)$ 2, with one-sided finite differences for boundary derivatives and composite Simpson’s or trapezoid rules for the integrals (Altybay, 26 May 2025).

A plausible implication is that the dominant numerical distinction across these works is not between “analytic” and “data-driven” methods, but between formulations that preserve the forward operator structure and those that obscure it. The cited schemes all preserve that structure explicitly.

4. Experimental identification from concentration, image, and carrier profiles

One direct experimental route identifies diffusion coefficients from the relaxation of spatial profiles. In the microfluidic free interface diffusion technique, pure water and solution are brought into contact at $p(t)$ 3 in a dead-end microchannel, and under dilute conditions the one-dimensional mass transport is purely Fickian: $p(t)$ 4 For the step initial condition, the solution is the error-function profile

$p(t)$ 5

or the equivalent sign convention. Identification proceeds by fitting the normalized profile to

$p(t)$ 6

extracting $p(t)$ 7, and plotting $p(t)$ 8 versus $p(t)$ 9, whose slope is $D$ 0. The reported range is $D$ 1, with typical relative error in $D$ 2 of a few percent. The same experiments reveal diffusio-phoresis and diffusio-osmosis, but under the stated conditions the Taylor-dispersion correction satisfies $D$ 3 at most, and the overall impact on the solute diffusion measurement is negligible $D$ 4 in the explored dilute regime (Nguyen et al., 2022).

A closely related identification problem is single-image diffusion coefficient measurement for freely diffusing fluorescent proteins. The recorded image is modeled as a convolution $D$ 5 of an axial-projected PSF with a lateral pathway distribution function. For short exposure times both are approximately Gaussian, so the measured width obeys

$D$ 6

For the stated TIRF setup,

$D$ 7

which gives

$D$ 8

For eGFP at $D$ 9, the extracted single-image value is $u(x)$ 00, and with $u(x)$ 01 molecules per frame the uncertainty shrinks to $u(x)$ 02, matching typical FCS precision (Zareh et al., 2011).

In semiconductor optical microscopy, the standard MSD relation

$u(x)$ 03

is widely used for expanding carrier profiles, but it can significantly overestimate $u(x)$ 04 because second-order and third-order recombination artificially broaden the distribution. Revised MSD models add explicit corrections for bimolecular and Auger terms through integral expressions involving $u(x)$ 05, $u(x)$ 06, $u(x)$ 07, and $u(x)$ 08. The stated guidance is that for perovskite thin films the standard MSD model can produce large fitting errors at carrier densities $u(x)$ 09, and that the commonly deployed MSD models are not well-suited when feature sizes are comparable to the diffusion length and boundary behavior is unknown. Photon recycling is reported to impact energy carrier profiles only on ultrashort time scales or for materials with fast radiative decay times (deQuilettes et al., 2020).

These cases show that profile broadening is often the most accessible route to $u(x)$ 10, but also the easiest place to conflate diffusion with other transport or loss mechanisms. The literature repeatedly treats that conflation as the central experimental pitfall.

5. Trajectory-, fluctuation-, and simulation-based estimation

When trajectories are available, diffusion coefficient identification is usually cast as a statistical estimation problem. For molecular dynamics, the basic observable is the discrete-lag MSD,

$u(x)$ 11

fitted with

$u(x)$ 12

The key improvement is to account for the strong correlations between MSD values at different times by minimizing a GLS $u(x)$ 13 based on the covariance matrix $u(x)$ 14. The framework also introduces a quality factor

$u(x)$ 15

to choose an optimal sub-sampling interval, and a Kolmogorov-Smirnov test on endpoint displacements to assess whether the long-time dynamics is diffusive. In the cited applications, the method yields $u(x)$ 16 for TIP4P-D water and $u(x)$ 17 for ubiquitin (Bullerjahn et al., 2020).

A different trajectory-based route uses equilibrium fluctuations in a trap. For overdamped Langevin dynamics in a confining potential $u(x)$ 18, one defines the time-averaged position

$u(x)$ 19

and obtains the lower bound

$u(x)$ 20

For a harmonic trap $u(x)$ 21, the bound is asymptotically exact: $u(x)$ 22 Weak interactions and anharmonic corrections make the estimate less tight, but numerical tests show that $u(x)$ 23 remains close to one in several moderately perturbed cases (Dechant, 2018).

High-frequency observation of branching diffusions with immigration introduces a nonparametric identification problem for the one-particle diffusion coefficient $u(x)$ 24. Because particle identities are lost between snapshots, the method first reconstructs correct identities on “good” short intervals by nearest-neighbor matching under $u(x)$ 25-identifiability and well-spreadness assumptions. The recovered increments populate a local regression scheme on a cube $u(x)$ 26 where the invariant occupation density $u(x)$ 27 is strictly positive. In one dimension, with $u(x)$ 28 in a Hölder class $u(x)$ 29, a Nadaraya-Watson estimator based on $u(x)$ 30 achieves

$u(x)$ 31

or equivalently

$u(x)$ 32

described as the minimax-optimal rate for nonparametric estimation under mean-square loss (Hammer et al., 2019).

At the kinetic-theory end, gaseous diffusion can be identified from single-collision statistics. The correlated-random-walk formulation introduces the persistence ratio

$u(x)$ 33

and the exact mean-square free path factor

$u(x)$ 34

Using the near-geometric decay of the reduced collisional series, the improved formula becomes

$u(x)$ 35

For hard spheres, the improved random-walk formula lies within a few percent of DSMC data for all $u(x)$ 36, and for $u(x)$ 37 it even outperforms the first-Sonine approximation (Yuste et al., 2024).

A recurring misconception is that linear fitting of MSDs is intrinsically sufficient once enough lag times are available. The cited work argues the opposite: covariance structure, confinement, branching identity loss, and step-step correlations can all dominate the estimator design.

6. Coupled, multicomponent, and phenomenological diffusion parameters

Some identification problems concern transport coefficients that are coupled by geometry, friction, or conservation laws. In bilayer membranes, the bilayer is modeled as two incompressible 2D fluids with surface viscosities $u(x)$ 38, coupled by inter-leaflet friction $u(x)$ 39 and by solvents of viscosities $u(x)$ 40 and thicknesses $u(x)$ 41. For a rigid circular domain of radius $u(x)$ 42 in leaflet $u(x)$ 43, the diffusion coefficient is

$u(x)$ 44

Identification proceeds by fitting $u(x)$ 45 versus $u(x)$ 46 and, more robustly, by fitting longitudinal and transverse correlated diffusivities $u(x)$ 47 and $u(x)$ 48. In a symmetric membrane, the crossover from $u(x)$ 49 to $u(x)$ 50 is a factor of two, and the inflection in $u(x)$ 51 occurs when the leaflet-sliding length $u(x)$ 52 is comparable to the hydrodynamic screening length $u(x)$ 53. The work also shows that if $u(x)$ 54, the integral diverges as $u(x)$ 55: inter-leaflet friction alone does not regularize the two-dimensional Stokes divergence (Seki et al., 2014).

For several conserved charges, diffusion is described by a matrix $u(x)$ 56 rather than a scalar. With charge densities $u(x)$ 57, currents $u(x)$ 58, susceptibility matrix $u(x)$ 59, and conductivity matrix $u(x)$ 60, the constitutive law is

$u(x)$ 61

The Kubo formula gives

$u(x)$ 62

or equivalently the fully covariant expression in terms of $u(x)$ 63. In the weak-coupling quasiparticle limit, the Kubo result matches kinetic theory with the identification $u(x)$ 64 (Dey et al., 2024).

A more phenomenological use of diffusion identification appears in rapidity-space analyses of heavy-ion collisions. There, the measured object is the diffusion parameter $u(x)$ 65 defined through

$u(x)$ 66

with the interpretive relation

$u(x)$ 67

The paper emphasizes that $u(x)$ 68 and $u(x)$ 69 are not extracted separately; $u(x)$ 70 is the single fit parameter. In HIJING, the averaged values are $u(x)$ 71, $u(x)$ 72, $u(x)$ 73, and $u(x)$ 74; UrQMD gives similar values and the extracted $u(x)$ 75 is described as essentially independent of collision energies over $u(x)$ 76– $u(x)$ 77 GeV (Singh et al., 2020).

Taken together, these examples show that “diffusion coefficient identification” often means identifying a transport law under coupling constraints rather than estimating an isolated scalar $u(x)$ 78. This suggests that in many current applications the decisive issue is not the inverse step itself, but the correct constitutive level at which the coefficient is defined.