1  Introduction

Many physical phenomena such as fluid flow, heat conduction and semiconductor growth include a diffusion process. The action of the diffusion is a continuous process that has an increasing effect with time and typically manifests itself as a spreading and smoothing of the original distribution. Let the full effect of a diffusion process be written in operator notation as follows:
K f = g
(1)
where f is the original distribution, K: X ® Y represents the operation of diffusion over the time interval and g is the final distribution.
Physical problems that include diffusion processes, such as those stated, can often be modelled by a parabolic partial differential equation of the following general form
u(x,t)

t
= D Ñ2 u(x,t) +   first directional derivatives of    u.
(2)
The first term on the right hand side of the equation is the diffusion term and the coefficient D ( ³ 0) represents the rate of diffusion. Note that D need not be constant, it may for example be a function of position or vary with u as in material-dependent diffusion, the latter resulting in a non-linear equation: K = K(f). Often the first term is written Ñ.(D Ñu), but by product differentiation the equation would again take the form of (2). In the terminology of equation (1), u(x,0) = f(x) and u(x,T) = g(x).
In general, equations of the form (2) that model physical phenomena cannot be solved analytically. Numerical methods such as finite difference and finite element methods have been developed for their solution, simulating the relevant physical process through stepping forward in time t and re-computing the function u at the each sucessive time level ( see ref. [10], [12], for example).
The inverse problem of solving diffusion equations backwards in time - determining the initial distribution f from final distribution g - is fundamental to much scientific enquiry. The classical example of inverse diffusion problems is the backward heat conduction problem in which the original temperature distribution in a material is to be determined from the distribution after an elapsed time. Such a problem is considred in Beck [1], although this book is more concerned with the inverse heat conduction problem, the estimation of the surface flux history of a heat conducting biody. The authors' interest in the inverse diffusion problem arises from their work in atomic mixing modelling [7],[8],[14] and the inverse problem of recovering the original material distribution [6]. In the forward solution of parabolic equations, it is generally found that the presence of the diffusion term (the first term on the right hand side of (2)) improves the stability properties of the numerical method as it tends to damp away the numerical oscillation that may otherwise arise. However, in the backward solution of equations of the form (2), it is the diffusion term that causes the difficulty: simply employing the finite difference or finite element methods with negative time steps is numerically unstable. The other terms on the right hand side of (2) simply translate the solution and their effects are easily reversible.
Inverse diffusion problems are examples of ill-posed problems. Such problems - where the conditions placed on an equation are not classical and do not determine a unique solution - are also termed Cauchy problems (see Hadamard [4]). For diffusion problems, the smoothing that is inherent in the diffusion process results in a loss of resolution in the details of the original distribution. The inverse problem effectively suffers from insufficient information in g to determine a unique f. A further consideration is that in all practical applications the input to the inverse problem g will generally be measured data. It follows that g will be affected by noise and g could then easily fall outside of the set Y, the range of K, and there will then be no f such that K f=g, f Î X.
The inversion method can only succeed if the original inverse problem - which will generally have no unique solution - is replaced by one that always has a unique solution. The substitution of a nearby problem for the original inverse problem is often termed (or is equivalent to) the regularization of the operator K. The inversion method must converge to the solution of the nearby problem. In section 2 the general properties of the inverse problem and approaches to its solution are considered.
The purpose of this paper is to introduce an operator-splitting or defect correction method for the iterative solution of inverse diffusion problems. The convergence of the resultant method is analysed, particular in regard to observing Fourier solutions of the one-dimensional simple diffusion equation. The operator-splitting methods are applicable to general non-linear diffusion problems. Results from the application of the method to Fourier components of the simple diffusion equation to illustrate the points of the analysis. The method is also applied to problems arising from selecting a general original f on the simple diffusion equation and it is applied to a typical non-linear diffusion equation to illustrate the usefulness of the method.
An Excel spreadsheet demonstration of the inversion method can be downloaded free from the www.scientific-computing.info web site.