Reduction and identification methods for Markovian control systems, with application to thin film deposition

Abstract

SUMMARYDynamic models of nanometer-scale phenomena often require an explicit consideration of interactions among a large number of atoms or molecules. The corresponding mathematical representation may thus be high dimensional, nonlinear, and stochastic, incompatible with tools in nonlinear control theory that are designed for low-dimensional deterministic equations. We consider here a general class of probabilistic systems that are linear in the state, but whose input enters as a function multiplying the state vector. Model reduction is accomplished by grouping probabilities that evolve together, and truncating states that are unlikely to be accessed. An error bound for this reduction is also derived. A system identification approach that exploits the inherent linearity is then developed, which generates all coefficients in either a full or reduced model. These concepts are then extended to extremely high-dimensional systems, in which kinetic Monte Carlo (KMC) simulations provide the input-output data. This work was motivated by our interest in thin film deposition. We demonstrate the approaches developed in the paper on a KMC simulation of surface evolution during film growth, and use the reduced model to compute optimal temperature profiles that minimize surface roughness.

Effective transition rates for epitaxial growth using fast modulation

“…29 Other recent work by the authors describes a method for designing process parameters that vary more slowly, on the time scale of the overall process. 30,32 This approach has also been applied to germanium homoepitaxy by Gallivan and Atwater. 31 Thus, the averaging analysis presented here is not the only tool needed to compute process inputs, but instead describes one particular aspect.…”

Current challenges in the design and control of multiscale systems

“…Often, the information obtained from these analyses is used to build approximated models that can be later used to study the controllability of the multiscale process via simulations (e.g. Siettos et al, 2003b;Gallivan and Atwater, 2004;Gallivan and Murray, 2004;Armaou et al, 2005;Rusli et al, 2006). The key step in this approach is that the manipulated variables must have an effect on the key process variables at the fine scale.…”

Model Reduction Using Multiple Time Scales in Stochastic Gene Regulatory Networks

“…By expanding transformation S to include eigenvectors corresponding to these eigenvalues, we essentially treat them as if they were part ofṼ . This procedure adds robustness to the method with respect to separating fast and slow reactions in (8).…”

“…Since the two transcription factors bind at significantly different rates, a smart bookkeeping practice would be to record Lrp binding propensities in the matrix H and methylation propensities in V as defined in (8). With a convenient labeling scheme, as shown in Figure 4, we can express H in a simple block diagonal form,…”

“…Model reduction approaches based on singular perturbation theory have been used in various areas of engineering and science 5,6,7,8 , but the overwhelming complexity of some biological models has severely limited applicability of these approaches in some areas in biosciences. The Finite State Projection retains an important subset of the state space and projects the remaining part (which can be infinite) onto a single state, while keeping the approximation error strictly within prespecified accuracy.…”