Empirical Modeling of Nanoscale Dynamics using Solution Mapping

Computer simulations provide useful predictions of complex system dynamics, but they cannot be easily inverted for use in control and optimization. When the computational time to run a single prediction is high, approximate models with reduced computation are required. These models must be straightforward to build and have quantified bounds on their accuracy. With support from this grant two automated methods were developed for building empirical models from simulation data. These methods were subsequently applied to stochastic simulations of nanoscale dynamics. In the first method, the simulation dynamics are modeled on a discrete state space, with input-dependent transitions between the states. This approach was used for dynamic optimization of a gallium arsenide surface deposition process, which was not computationally feasible for the full simulation. The second method, based on Gaussian process modeling, was developed to further improve the prediction accuracy of the first method, which was limited by the discrete state space. Moreover, Gaussian process modeling enabled a quantification of the prediction variance, which is necessary so that the dynamic model can be used with confidence in control applications.