(364m) Hyper-Sample-Efficient Optimization of Expensive Simulation-Based Models for Process Design Under Uncertainty

Many decision-making problems require global optimization of expensive-to-evaluate functions, in which there is a complex relationship between tunable parameters and the objective/goal. For example, pharmaceutical researchers are interested in designing drugs to fight diseases, and process engineers must synthesize flowsheets to achieve desired levels of profitability and yield. Under perfect knowledge of future conditions, one can straight-forwardly formulate a nominal optimization problem and obtain an optimal design. However, uncertainty is inevitably present in real-world problems, due to several sources including noisy and incomplete data, unknown or fluctuating model parameters, implementation errors and exogenous disturbances. Decision making under uncertainty is one of the biggest business challenges across industries including chemical, pharmaceutical, oil & gas, etc. because the so-called optimal solutions identified by solving a nominal optimization problem can be potentially suboptimal or even infeasible when implemented in practice. We utilize robust optimization (RO), a framework that aims to overcome this challenge of uncertainty, by recommending a robust optimal design that achieves the best worst-case value of objective while satisfying the constraints for all possible uncertainties. While there have been several ongoing works on development of RO algorithms, most existing methods are not applicable to high-fidelity simulation software which has quickly grown in scope due to advances in computing and increased availability of simulation software. Some examples include the simulation of: (i) multi-product plants that often requires proprietary (e.g., Aspen Plus) models of thermodynamics or kinetics, (ii) environmental models (e.g., CALPUFF, SEEMAC, iTree) that can be used to evaluate important sustainability metrics, and (iii) multiphysics phenomena (e.g., COMSOL) that is needed to represent important spatial variability and boundary effects. In this work, we develop a family of RO algorithms for expensive-to-evaluate high fidelity simulators with limited computational cost/time budget.

Contributions and Thesis Overview

1. We proposed constrained adversarially robust Bayesian optimization (CARBO), an algorithm that is well-suited for constrained robust min-max optimization problems whose objective and/ or constraint functions are defined in terms of noisy expensive black-box functions. We also introduced a novel recommendation criterion which allows a domain expert to choose the “best” design out of a set of sampled points likely closest to the true global solution of the constrained robust optimization problem
2. Flexibility analysis provides a quantitative framework for systematically identifying if a particular design of interest is feasible over a range of uncertainty values given the potential for recourse. Previous derivative-free approaches neglect recourse variable by removing the intermediate “min”, effectively solving a single level maximization problem, which provides a conservative solution. We proposed BoFlex, the first algorithm to the best of our knowledge that solves the “true” flexibility test problem in DFO settings, with convergence guarantees
3. One of the most recent contributions is BONSAI, which represents the objective as a directed acyclic graph (DAG) where each node can be either a black- or white-box function, enabling it to leverage intermediate information in the graph. We show order of magnitude improvement over existing derivative-free methods for robust optimization

Research Interests

These contributions can be practically implemented to a wide-variety simulation-based optimization of high-fidelity simulation software. The effectiveness of these techniques has been illustrated with several interesting case studies that encompasses many sectors including energy, robotics, pharmaceuticals, machine learning, and controller tuning to name a few. Apart from my thesis, I have had the opportunity to work on other projects related to global optimization of Gaussian processes, machine learning for additive manufacturing, robust decisions making for supply chain optimization (internship), model-predictive controller auto-tuning, and integrated design and control for technoeconomic synergy. I am seeking opportunities within a dynamic R&D organization where I can leverage my skills and experience. I am particularly interested in joining a team that values innovation, continuous learning, and the application of cutting-edge technologies.