(4ly) New Approaches to Surrogate Modeling Under Uncertainty and Adaptive Learning in Systems Engineering

Research Interests: uncertainty quantification, mathematical modeling, Bayesian machine learning, optimization

The development of surrogate models is key to advancements in process systems engineering for the design and management of complex systems and critical assets. This poster discusses my research vision in the context of three challenges confronting the identification and use of artificial intelligence in systems engineering.

Novel Algorithm for Robust Sparse Model Development for Process Systems

The relative ease of development, simulation and maintenance of data driven models compared to their first-principles counterparts has encouraged the continuous development of new algorithms with enhanced performance and capabilities. However, many of these approaches either result in highly complex models with no physical significance on the influence of the input variables on the outputs, or remain sensitive to noise that may be present in the training data. One popular approach with great potential for uncertainty quantification while providing opportunity to incorporate user belief and/or prior knowledge into the estimation process is the Bayesian approach to machine learning. However, the high computational cost associated with this approach makes it unsuitable for many online applications such as model-based control and optimization. Hence, we recently proposed an approximate Bayesian approach with much less computational cost where inferencing of model parameters is implemented in an Expectation Maximization (EM) algorithm framework simultaneously with uncertainty quantification. While the developed algorithm does not assume identical distribution of noise in the measured variables, possible noise correlations are also accounted for.

To achieve model interpretability, our approach seeks to approximate system nonlinearities by a set of nonlinear transformations of input variables and their interactive effects. The optimal subset of the resulting large family of basis functions is selected using a global optimization algorithm by employing Branch and Bound (B&B) algorithm with efficient pruning strategies. The resulting Bayesian Identification of Dynamic Sparse Algebraic Model (BIDSAM) framework has shown superior performance when compared to existing methods in terms of sparsity and prediction capability of resulting models.

Satisfaction of Mass, Energy and Other Physics-Constraints by Sparse Data-Driven Model

While it is desired that data driven models conserve mass and energy, many existing approaches that pursue the satisfaction of physics constraints by data-driven models are not generalizable as they require some detailed knowledge of the system such as partial differential equations describing the system. Here we propose two methods that utilize only boundary conditions to guarantee exact satisfaction of mass and energy constraints even when the training data is noisy and contains some bias. In the first method, a reconciliation step is introduced into the EM algorithm while the second approach exploits the unique model structure to impose a set of equality constraints that ensure that mass constraints are exactly satisfied. Each of the proposed approach also applies to simultaneous satisfaction of mass and energy conservation laws.

Hybrid Modeling for Industrial Boiler Health Monitoring and Adaptive Learning of Sparse Models

There is a growing interest in the synergistic coupling of first-principles models and their data-driven counterparts as this brings together the computational efficiency of the latter and the high fidelity of the physics-based model for enhanced performance. One limitation that may arise whenever data driven model is used for prediction in real operating systems such as industrial boilers is possible frequent retraining that may be required due to significant variations in the system behaviors as a result of time varying uncertainties in the system. To minimize such requirement, an adaptive learning technique is necessary.

The novel BIDSAM algorithms we have proposed have found utility in the development of a health monitoring tool for industrial boilers on hybridization with mechanistic model. In collaboration with Electric Power Research Institute (EPRI) and Southern Company, BIDSAM models are used for predictions of uncertainties in the convective heat transfer coefficients and spatial distribution of temperatures inside the boiler where measurements are difficult or impossible to obtain. The predictions are in turn used for evaluation of boiler tube health conditions by estimating degradations due to stress and corrosion. Data from both NGCC and coal fired power plants which are characterized with much noise due to unstable ash conditions have been used to validate the developed models. Future work includes the formulation and implementation of brain-inspired adaptive technique as discussed in the computational neuroscience literature where predictive coding is used to explain the efficient and adaptive learning procedure of the human brain. Furthermore, I have keen interest in the optimization, efficient operation, health monitoring and management of energy systems and critical assets such as gasifiers, electrolyzers and solid oxide fuel cells under uncertain and harsh operational conditions.

Teaching Interests:

I am well qualified to teach all core courses in the undergraduate and graduate programs in chemical engineering with special interest in teaching process control, statistics, design and transport phenomena. I am also excited to teach graduate courses in process optimization, thermodynamics, and optimal control. I also look forward to developing new courses that would be helpful to students in gaining hands-on experience in the use of computational tools like MATLAB and Python.