Dr. Ranjit Chiplunkar

Data-Driven Probabilistic Modeling for Design and Monitoring of Process Systems.

Chemical and biochemical processes often operate under significant uncertainty arising from limited mechanistic knowledge, noisy data, and evolving process behavior. In this talk, I present two distinct but related research efforts that apply probabilistic modeling to different challenges in process systems engineering. 

The first part of the talk focuses on process development, where identifying suitable operating conditions is a central challenge. In this context, design spaces define admissible operating regions that meet predefined performance and feasibility criteria. In many practical settings, mechanistic models required to identify such regions are incomplete or unavailable, and experimental data are costly, limited, and exhibit variability. For such systems, I present a data-driven probabilistic design space identification framework based on Gaussian process surrogate models to characterize feasible operating regions under uncertainty. The estimated design space is then used to guide experimental design through a Delaunay triangulation–based strategy, enabling efficient selection of informative experiments and progressive refinement of the feasible region. The approach is demonstrated using benchmark numerical examples and a batch reactor case study. 

The second part of the talk focuses on process monitoring, where the objective is to track system behavior over time and detect changes arising from degradation, wear, or other non-stationary effects. In many industrial settings, such changes are not directly observable and must be inferred from process measurements through latent variables. I present a state-space modeling framework in which non-stationary behavior arises from a monotonically evolving dynamic latent variable modeled as a closed skew-normal random walk, while stationary process behavior is captured using Gaussian state-space components. The latent variables and model parameters are jointly estimated using the expectation–maximization algorithm, leading to a skew-normal filtering and smoothing problem. The approach is validated through numerical simulations and applied to a fouling monitoring problem in a hot lime softener.

Together, these two case studies illustrate how probabilistic modeling can support different stages of the process lifecycle, from process development to process monitoring.