Graduate Students Seminar

Speaker 1: Gargi Chaudhuri

Title

Observer design for Chemical Reaction Networks

Abstract

Chemical Reaction Networks (CRNs) modeling biochemical and industrial processes may exhibit complex non-linear dynamics, including multi-stability, oscillations and chaos, making the real-time estimation of unmeasured species concentrations a significant challenge. Our work presents a linear algebraic framework for the design of observers for CRNs.

Using the algebraic structure of the reaction stoichiometric compatibility classes, we introduce a decomposition of the state space into two subspaces. And we derive sufficient condition for the asymptotic stability of the error dynamics, reducing the complex non-linear estimation problem to the verification of a Lie algebraic property of derived system submatrices.

The efficacy of this observer is explicitly demonstrated through its application to some real world CRN models. And we show that our observer guarantees exponential convergence of the estimation error. This result highlights the capability of this proposed observer to robustly reconstruct hidden trajectories even in chaotic cases.

Speaker 2: Mesfin Haileyesus

Title

Time-Indexed Split Conformal Classification Based on Prediction Bands in Survival Analysis

Abstract

Uncertainty quantification for survival predictions remains challenging in the presence of censoring. While existing conformal approaches for survival analysis typically target event times or entire survival curves, such targets are often difficult to interpret and may require strong modeling assumptions. In this talk, I will present a proposed conformal prediction framework for constructing time-specific prediction bands for the binary survival indicator. Under independent censoring and mild regularity conditions on the censoring estimator, the resulting bands achieve asymptotically valid marginal coverage at each fixed time point, regardless of whether the assumed survival model is correctly specified. The proposed approach naturally induces time-indexed classification sets that distinguish confident survival, confident failure, and explicit regions of uncertainty. An application to the METABRIC breast cancer dataset illustrates how the resulting classification paths provide interpretable, subject-specific, and uncertainty aware survival predictions.