Title: Towards a Universal Electronic Nose: Algorithms for Designing & Optimizing Large Gas Sensor

Abstract: Gas sensing arrays—often referred to as electronic noses—consist of multiple sensing elements capable of collectively identifying complex gas environments. Such systems are essential for applications including toxic gas leak detection, food quality monitoring, and early medical diagnostics. Achieving high accuracy in these arrays requires selecting an optimal combination of sensing elements. However, current practices largely rely on empirical trial-and-error approaches, which become infeasible as the number of sensors and the complexity of gas mixtures increase.

Whereas prior research from our group focused on optimizing sensor arrays under conditions where brute force search is possible (i.e., finding the best 4-element array from a library of 50 elements, yielding 250,000 possibilities), this proposal aims to develop methods for finding optimal sensor arrays when there are over 10³⁵ possibilities (i.e., best 20-elements from a library of 500). At these scales prior optimizations methods are infeasible. We introduce two recommendation algorithms designed to identify the most effective set of sensors for a targeted gas mixture, along with a random selection strategy used as a baseline used to mimic the trial-and-error approach still in wide use. Whereas prior work in this area used molecular simulations of gas adsorption to predict sensor performance, such simulations would be too costly to use to explore this larger combinatorial space. To address this limitation, we further propose a computationally inexpensive data-generation algorithm capable of synthesizing artificial yet physically meaningful sensor–gas interaction data.

The proposed work has three aims. Aim 1, currently in progress, focuses on a reinforcement-learning-based recommendation algorithm that will be systematically evaluated against the other two proposed approaches. Future work (Aim 2) will extend this framework to determine sensor arrays that are optimal not for a single composition, but for broader regions within gas-mixture space. Finally, (Aim 3) explores grouping gases based on similarity such that a single sensor can reliably respond to an entire class of gases. This sensor-grouping strategy, inspired by mixture-of-experts architectures in contemporary generative AI, is expected to lay the foundation for a universal MOF-based gas sensing platform will be used to (1) conduct long-timescale molecular dynamics simulations to study product distributions of polyurethane pyrolysis at different temperatures (at constant volume) as a function of time; and (2) couple with kinetic Monte Carlo simulations to capture rare events and extend simulation timescales into the micro to millisecond regime.

Chair:

Dr. Christopher E. Wilmer

Department of Chemical and Petroleum Engineering, University of Pittsburgh

Committee Members:

Dr. Jason Shoemaker

Department of Chemical and Petroleum Engineering, University of Pittsburgh

Dr. Susan Fulllerton

Department of Chemical and Petroleum Engineering, University of Pittsburgh

Dr. Nikhil Bajaj

Department of Mechanical Engineering & Materials Science, University of Pittsburgh