By Danielle Fretwell

The Francis College of Engineering, Department of Electrical & Computer Engineering, invites you to attend a Doctoral Dissertation defense by Travis Kessler on "Predictive Modeling and Statistical Evaluation of Chemical Properties Relevant to the Combustion of Alternative Fuels and Fuel Blends."

Candidate Name: Travis Kessler
Degree: Doctoral
Defense Date: Thursday, March 23, 2023
Time: 9:30 to 11:30 a.m.
Location: ETIC 445

Those interested in attending virtually should contact the student (travis_kessler@uml.edu) and committee advisor at least 24 hours prior to the defense to request access to the meeting.

Committee Members:

  • Advisor Prof. Vinod Vokkarane, Electrical & Computer Engineering, UMass Lowell
  • Assoc. Prof. John Hunter Mack, Mechanical Engineering, UMass Lowell
  • Prof. Yan Lou, Electrical & Computer Engineering, UMass Lowell
  • Assoc. Prof. Hsi-Wu Wong, Chemical Engineering, UMass Lowell

Brief Abstract:

This dissertation investigates the use of computational models and various statistical methods to predict and analyze numerous chemical properties relevant to the combustion of alternative fuels and fuel blends. Discovering alternative liquid fuel sources is paramount in reducing emissions from existing engine architectures while society transitions to cleaner methods of energy generation. Alternative fuel research has traditionally been constrained to trial-and-error approaches driven by chemical intuition; predictive models, specifically state-of-the-art machine learning models, offer methods to screen fuels/molecules prior to experimental testing, drastically reducing the amount of time required to find cleaner, more efficient fuels.

The primary intent of the present work is to illustrate a standard methodology for accurately predicting chemical properties, specifically combustion-related properties, using modern feature selection/dimensionality reduction and machine learning techniques. Trained models are used to predict relevant properties for molecules resulting from the catalytic upgrading of bio-oil produced from fast pyrolysis. Trained models are accurate to within 5% of experimental property values, indicating accurate predictions for products resulting from catalytic upgrading.

Additional studies carried out in the present work aim to (1) compare a variety of machine learning/regression models, namely feed-forward neural networks, graph neural networks, and other multivariate regression techniques, (2) investigate the role of outlier data points, specifically how predictive model accuracy is affected, and (3) highlight key statistical relationships between molecular structure and combustion-related properties, ultimately providing a better understanding of how a molecule’s structure affects its performance as a fuel or fuel additive in existing engine architectures.