04/30/2024
By Danielle Fretwell

The Francis College of Engineering, Department of Energy Engineering - Nuclear, invites you to attend a Doctoral Dissertation Proposal defense by Julian Eduardo Barra Otondo on "Inverse Design of Material Mixtures."

Candidate Name: Julian Eduardo Barra Otondo
Degree: Doctoral
Defense Date: Monday, May 6, 2024
Time: 9 to 11 a.m.
Location: Perry Hall 115, North Campus

Committee:

  • Advisor: Stephen Lam, Assistant Professor, Department of Chemical Engineering, University of Massachusetts Lowell
  • Fanglin Che, Assistant Professor, Department of Chemical Engineering, University of Massachusetts Lowell
  • Jerome Delhommelle, Associate Professor, Department of Chemistry, University of Massachusetts Lowell

Brief Abstract:
Material scientists are constantly in search of new materials with properties that could enable better solutions to engineering problems. A bottleneck in the discovery of new materials is the high cost involved in the synthesis and experimentation over an essentially infinite design space. The field of Material Informatics (MI), referring to the use of high-throughput computation to analyze material property databases, has evolved out of the necessity to evaluate material properties and their performance without having to incur in such expensive and time-consuming experimentation, with Artificial Intelligence (AI) and Machine Learning (ML) being recently incorporated into the field.

MI has opened the possibility of Inverse Design (ID) of materials, that is, taking the desired properties of a material as an input and generating a material which possesses these properties as an output. Inverse design can be achieved through several non-AI/ML methods, but the application of AI/ML methods has led to unprecedented accomplishments. ML property prediction models have allowed for high-throughput screening of materials according to their material properties, allowing to forego the costs involved both in the synthesis and experimentation over the wide compositional space of materials and in the simulation of these material properties through more computationally expensive methods like Density Functional Theory (DFT) and Molecular Dynamics. Many achievements have already been had by high-throughput screening implemented through ML, and these achievements have motivated research into further applications of AI/ML in materials science.

Newer approaches for ID through AI rely on the same method as the more known applications of AI: Generative Modeling (GM). By training specific Neural Network (NN) architectures on a dataset of material representations, these networks can generate new unseen materials. GM can be combined with the aforementioned property prediction models, both to perform high-throughput screening of the materials generated through GM, and also to incorporate them into the GM architecture to make models generate new materials with desirable values for a particular material property directly. The greatest successes of AI-enabled ID of materials have been achieved through implementing GMs in this manner, amongst which are the discovery of a new class of antibiotics and the prediction of protein structures. All of these successes demonstrate the effectiveness of AI/ML as a tool for ID, heralding continued research interest in applying these methods to further explore the design space.

While the aforementioned ID methods have been applied in the discovery of pure compounds both organic and inorganic, they have been scarcely applied in the discovery and optimization of material mixtures. This is specially the case with GM methods, and despite the success NNs have shown in the prediction of material properties in mixtures before. For this reason, the research proposed here would develop a workflow for the inverse design of material mixtures, with the ultimate goal of applying it in the optimization of molten salts. Ideally, the workflow will be implemented as a new GM architecture conjoined with a prediction model, so the materials can be used to generate the materials with desired properties directly, but difficulties involved in encoding the information of the materials might leave the option of high-throughput material screening, with the associated higher computational cost, as the only option. The design and optimization of material mixtures would allow for improvements in Generation IV nuclear reactors, Thermal Energy Storage, Molten Salt Batteries, in oxide fuel cells and solar cells, among other energy applications which are very relevant in the decarbonization of electricity grids so important in our age of climate crisis, and many other applications mixtures and alloys owe to their extremely tunable nature.