09/03/2025
By Danielle Fretwell

The Francis College of Engineering, Department of Energy Engineering - Nuclear, invites you to attend a Doctoral Dissertation defense by Shubhojit Banerjee on: "Accelerating material design for energy applications with molecular dynamics and machine learning."

Candidate Name: Shubhojit Banerjee
Degree: Doctoral
Defense Date: Friday, Sept. 12, 2025
Time: 2 - 3 p.m.
Location: Perry Hall, Room 415

Committee:
Advisor: Stephen T. Lam, Ph.D., Assistant Professor, Chemical Engineering, UML

Committee Members

  • Jerome Delhommelle, Ph.D., Associate Professor, Chemistry, UML
  • Zhiyong Gu, Ph.D., Chair, Professor, Chemical Engineering, UML
  • Rajni Chahal, Ph.D., Assistant Professor, Mechanical and Nuclear Engineering, Tennessee Tech University

Abstract:

Advancing materials for clean energy systems requires accurate, generalizable models capable of capturing complex atomic interactions and transport phenomena. In this work, we combine ab initio and machine learning methods to deepen fundamental understanding and to develop tools that accelerate the materials-development cycle.

To that end, first, ab initio molecular dynamics (AIMD) simulations were employed to investigate Group I and II molten fluorides, with a focus on assessing the impact of dispersion corrections on structural, thermodynamic, and thermophysical properties. This demonstrates that dispersion interactions are essential for reproducing experimental densities and local coordination environments, particularly in highly polarized salts such as BeF₂.

The real-space dynamic behavior of molten salts was then explored. AIMD was implemented to model the Van Hove function (VHF), unravel its partial contributions, and elucidate the underlying ionic transport mechanisms. Slow decorrelation is revealed for oppositely charged ions (Mg²⁺ and Cl⁻), caused by ion exchange across the solvation shell between adjoining ionocovalent complexes. Furthermore, transport coefficients are accurately recovered, and clear connections between macroscopic properties and ion dynamics are established. This study demonstrates the potential of ab initio-informed VHF to resolve long-standing challenges in uncovering relationships between picosecond-scale ion dynamics, mechanisms, and emergent physical properties of molten salts.

A robust and transferable Neural Network Interatomic Potential (NNIP) is developed for the LiF–BeF₂ binary system using the DeepMD-smooth edition framework. The potential was trained on a diverse dataset spanning a wide range of temperatures and compositions, and its performance was validated against data beyond the training domain. The choice of the training composition was further rationalized by understanding the neural network’s high-dimensional latent space in a 2D space. The resulting model successfully reproduces experimental salt structures and captures key thermodynamic and transport properties relevant to molten salt applications. Overall, this work lays the groundwork for building accurate and transferable NNIPs for chemically complex ionic systems, enabling large-scale simulations of molten salts in energy-relevant environments.

Finally, while NNIPs successfully speed up the materials-development process, they remain impractical for screening large chemical spaces. Herein, a machine learning framework was developed to predict metal–ligand stability constants (log β for Ga³⁺ complexes. Trained on over 1,800 curated complexes and validated via DFT, the models enabled high-throughput screening of ligands, identifying several promising Ga-selective candidates for critical-material separation.

Together, this work demonstrates a multi-scale, data-driven modeling strategy that accelerates the design of materials, advancing both reactor materials and separation technologies critical to clean energy deployment.