Research Scientist

Ankita Biswas

*****@********.*** Charlottesville, VA linkedin.com/in/ankita-biswas-b2312520/ github.com/Anny-tech Professional Summary

- AI & Materials Researcher with 5+ years of experience at the intersection of Computational Physics and Modern Machine Learning. Expert in electronic structure methods and data-driven optimization for complex structure-property relationships.

- Specialized in Geometric Deep Learning, including Equivariant Neural Networks (MACE) and Graph Neural Networks (GNNs), to accelerate quantum mechanical simulations (DFT, MD) with high-fidelity inference.

- Experienced in building Autonomous Discovery frameworks using Active Learning, Bayesian Optimization, and Agentic Workflows (LLMs/LangGraph) to bridge the gap between in-silico prediction and experimental synthesis.

- Proficient in Python, JAX and PyTorch for developing scalable high performance computing (HPC) workflows and ML accelerated discovery pipelines.

Technical Skills

- Machine Learning & AI: JAX, PyTorch, TensorFlow, Geometric Deep Learning (GNNs, Equivariant Net- works), Diffusion Models, Bayesian Optimization, Active Learning, Reinforcement Learning, LLMs (LangChain, Lang- Graph), RAG.

- Physics & Simulation: MACE (Equivariant MLIP), Quantum ESPRESSO, VASP, LAMMPS, ABINIT, Vesta, Pymatgen, ASE.

- Data & Infrastructure: Python, SQL, SLURM (HPC), Git, Materials Project, GNoME, NOMAD, NIST- JARVIS, AiiDA.

Education

PhD in Materials Sc. & Engg., University of Virginia 02/2021 to Present MS in Computational Materials Sc., RUB, Germany 10/2017 to 10/2020 BS in Ceramic Engineering, WBUT, India 08/2009 to 07/2013 Work Experience

Graduate Research Assistant, Dept. of Mat. Sc. & Engg, UVA Jan 2021 – Present

- Project 1 (End to End Discovery): Conceptualized a defect physics driven workflow for Bi2Te3 thermoelectrics. Per- formed 500+ high-throughput DFT calculations on HPC, utilizing PCA and auto-encoders for structure representation. Successfully linked local atomic structures within defect configurations to transport properties, creating a generalizable model for additively manufactured thermoelectrics.

- Project 2 (ML Accelerated Simulation): Developed a Constrained Diffusion-based surrogate model to replace expensive DFT structure relaxations. Utilized Equivariant Neural Networks (MACE) to achieve orders of magnitude faster inference while maintaining quantum level fidelity.

- Project 3 (Autonomous Frameworks): Designed an AI framework combining 1000+ micromagnetic simulations with Active Learning surrogates. Used black-box model interpretation and symbolic regression to identify analytical equations for skyrmion stability, enabling the design of low power memory storage.

- Project 4 (Intelligent Agents): Leading a multi-objective regression project using RAG optimized LLMs for scientific literature parsing and initial dataset generation. Built a pipeline consisting of exploratory data analysis, di- mensionality reduction (PCA), and multi-objective regression modeling to predict photonic and laser material properties to significantly reduce experimental characterization bottlenecks. Graduate Research Fellow, Oak Ridge National Laboratory May 2024 – August 2024

- Developed a Bayesian-based self-driven framework to automate virtual experimentation for 2D MoS2 heterostruc- tures. Integrated Molecular Dynamics (MD) with Active Learning to guide real-time experimental synthesis, demon- strating a closed-loop discovery cycle for quantum devices. Graduate Research Assistant, Ruhr University Bochum, Germany December 2018 – October 2020

- Performed Ab initio simulations (DFT-MD) to parameterize interatomic potential functions for ferroelectric perovskites.

- Developed Python-based ML models for automated analysis of experimental nano-indentation data. Graduate Research Assistant, Max Planck Institut for Iron Research, Germany April 2018 – September 2018

- Developed automated Python workflows to compare interatomic potentials based on vacancy interactions in bcc-Fe. Publications & Conferences

- Mentorship/NeurIPS: Jingyi Cui (M.S.)—short paper accepted at NeurIPS 2025 AI4Mat Workshop; I provided dataset, research direction, and editorial oversight (Constrained Diffusion for Accelerated Structure Relaxation of Inor- ganic Solids with Point Defects). Link

- Invited Speaker: Invited oral presentation at the 2025 MRS Fall Meeting & Exhibit (Boston, MA) in Symposium EN05: Advancements in Thermoelectric Materials—From Conventional to Cutting Edge. Title: “Uncovering Insights into the Role of Point Defects in Determining Carrier Type in Bi2Te3 Thermoelectrics Using Density Functional Theory and Machine Learning” (Session EN05.03: Theory and Computational Modeling of Thermoelectrics, 2025)

- Ankita Biswas, Prasanna Balachandran; Uncovering Insights into the Role of Point Defects in Determining Carrier Type in Bi2Te3 Thermoelectrics Using Density Functional Theory and Machine Learning; Under Preparation

- Soumendu Bagchi, Ankita Biswas et al.; Towards "on-demand" van der Waals epitaxy with hpc-driven online ensemble sampling; arXiv:2504.05539 [cond-mat.mtrl-sci], Under Review: npj Comput. Mater.

- Ankita Biswas, et al.; Integrating adaptive learning with post hoc model explanation and symbolic regression to build interpretable surrogate models.; MRS Communications 14, 983–989 (2024)

- Aris Dimou, Ankita Biswas, Anna Gruenebohm; Ab initio based study on atomic ordering in (Ba,Sr)TiO3; Phys. Status Solidi RRL, 18: 2300380 (2024).

- Callista M. Skaggs, Andrew P. Justl, Ankita Biswas et al.; Ultralow Lattice Thermal Conductivity in Metastable Ag2GeS3 Revealed by a Combined Experimental and Theoretical Study; Chemistry of Materials, 34(14):6420–6430, Jul 2022.

Open-Source Projects & Awards

- Scientific Discovery Agent: Developed a hypothesis generating platform for materials showing special features in Density of States (DOS) at fermi level using LLMs - Developed during the 2025 LLM Hackathon for Materials Science Chemistry. Published repo demonstrates autonomous hypothesis generation framework for scientific discovery of materials with high DOS at fermi level.

- Entrepreneurship: Awarded $15,000 by the Batten Institute (i.Lab) and won the UVA Entrepreneurship Cup for

’MatDash’, a Materials Informatics startup concept.