Research

My research focuses on generative AI and machine learning for scientific discovery, with thermoset shape-memory polymers (TSMPs) serving as a primary testbed. I study how AI systems can move beyond conventional property prediction to generate, evaluate, and refine scientific candidates under multiple constraints. My work spans constrained generative modeling, large language models (LLMs), reinforcement learning, predictive modeling, and emerging tool-augmented agentic AI systems.

A central question across my research is how generative models can produce useful scientific candidates when several requirements must be satisfied simultaneously. In polymer design, these requirements may include target thermal and mechanical properties, reactive-group compatibility, chemical validity, novelty, and synthesizability. Rather than treating these considerations only as post-generation filters, I investigate methods that incorporate constraints and feedback directly into the design process.

My doctoral research has developed through several complementary directions. I first used conditional variational autoencoder (CVAE)-based generative modeling to incorporate reactive functional-group constraints into two-monomer TSMP generation. I subsequently developed Prompt2Poly, an LLM-based framework that translates natural-language design requirements into constraint-aware polymer generation and supports multi-turn refinement. My current work extends this direction through reward-guided reinforcement learning, where explicit evaluation signals are used to improve generation under multiple design objectives.

I am also investigating tool-augmented agentic AI for scientific design. In this setting, an LLM-based system can interact with specialized tools for tasks such as property prediction, chemical validation, novelty assessment, and candidate evaluation. The goal is not to assume fully autonomous scientific discovery, but to study how generation, external evaluation, and iterative refinement can be coordinated within more reliable scientific workflows.

Beyond candidate generation, I am interested in understanding the structure of sparse scientific datasets themselves. Through PolyPair, I study whether pair-specific interactions between reactive polymer components can be separated from the usual effects of the individual components and whether those interactions can be predicted from molecular structure. Through RePair-Gen, I am exploring reaction-aware generative models that modify polymer building blocks while preserving chemical compatibility and targeting desired material properties.

More broadly, I am interested in developing transferable, constraint-aware and feedback-driven AI methods that can extend beyond polymer informatics. Emerging directions include biomedical materials, medical text analysis, healthcare decision support, scientific imaging, and other domains in which reliable AI systems must operate under limited data, domain constraints, and human oversight.

Current Research

1. Reward-Guided LLMs for Constrained Polymer Generation

I am developing reward-guided methods for improving LLM-based thermoset shape-memory polymer generation. The framework evaluates generated candidates using criteria such as property alignment, chemical validity, functional-group compatibility, and novelty, and incorporates these signals into reinforcement learning using approaches such as Group Relative Policy Optimization (GRPO).

This work investigates whether explicit feedback can move scientific generation beyond one-shot prompting toward iterative improvement under multiple simultaneous constraints. A journal-oriented version of this research is also being extended with experimental validation of generated polymer candidates.

2. RePair-Gen: Reaction-Aware Generative Polymer Design

RePair-Gen investigates generative formulation refinement under explicit reaction and property constraints. Given an existing polymer formulation and desired changes in performance, the framework studies how one or both molecular components can be modified while preserving chemical compatibility.

The project explores reaction-conditioned, graph-based generative modeling and transfer between related polymer systems. Its broader goal is to develop AI methods that support property-constrained molecular editing rather than unrestricted generation from scratch.

3. PolyPair: Learning Pair-Specific Effects in Sparse Polymer Networks

PolyPair studies an important problem in sparse materials datasets: whether the behavior of a particular pair of polymer components can be distinguished from the usual contribution of each component individually.

The project investigates how molecular structure, reaction compatibility, and targeted data acquisition can improve learning when only a small fraction of possible component combinations have been experimentally observed. This work connects polymer informatics with graph-based learning, sparse experimental design, and structure-conditioned prediction.

4. Physics-Informed Machine Learning for Structural Dynamics

I am also exploring physics-informed neural networks (PINNs) for structural-dynamic response prediction. This work examines how governing physical equations can be incorporated into machine-learning models and how such models should be verified and validated beyond simply satisfying equation residuals.

This direction has also motivated a focused Mini Review on verification and validation practices for PINNs in forward structural dynamics.

Emerging Research Directions

My longer-term research goal is to develop AI systems that combine generation, prediction, external tools, feedback, and human expertise in a principled way. While polymer and materials design provide my primary experimental setting, many of the underlying methodological challenges are shared across other scientific domains.

I am particularly interested in extending these methods to biomedical materials, medical text analysis, scientific imaging, and healthcare decision support. My prior work also includes computer-vision research in healthcare, including deep-learning-based face-mask detection during the COVID-19 pandemic. I view healthcare as an emerging research direction rather than an extension of my current polymer systems, with future work requiring close collaboration with medical and domain experts.

Grants & Proposal Development

• Contributed AI methodology sections to interdisciplinary research proposals submitted to NSF and the U.S. Department of Energy (DOE) in AI-driven materials discovery, working under the supervision of the Co-Principal Investigator and with collaborators across universities and national laboratories.

Ongoing Research Projects

• Reward-Guided Thermoset Shape-Memory Polymer Generation through Reinforcement Learning in LLM-Driven Polymer Design. Developing GRPO-based reward optimization for property alignment, chemical validity, and novelty. Manuscript prepared; experimental validation of generated candidates is in progress.

• RePair-Gen: Reaction-Transferred Adaptive Dual-Graph GFlowNet for Property-Constrained 1W- and 2W-SMP Design. Developing a generative AI framework that modifies polymer building blocks while preserving chemical compatibility and targeting desired material properties. 

• Learning Structure-Conditioned Pair Effects in Sparsely Sampled Reactive Polymer Networks (PolyPair). Studying whether specific combinations of polymer components produce behavior beyond the effects of the individual components and how those interactions can be learned from sparse data. 

• Physics-Informed Neural Networks for Structural Dynamic Response Prediction. Investigating physics-informed learning for vibration and dynamic-response prediction, with emphasis on model accuracy, reliability, and validation. 

Manuscripts Under Review and in Preparation

• Reward-Guided LLMs for Constrained Scientific Generation: Lessons from Thermoset Shape-Memory Polymer Design. Under review at the AAAI Conference on Artificial Intelligence, 2027.

• RePolyGen: Reward-Guided LLMs for Multi-Constraint TSMP Generation. Manuscript prepared; experimental validation in progress prior to journal submission.

• Beyond Governing-Equation Residuals: Verification and Validation of Physics-Informed Neural Networks for Forward Structural-Dynamic Response Prediction. Mini Review under review at Frontiers in Mechanical Engineering for the Research Topic On the Application of Theory-Trained Neural Networks to Solid and Structural Mechanics.

• Toward Intelligent Telesurgery: Generative and Agentic AI for Predictive and Autonomous Surgical Assistance. Review manuscript in preparation.

Selected Publications

Peer-Reviewed Journal Articles — 2025

1. Borun Das, Andrew Peters, Guoqiang Li, and Xiali Hei.
   “Prompt2Poly: Ask, Specify, Create—A Dialogue-Based Large Language Model for Targeted TSMP Design.”
   Polymer Chemistry, 16(45), 4918–4933, 2025.

2. Borun Das, Andrew Peters, Guoqiang Li, and Xiali Hei.
   “Generative Design of Thermoset Shape Memory Polymers Driven by Chemical Group: A Conditional Variational Autoencoder Approach.”
   Journal of Polymer Science, 63(6), 1334–1344, 2025.

Conference Paper

2022

1. Borun Das, Mahmud Hasan, and Rafi Belal.
   “A Deep Learning Approach for Detecting Medical Face Masks on Human Faces in Response to COVID-19.”
   In Proceedings of the 11th International Conference on Robotics, Vision, Signal Processing and Power Applications, Springer, pp. 727–732, 2022.

Research Presentations

2024

• Borun Das, Guoqiang Li, and Xiali Hei.
  “Chemical Group-Driven Generation of Multi-Monomer Thermoset Shape Memory Polymers Using Generative Conditional Variational Autoencoder.”
  Louisiana Materials Design Alliance (LAMDA) Symposium, 2024.

2023

• Borun Das, Guoqiang Li, and Xiali Hei.
  “Unveiling Thermoset Shape Memory Polymers through RNN.”
  Louisiana Materials Design Alliance (LAMDA) Symposium, 2023.

Academic Service

• Reviewer, Cleaner Materials, Elsevier, 2026.

• Reviewer, Special Session on Anomaly Detection in Complex and Evolving Systems, 24th IEEE International Conference on Machine Learning and Applications (ICMLA), 2026.