Research

My research focuses on reliable generative AI and machine learning for materials design, with thermoset shape-memory polymers (TSMPs) serving as my primary testbed. I study how AI systems can generate, evaluate, and refine scientific candidates when multiple constraints interact, feedback is incomplete or imperfect, and experimental validation is expensive. My work spans generative modeling, large language models (LLMs), reward-guided learning, predictive modeling, sparse-data learning, and tool-augmented AI systems.

My doctoral research has progressed through several complementary directions. I first developed a conditional variational autoencoder (CVAE)-based framework that incorporated reactive functional-group constraints into two-monomer TSMP generation. I then developed Prompt2Poly, an LLM-based framework for natural-language-guided, constraint-aware polymer design and multi-turn refinement. More recently, I developed and evaluated reward-guided LLM methods that combine chemical validity, compatibility, target-property alignment, novelty, and training-distance signals. This work also examines when reinforcement learning actually improves constrained scientific generation and when simpler supervised or sampling-based approaches remain competitive. Generated candidates from this research are currently undergoing experimental synthesis and characterization.

A major focus of my current research is completion-based process rewards for property-constrained molecular generation. Instead of providing feedback only after a complete molecule or formulation has been generated, I study whether meaningful intermediate design decisions can be evaluated using their downstream completion success. In the TSMP setting, a Planner proposes decisions such as reaction family, structural rigidity, network functionality, molecular flexibility, and composition ratio, while a Completer generates molecular candidates consistent with partial or complete plans. The resulting completion behavior provides a way to evaluate intermediate decisions according to whether they can ultimately lead to chemically valid candidates with the desired properties. More broadly, this project asks how process supervision can be constructed when intermediate reasoning steps do not have an obvious ground-truth label.

I am also investigating learning under sparse combinatorial data through my PolyPair project. Many reactive-material datasets observe only a very small fraction of the possible component combinations. PolyPair studies whether the effect of a specific component pair can be separated from the individual effects of its components, whether these pair-specific interactions are recoverable from sparse observations, and how molecular coverage, sampling topology, structural information, and targeted data acquisition affect what machine-learning models can reliably learn. This work is motivated by a broader question: what conclusions can we trust when scientific datasets contain many possible interactions but very few measured combinations?

A complementary direction of my research examines the reliability of physics-informed machine learning for structural-dynamic response prediction. Rather than treating low governing-equation residuals as sufficient evidence of model quality, I am interested in how physics-informed neural networks should be verified and validated against quantities that matter for actual structural prediction. This work has also motivated a focused review of verification and validation practices in physics-informed structural dynamics.

More broadly, I am interested in AI systems that remain reliable under interacting constraints, sparse data, imperfect feedback, evaluator disagreement, and costly validation. Although materials design provides my primary experimental setting, these methodological challenges arise across many scientific and engineering domains. I am therefore exploring emerging translational directions in biomedical materials, medical NLP and imaging, reliable clinical decision support, and generative and agentic AI for robot-assisted surgery and telesurgery. My current telesurgery work is review-driven and focuses on understanding how generative and agentic AI may support areas such as surgical scene understanding, latent-state estimation, and constraint-aware teleoperation.



Current Research

1. Completion-Based Process Rewards for Property-Constrained Polymer Generation

I am developing a completion-based process-reward framework for property-constrained thermoset shape-memory polymer generation. Instead of evaluating only the final generated candidate, the framework studies whether meaningful intermediate design decisions can be assessed using their downstream completion success.

In the current TSMP setting, a Planner proposes decisions such as reaction family, structural rigidity, network functionality, molecular flexibility, and composition ratio, while a Completer generates molecular candidates consistent with partial or complete plans. The goal is to determine whether intermediate decisions can receive useful process-level feedback even when no direct ground-truth label exists for each reasoning step.

2. PolyPair: Learning Pair-Specific Effects from Sparse Polymer Data

PolyPair studies whether the behavior of a particular pair of reactive polymer components can be separated from the individual effects of the components themselves.

The project investigates when these pair-specific effects are learnable under sparse combinatorial data and how molecular structure, component coverage, sampling topology, and targeted data acquisition affect what machine-learning models can reliably recover. The broader goal is to understand the limits of reliable prediction when only a small fraction of possible component combinations has been experimentally observed.

3. Physics-Informed Machine Learning for Structural Dynamics

I am investigating physics-informed neural networks (PINNs) for forward structural-dynamic response prediction, with particular emphasis on model reliability, verification, and validation.

Rather than treating low governing-equation residuals as sufficient evidence of model quality, this work examines how physics-informed models should be evaluated against physically meaningful response quantities and independent validation evidence. This direction has also motivated a focused Mini Review on verification and validation practices for PINNs in structural dynamics.

4. Reward-Guided LLMs for Constrained Polymer Generation

My reward-guided LLM research investigates how reinforcement learning can improve scientific generation under multiple simultaneous constraints, including chemical validity, functional-group compatibility, target-property alignment, novelty, and training distance.

This work also examines when reward-guided optimization actually improves over supervised fine-tuning and sampling-based alternatives. A journal-oriented extension is currently being strengthened through experimental synthesis and characterization of generated polymer candidates.

Emerging Research Directions

My broader research goal is to develop reliable generative and decision-making AI systems that can operate under interacting constraints, sparse data, imperfect feedback, and costly validation. While polymer and materials design provide my primary testbed, these methodological challenges arise across many scientific and engineering domains.

I am particularly interested in extending these methods toward biomedical materials, medical NLP and imaging, and reliable clinical decision support. I am also exploring generative and agentic AI for robot-assisted surgery and telesurgery through ongoing review work focused on areas such as surgical scene understanding, latent-state estimation, and constraint-aware teleoperation.

My earlier healthcare-related work includes deep-learning-based computer vision for medical face-mask detection during the COVID-19 pandemic. I view these biomedical and healthcare directions as emerging and translational areas that can build on my core work in reliable AI, while requiring close collaboration with clinical and domain experts.


Grants & Proposal Development


Academic Service