My research develops artificial intelligence methods for scientific discovery, with semiconductor materials and devices as the central application. Beyond semiconductors, my research also spans AI for healthcare, LLM multi-agent simulation of marketing and consumer behavior, and energy-efficient AI systems.

Research Areas

1. AI for Semiconductor Materials and Device Design

Machine learning and physics-based simulation for wide-bandgap semiconductor device design, including III-nitride tunnel junctions, AlGaN deep-ultraviolet LEDs with superlattice electron-blocking layers, GaN/AlGaN heterostructure transistors, and β-Ga₂O₃ thin films and photodetectors, combining TCAD simulation (Silvaco, APSYS, Sentaurus) with machine learning surrogate models. Our machine learning frameworks have, for example, guided the design of asymmetric III-nitride tunnel junctions with substantially lower resistance than conventional designs. For two-dimensional materials, we work on property prediction and first-principles calculation, automated identification of exfoliated flakes, and modeling of 2D memristors and graphene diodes. Current work develops multi-fidelity and LLM-agent frameworks that link atomic-scale prediction, device simulation, and experimental characterization for automated device design.

2. Multimodal AI for Medication and Drug Recommendation

Interpretable, multimodal models that combine patient records, molecular structure, and drug-drug interaction knowledge to recommend safe and effective medications, using heterogeneous representations and mixture-of-experts architectures.

3. LLM Multi-Agent Simulation for Marketing and Consumer Behavior

Generative agents that simulate consumer decisions, social interaction, and marketplace dynamics to test marketing strategies before deployment, including visual LLM agents for advertisement analysis and audits of how model choice changes simulation outcomes.

4. Energy-Efficient AI Systems

Reducing the energy consumption of foundation models and agentic AI across the stack, from hardware and system architecture to agent orchestration and harness design, model fine-tuning, and smaller, more efficient models, together with benchmarks for evaluating agentic AI systems.

Current Projects

AI for MOCVD Growth of β-Ga₂O₃

Surrogate models and LLM-based growth agents that learn from multimodal characterization data (XRD, UV-Vis, XPS, SEM/AFM, Raman, and photoluminescence) to predict film quality and recommend growth conditions for β-Ga₂O₃ epitaxial films, in collaboration with the University of Bristol.

Machine Learning Design of AlGaN Graded Tunnel Junctions

A systematic machine learning framework for the design of asymmetric AlGaN graded tunnel junctions for ultraviolet optoelectronics, extending our earlier machine-learning-designed III-nitride tunnel junctions.

Graphene Geometric Diodes

Combining drift-diffusion simulation with deep neural networks to accelerate the design of graphene geometric diodes, in collaboration with KAUST.

LLM Agents for TCAD Device Simulation

LLM agents that set up, run, and analyze TCAD device simulations with Synopsys Sentaurus, linking device simulation with machine learning surrogate models for automated device design.

Gloss2Motion

Animating American Sign Language glosses into robot joint trajectories for humanoid sign-language translation (with Prof. Cameron LaMack).

Agentic Marketplace Simulation

Auditing how the choice of open-weight language model changes the outcomes of a frozen LLM agent market, and decomposing first-offer bias in agentic marketplaces with calibrated decision models.

QBench

A reproducible, continuously running benchmarking platform for coding-agent harnesses.

Energy-Efficient AI for Foundation Models

A review of energy efficiency across algorithms, hardware, and data center infrastructure, published in Computers 15(10), 660 (2026).

Interpretable Medication Recommendation

Heterogeneous-representation and mixture-of-experts models for safe, interpretable drug recommendation.