03/24/2026
By Danielle Fretwell

The Francis College of Engineering, Department of Electrical and Computer Engineering, invites you to attend a Doctoral Dissertation defense by Yi Huang on: "Automating Meta-Optics Discovery: Harnessing AI through Physically Grounded Agentic Workflows"

Candidate Name: Yi Huang
Degree: Doctoral
Defense Date: Tuesday, April 7, 2026
Time: 10 a.m. - noon
Location: Perry 115

Committee:

  • Advisor: Hualiang Zhang, Professor, Electrical and Computer Engineering, University of Massachusetts Lowell
  • Viktor A Podolskiy, Professor, Physics, University of Massachusetts Lowell
  • Alkim Akyurtlu, Professor, Electrical and Computer Engineering, University of Massachusetts Lowell
  • Xingwei Wang, Professor, Electrical and Computer Engineering, University of Massachusetts Lowell

Brief Abstract:
Metasurface inverse design holds considerable promise for engineering electromagnetic responses on demand. However, it requires deep expertise in computational electromagnetics, specialized simulation software, and extensive manual tuning, creating barriers that limit broader adoption and cross-disciplinary innovation. The rapid advances in artificial intelligence (AI), particularly the emergence of large language models (LLMs), have opened new possibilities for AI-assisted and autonomous scientific discovery. However, how to effectively architect AI-driven workflows that combine efficient simulation, intuitive interaction, and autonomous design intelligence for photonic discovery remains largely unexplored.

The eigendecomposition at the heart of rigorous coupled-wave analysis (RCWA) imposes a prohibitive per-step cost on gradient-based metasurface inverse design. To address this bottleneck, this dissertation introduces TorchRDIT. This fully differentiable electromagnetic solver replaces eigendecomposition with a Taylor series approximation, enabling end-to-end gradient computation throughout the simulation pipeline. Implemented in PyTorch with GPU-accelerated batch evaluation, TorchRDIT achieves an order-of-magnitude speedup over conventional RCWA while maintaining comparable accuracy, enabling end-to-end inverse design grounded in rigorous full-wave physics.

Despite this computational advance, performing inverse design continues to require specialized knowledge in computational electromagnetics and proficiency in solver-specific programming, thereby restricting accessibility to researchers outside the field. Recent LLMs have demonstrated code-generation capabilities approaching human-level performance on routine programming tasks. Building on this, this dissertation leverages the Model Context Protocol (MCP) to provide a standardized interface through which an LLM can autonomously retrieve domain knowledge and invoke simulation tools, enabling users to specify design objectives in natural language and obtain complete inverse design solutions without writing solver code.

The broader AI community is advancing agents from single-turn tool invocation toward systems capable of autonomous planning, reasoning, and continual learning with progressively less human supervision. This dissertation introduces a self-evolving agentic workflow in which the agent autonomously generates, evaluates, and refines persistent skill artifacts through closed-loop electromagnetic simulation feedback. By distilling error patterns and successful strategies into reusable expertise, the framework improves design pass rates, lowers computational cost, and progressively reduces human intervention across heterogeneous tasks. Together, the three layers presented in this dissertation, efficient differentiable simulation, language-driven solver access, and self-evolving design intelligence, chart a viable path toward automating meta-optics discovery and offer a generalizable blueprint for AI-driven autonomous discovery across scientific domains.