Physics foundation models
Studying how models represent physical laws, from mechanistic discovery in orbital dynamics to differentiable solvers for electromagnetics.
Research Scientist · IHPC, A*STAR
Xinyu Yang
AI for science researcher working on physics foundation models, differentiable simulation, and agentic AI for scientific workflows.
Research thread
Structured inductive biases for trustworthy simulation, inverse design, and scientific decision support.
I am Xinyu Yang, a Research Scientist at the Institute of High Performance Computing (IHPC), Agency for Science, Technology and Research (A*STAR), Singapore, with a background in computational physics and simulation.
My research focuses on AI for science, particularly physics foundation models and agentic AI for scientific workflows. I work on methods that combine neural operators, graph-based formulations, structured physical kernels, and LLM-driven agents.
Across these systems, I care most about structured inductive biases: choosing architectures that align learning dynamics with physical principles, and turning them into trustworthy tools for simulation, inverse design, and scientific decision support.
Research themes
What the work is organized around
The current arc of my work connects AI for science with deployable systems for scientific use.
Structured physical kernels
Designing architectures where neural operators, graph-based formulations, and fixed update rules are chosen to match the structure of the underlying physics.
Scientific workflow systems
Building agentic and knowledge-centric tools that help researchers run simulations, organize sources, and turn outputs into reusable scientific infrastructure.
Selected work
Writing, models, and scientific systems
A small set of essays and systems that best captures how the research turns into working artifacts.
Research article · AI4X 2026
NeCLO: Differentiable Electromagnetics
A research-facing writeup on reframing Maxwell solvers as structured tensor programs, replacing graph overhead with convolutions and tensor slicing while retaining bit-exact fidelity to FDTD.
Read the articleResearch essay · Apr 2026
Drawing Auxiliary Lines
A technical essay on structured inductive bias, asking why GNN input enrichment helps long-context Transformers recover Newtonian structure better than optional cross-attention or output-side constraints.
Read the articleScientific platform
Urban Cooling Agent
A full-stack urban microclimate platform that lets users describe a scenario in plain English, then runs coupled CFD, solar, and thermal comfort analysis through a map-based workflow.
Workflow infrastructure
Personal Research Wiki
An Obsidian-based research workspace for collecting raw material, compiling it into linked markdown knowledge, and querying it through LLM-assisted workflows and exports.
View repositoryAgentic workflow system
Agentic A*STAR
A multi-agent organizational simulation where AI agents take on institutional roles, deliberate in public channels, delegate work, and expose scientific workflow coordination as a visible system.
View repositoryRecent activity
Recent writing and project activity
A compact timeline of the essays and systems currently shaping this homepage.
Apr 17, 2026
NeCLO: published a new article on differentiable electromagnetics, bit-exact tensor reformulations, and inverse design workflows for Maxwell solvers.
Apr 17, 2026
Drawing Auxiliary Lines: published a new essay on structured inductive bias for Newtonian discovery in long-context Transformers.
Apr 15, 2026
urban-cooling-agent: updated the full-stack simulation platform and live demo for conversational urban microclimate analysis.
Apr 15, 2026
personal-research-wiki: continued developing an Obsidian-centered research workflow with compile, search, ask, and export tooling.
Contact
Get in touch
Email is the best way to reach me for AI for science, simulation, and scientific workflow collaborations. Reach me at yang_xinyu@a-star.edu.sg . Based in Singapore and open to thoughtful technical conversations.