Tianxiang Dai
PhD candidate in Electrical Engineering, Stanford University
Representations for physical systems
- txdai@stanford.edu
- Phone
- (650) 460-9218
- Scholar
- Google Scholar
- ORCID
- 0000-0002-9403-7511
- Code
- github.com/txdai
- tx-dai
- CV
Research
Optimization and system design begin with the representation, not the model. I translate systems between physics, neural networks and language, and solve each problem in the form where it becomes simple.
News
2026-08
Started an NSF ACCESS allocation as principal investigator, for large-scale electromagnetic simulation on GPUs.
2026-06
Research internship at Apple, Hardware Technologies (Display Technologies), Cupertino, through September.
2026-05
Preprint: Language models fail at extended rule following, with a benchmark and a prompt generator.
2026-04
ICLR 2026 poster: the spatial kernel of multi-resolution hash encodings. Recognized as one of the top 200 reviewers.
2026-03
Nano Letters: Freeform mode-engineered metasurfaces.
2026-01
Laser & Photonics Reviews: Conformal aberration-correcting spaceplates.
2025-12
MetaChat in the Stanford Report. Received a Tinker research grant from Thinking Machines Lab.
Earlier
2025-10
2025-09
M.S. in Electrical Engineering, Stanford University.
2025-08
npj Computational Materials: Neuroshaper, geometric neural parameterization of freeform devices.
2025-06
Talk at the Optica Design and Fabrication Congress, Denver: neural level sets for high-Q nonlocal metasurfaces.
2024-07
ICML 2024 poster: general neural surrogate solvers with specialized neural accelerators.
Publications
* equal contribution
Software and data
Merged contributions to two open-source electromagnetic solvers: Tidy3D, Flexcompute's FDTD solver, and FDTDX, FDTD in JAX.
| Repository | Pull request | Change | Merged |
|---|---|---|---|
| flexcompute/tidy3d | #2607 via #2608 | Correct gradients for array-valued outputs of the PyTorch wrapper | 2025-07-01 |
| ymahlau/fdtdx | #19 | Diffractive detector and multi-frequency phasor | 2025-03-03 |
| ymahlau/fdtdx | #18 | Continuous devices beside discrete ones | 2025-03-03 |
| ymahlau/fdtdx | #17 | Temporal source profiles | 2025-03-03 |
| ymahlau/fdtdx | #13 | Periodic boundary conditions; a plotting fix | 2025-02-07 |
| flexcompute/tidy3d | #2202 | Twin-axis plot of n and k | 2025-01-30 |
| ymahlau/fdtdx | #12 | Automatic Metal backend on macOS | 2025-01-23 |
| flexcompute/tidy3d | #1989 | Distinct file names for forward and adjoint simulations | 2024-10-09 |
Bug report
pytorch/pytorch #180201: MPS autograd returned corrupted gradients for batched inputs.
Traced by a core maintainer to stride-0 matrix multiplication and fixed within a day (#180236, release 2.12).
Data
Data
FiLM WaveY-Net training and validation data Stanford Digital Repository, 2026
DOI 10.25740/dq123fg9049
Code
llm-extended-rule-following: the Stable Counting Capacity benchmark and prompt generator.
Service
Conferences
Reviewer, ICLR 2026, top 200 reviewers
Reviewer, ICML 2026
Journals
Reviewer for Science Advances, Nature Communications, Optics Express, Applied Optics and Photonics Research
Reviews deposited by the publishers on ORCID.
Talks and press
2025-12
New agentic AI platform accelerates advanced optics design
Stanford Report, on MetaChat
2025-06
Neural level set representation for high-Q nonlocal metasurface design
Optica Design and Fabrication Congress (Flat Optics), Denver, paper FTh1A.3
Grants and resources
2026 – 2027
NSF ACCESS Discover allocation, principal investigator
EMScale: high-memory GPU computing for large-scale electromagnetic simulation. ELE260067
2025 – 2026
Tinker research grant, Thinking Machines Lab
Compute for fine-tuning language models
Patents
2024
On-chip metasurface photonic diffractive neuron and its implementation
CN114819131B, granted 2024-06-21. X. Hu, Q. Gong, K. Liao, T. Dai; Peking University.
2022
All-optical nonlinear activation layer for optical neural networks and its implementation
CN112882307B, granted 2022-07-26. X. Hu, Q. Gong, K. Liao, T. Dai; Peking University.
Experience
Jun – Sep 2026
Research Intern, Apple, Cupertino
Hardware Technologies, Display Technologies
2023 – now
Doctoral Researcher, Fan Group, Stanford University
Theory-grounded scientific machine learning, neural fields, nanophotonics and autonomous design.
Education
2023 – 2028
Ph.D., Electrical Engineering, Stanford University expected
Advised by Jonathan A. Fan. Coursework in deep generative models, reinforcement learning, computational imaging, nanophotonics and electromagnetic waves.
2023 – 2025
M.S., Electrical Engineering, Stanford University
2019 – 2023
B.S., Physics and Computer Science (dual degree), Peking University
GPA 3.8 / 4.0
Honors
2023
Weiming Bachelor, Peking University
50 of 3,826 graduates
2021
Leo KoGuan Scholarship
7 of 207 in the School of Physics
2020, 2021
Merited Student, Peking University
2020
Second Class Scholarship, Peking University
2018
Silver Medal, 35th Chinese Physics Olympiad (final)