Disclaimer: This is a minimal, open-source baseline implementation of the core idea behind FieldSeer I (Guo, 2025). It demonstrates the feasibility of using world models to predict spatiotemporal field evolution in 2D FDTD simulations. This version is intentionally simplified and does not include the full suite of physics-guided components, geometry-aware conditioning, or online adaptation mechanisms described in the paper — those are reserved for proprietary and patent-pending systems.
- Generates synthetic 2D FDTD waveguide data with circular inclusions.
- Trains a transformer+RNN world model to predict future field frames from a prefix.
- Visualizes real vs. predicted field evolution via animated GIFs.
- Handle multi-material or complex geometries (no GNN encoder).
- Guarantee long-horizon physical stability (no symlog/symexp or physics-based regularization).
- Support real-time geometry editing or online adaptation.
- Evaluate optical responses (R/T/A) — only pixel-level MSE is reported.
This repository serves as:
- A proof-of-concept for applying world models to computational electromagnetics.
- An open benchmark for the community to build upon.
- A minimal baseline for comparison with future work.
For the full, production-grade system with physics-guided stability and geometry-aware control, please refer to the upcoming publication:
FieldSeer I: Data-Efficient World Models for Long-Horizon Electromagnetic Dynamics.