How thuml/transolver works
Project type and purpose
• This is an open-source research code base (“Transolver”) for solving partial-differential equations (PDEs) on general geometries using a Transformer-based neural operator.
• It provides a core library (the “Transolver” operator), plus three experiment suites:
· Standard benchmark PDE problems
· Car design on ShapeNet
· Airfoil design on AirfRANS
File structure
• Top-level items:
· Physics_Attention.py ‑ the core implementation of the Transolver operator
· README.md, LICENSE, requirements.txt, params.yaml
· Three experiment folders:
– PDE-Solving-StandardBenchmark
– Car-Design-ShapeNetCar
– Airfoil-Design-AirfRANS
• Each experiment contains:
· dataset/ (data loaders and preprocessing)
· models/ (experiment-specific model wrappers, including Transolver variants)
· train.py, main.py, main_evaluation.py (training and evaluation entry points)
· scripts/ (shell scripts to launch experiments)
· utils/ (metrics, normalization, naca_generator, etc.)
· fig/ (output plots)
• Shared configuration: params.yaml at root; requirements.txt in each folder
README for additional insights
• The README emphasizes:
· Core operator in Physics_Attention.py
· Modular experiments in their own subdirectories
· Use of Python, PyTorch, YAML configs, shell scripts for orchestration
· Outputs: figures, tables, metrics
Detail
• That Params.yaml feeds into all train.py and main_evaluation.py
• That scripts/*.sh wrap command-line calls to the training and evaluation scripts
• That fig/ directories contain the final visual results