Do Physics Foundation Models Learn Generalizable Physics? A Bias-Aware Benchmark Across Physical Regimes and Distribution Shifts

1. 1Construct a benchmark with diverse physical dynamics and training-data mixtures
2. 2Evaluate model performance across multiple test regimes
3. 3Analyze results to identify biases and areas for improvement
4. 4Apply the benchmark to existing physics foundation models
5. 5Refine models based on benchmark results

Researchers and developers working on physics foundation models can benefit from this benchmark to assess their models' generalizability, while data scientists can use it to identify potential biases in their models

💡 A single average score is not enough to determine a model's generalizability, a comprehensive benchmark is needed

🚀 Evaluate physics foundation models' generalizability with a bias-aware benchmark! 📊