Benchmark RadarPhysicsBench
PhysicsBench is a unified benchmark and leaderboard for generative and predictive AI models in engineering design and simulation. It spans seven tasks across 1D, 2D, and 3D domains, ranks 66 models on nine datasets, and uses a common metric suite including BenchRank for debiased ranking. Evaluation covers data scales from S to XL.
- Released
- 2026-08-25
- Readiness
- Runnable
- Primary field
- General AI
Why it matters
Existing evaluations of generative and predictive AI in engineering are isolated with inconsistent metrics. PhysicsBench provides a standardized leaderboard for model selection, revealing that academic performance poorly predicts small-data ranking, supporting data-efficiency decisions.
Motivation
Generative and predictive artificial intelligence models are increasingly used to generate geometry and to predict physical fields and scalar quantities in engineering design and simulation.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.