PhysicsBench: Unified Leaderboard for Engineering AI Models

The evaluation of generative and predictive AI models in engineering design and simulation has traditionally been fragmented, with models often assessed in isolation on academic datasets using inconsistent metrics. To address this, PhysicsBench has been introduced as a unified benchmark and leaderboard. This platform standardizes the evaluation procedure for both generative and predictive models. PhysicsBench covers seven generation and prediction tasks across 1D, 2D, and 3D domains, ranking 66 models on nine datasets. These datasets include industrial-scale CAD/CFD/FEA simulations and public references, configured into 28 distinct setups. A crucial aspect of PhysicsBench is its focus on realistic, limited data scales (from S to XL), contrasting with the unlimited training sets common in academic benchmarks. The benchmark employs a common metric suite that captures geometric fidelity using distributional distances, physical-field and scalar accuracy, and engineering-specific field and shape validity. BenchRank, a PageRank-based system over a head-to-head dominance graph, debiases correlated metrics and provides a comprehensive ranking, with computational cost presented separately. Findings indicate that an architecture's large-scale academic standing is a weak predictor of its small-data ranking, with the top model changing with data scale in six of seven tasks, and no single model leading more than one task. PhysicsBench aims to transform "state-of-the-art" claims into openly published, verifiable foundations for model selection.