Excluding Target Domain Improves AI Extrapolation in Physics Models
A new AI method, the Deconfounded Hierarchical Gate (DHG), improves extrapolation in physics-constrained deep generative models by addressing confounding variables and hierarchical physical constraints. Researchers found that excluding target-domain data during pretraining outperforms including it, a counter-intuitive result that enhances out-of-distribution performance. The DHG uses counterfactual estimation and backdoor adjustment to remove spurious temperature effects, applying coarse-to-fine constraints progressively. This work was published on arXiv under ID 2605.07485.
Key facts
- arXiv paper 2605.07485 introduces Deconfounded Hierarchical Gate (DHG).
- DHG addresses confounding variable problem in physics-constrained generative models.
- Excluding target-domain data during pretraining improves extrapolation.
- DHG uses counterfactual estimation via do-operator and backdoor adjustment.
- Coarse-to-fine physical constraints are applied progressively.
- The method identifies temperature confounding at each constraint level.
- Hierarchical gates reflect intrinsic physical inconsistency.
- Published as arXiv:2605.07485v1.
Entities
Institutions
- arXiv