Researchers have found that the cost of imposing a physics prior in machine learning models is largely dependent on the ablation gap, which measures how much the model's performance drops when certain features are removed. They show that this cost is not a property of the prior itself, but rather of the free features and validation split used to evaluate the model. The study uses an ordinal wildfire-severity task as an example, where they find that imposing a physics prior can significantly improve performance, especially when certain meteorological drivers are constrained. --- Why it matters: This research is important for engineers working on machine learning models that rely on physical constraints, as it provides a new understanding of the trade-offs involved in enforcing these constraints. By identifying the ablation gap as a key factor, researchers can better design and optimize their models to achieve improved performance. Source: https://arxiv.org/abs/2608.21059