I haven't been too close to this recently, but a common technique is called "Model Output Statistics" where the output from these computationally-intensive phys

I haven't been too close to this recently, but a common technique is called "Model Output Statistics" where the output from these computationally-intensive physical models are fed into another model, trained on historical observations and tuned to your specific risk profile and timeline. These secondary statistical models can be neural nets, or simpler models. So you're not going to try to teach a CNN about Coriolis forces or boundary-layer mixing physics, but you can teach it that the physical model is more likely to be wrong in this situation, or that the physical model is consistently wrong in your specific location compared to your observations.