Neural Spawning: Self-Growing Weather Models via Uncertainty-Triggered Specialist Birth with Independent Bayesian Self-Distillation

Neural Spawning is a method for letting a weather model grow itself over time. Here's the idea: you start with a trained base model (the "Mother") and freeze it permanently. Then, whenever the model encounters weather patterns it isn't confident about, it automatically creates a small specialist "child" network to handle that specific pattern — storms, heat waves, fog, or whatever it might be. Each child is tiny (around 189K parameters), trains itself in three phases by first imitating the Mother, then specializing on its specific pattern, and finally calibrating its uncertainty. Children that stop being useful get pruned; ones that become relevant again can be resurrected. The whole thing is managed through a routing tree so inference stays fast regardless of how many children exist. After a year of deployment, we project around 200 active children totaling about 53M parameters — all without ever touching the original model's weights. This approach avoids the catastrophic forgetting problem that plagues most continual learning methods, while continuously expanding what the system can handle.