This isn't something so fanciful as fusion power, this is reasonably something that might be within the capabilities of object detection transformers. Whether a different prompt/finetuning with a good dataset could make this work is very relevant here.