MIT neural physics model simulates disasters never seen in training

MIT built a neural model that predicts structural material failure in extreme disasters it never saw in training data. Standard models fail on rare edge cases. Enforcing physical stress rules directly in model architecture accurately simulates unobserved physical catastrophes.