Embedding machine-learnt sub-grid variability improves climate model biases