Summary: We developed invertible neural networks to accelerate model calibration and simulation of computationally efficient earth system models.

Accomplishment: We proposed an interpretable Long Short-Term Memory (iLSTM) network for time series prediction. iLSTM enables interpretability of variable importance and variable-wise temporal importance for target predictions by exploring internal network structures. First, it enables hidden states to encode individual variables, such that the contribution from individual inputs to the prediction can be distinguished. Then, it uses a mixture attention mechanism to summarize the variable-wise hidden states and jointly learns the network parameters for prediction and the importance weights for interpretation. The proposed iLSTM not only results in accurate time series prediction, but also interprets the relative importance of inputs to outputs and their timescales of influence. Scientific machine learning (SciML) is not only for improvement of predictions but more importantly for enhancing predictive understanding and advancing scientific discovery. The insight into dynamic systems provided by iLSTM opens the black-box of the LSTM network and answers the question why it works, which improves our understanding and guides scientific model development.

Reference: https://ai4earthscience.github.io/iclr-2022-workshop/accepted