MAGIC: Modular Auto-encoder for Generalisable Model Inversion with Bias Corrections

Scientists often use models to interpret physical processes and uncover the underlying causes of natural phenomena. However, these models, due to necessary simplifications, can exhibit systematic biases that distort their predictions compared to actual observations. Traditional methods for model inversion, such as Bayesian inference or regressive neural networks, may either overlook these biases or make specific assumptions about their nature, which can lead to inaccurate results.

To address this, the project proposes a novel approach inspired by inverse graphics: it replaces the decoder stage of a standard autoencoder with a physical model followed by a bias-correction layer. This method allows for simultaneous inversion and bias correction in an end-to-end manner, without making strong assumptions about the biases. It validates the approach using two diverse physical models: a radiative transfer model from remote sensing and a volcanic deformation model from geodesy. The method achieves results that match or exceed those of traditional approaches, offering a promising solution for accurately understanding physical processes and their underlying causes.