Why it is Unfortunate that Linear Machine Learning 'Works' so well in Electromechanical Switching of Ferroelectric Thin Films

摘要

Machine learning (ML) is relied on for materials spectroscopy. It is challengingto make ML models fail because statistical correlations can mimicthe physics without causality. Here, using a benchmark band-excitationpiezoresponse force microscopy polarization spectroscopy (BEPS) datasetthe pitfalls of the so-called “better”, “faster”, and “less-biased” ML ofelectromechanical switching are demonstrated and overcome. Using atoy and real experimental dataset, it is demonstrated how linear nontemporalML methods result in physically reasonable embedding (eigenvalues)while producing nonsensical eigenvectors and generated spectra,promoting misleading interpretations. A new method of unsupervisedmultimodal hyperspectral analysis of BEPS is demonstrated using longshort-term memory (LSTM) β-variational autoencoders (β-VAEs) . Byincluding LSTM neurons, the ordinal nature of ferroelectric switching isconsidered. To improve the interpretability of the latent space, a variationalKullback–Leibler-divergency regularization is imposed . Finally, regularizationscheduling of β as a disentanglement metric is leveraged to reduceuser bias. Combining these experiment-inspired modifications enables theautomated detection of ferroelectric switching mechanisms, including acomplex two-step, three-state one. Ultimately, this work provides a robustML method for the rapid discovery of electromechanical switching mechanismsin ferroelectrics and is applicable to other multimodal hyperspectralmaterials spectroscopies.