Data-Driven Reconstruction of Transient Lattice Dynamics in VO₂ Using Physics-Informed Deep Learning

Deep learning is increasingly applied to ultrafast diffraction and pump–probe experiments, where time-resolved measurements encode complex underlying physics. Leveraging these correlations through physics-informed architectures enables data-driven discovery of transient states. We present an iterative physics-informed deep learning framework for ultrafast electron diffraction (UED) studies of the photoinduced phase transition in VO₂, addressing the long-standing inverse problem of retrieving transient structures from incomplete, noisy data. The CNN–LSTM Encoder–Decoder model integrates convolutional layers with LSTM units for temporal correlations, yielding phase fraction dynamics and reconstructed Bragg peak intensities. An outer refinement loop couples retrieved phase fractions with lattice constants optimized from peak positions, minimizing a shared loss to capture how lattice distortions alter stacking and symmetry. The model recovers the transient polaronic phase fraction within 3% error and bridges UED structural shifts with optical dynamics. By exploring latent-space trajectories, it achieves robust inverse mapping of diffraction data to nonequilibrium coordinates, offering quantitative access to metastable phases in correlated materials.