SENSE: SELF-SUPERVISED NEURAL EMBEDDINGS FOR SPATIAL ENSEMBLES

Analyzing and visualizing scientific ensemble datasets characterized by high dimensionality and structural complexity remains a significant challenge.While dimensionality reduction methods and autoencoders are widely used for feature extraction, their performance often degrades in high-dimensional settings.In this work, we propose an enhanced autoencoder framework that integrates clustering and contrastive loss functions into the latent space to improve the interpretability and visualization of scientific ensemble data.Clustering is guided by a soft silhouette score, encouraging compact and well-separated latent representations.To address the presence of unlabeled data, EfficientNetV2 is employed to generate pseudo-labels for partially unlabeled ensembles.The model is trained by jointly optimizing reconstruction, clustering, and contrastive objectives, resulting in improved grouping of similar samples and clearer separation of distinct structures in the latent space.UMAP is subsequently applied to the learned embeddings to produce two-dimensional visualizations, which are quantitatively evaluated using silhouette scores.We evaluate multiple autoencoder variants on two scientific ensemble datasets: subsurface channel structures generated via Markov chain Monte Carlo simulations and droplet-impact dynamics on a liquid film.The results show that incorporating clustering or contrastive objectives yields marginal but consistent improvements over baseline autoencoders.