Optimizing Uncertainty Estimation on Scientific Visualizations Using Learning Models

Scientific visualizations (SciVis) translate numerical and spatial data as images, enabling scientists to better understand the phenomena described by the data and gain insights that may be overlooked by statistical methods alone. In recent years, machine learning, particularly Deep Learning (DL), has driven significant advancements in SciVis. By combining classical numerical techniques with DL methods, we can create models that potentially achieve the accuracies needed for real-world industrial applications and simulations. However, DL models often exhibit undue confidence, particularly when faced with inputs that differ significantly from the data they were trained on, known as out-of-distribution (OOD) inputs. This can lead to misclassifications where the model still produces high confidence scores, reflected in elevated softmax output values, even though its prediction is incorrect. Ideally, in such cases, the model should indicate uncertainty, effectively saying, "I do not know." In this work, we enhance the trustworthiness of DL models by quantifying their uncertainty and enabling selective classification, where the model abstains from making predictions when uncertainty is high. Instead of providing a single prediction, our model outputs a prediction distribution, allowing users to determine whether human intervention or other actions are necessary for handling high-uncertainty outputs. This paper evaluates uncertainty quantification across three different tasks: predicting airfoil pressure and velocity using a Reynolds-Averaged Navier-Stokes (RANS) model on a custom airfoil dataset, image classification using the ImageNet1K dataset, and digit recognition using the MNIST handwritten digits dataset.