Self-Supervised Scientific Machine Learning Framework for Data-Scarce Simulations
Keywords:
data-efficient learning, masked reconstruction, temporal prediction, computational simulations, physics-constrained optimization, generalization in low-data regimes.Abstract
Scientific simulations are an important tool in predicting complex physical systems but tend to be limited by the quality of available labelled data. In data-sparse cases, traditional supervised machine learning methods cannot attain accurate and efficient generalization. In this paper, a new proposal of self-supervised scientific machine learning framework has been provided to effectively utilize scarce labelled data to learn the model by using a vast amount of unlabelled simulation data. The proposed method is based on self-supervised learning schemes, such as masked field reconstruction and temporal prediction, and physics-constrained loss functions to make sure that these physical laws are followed. With data-grounded representation learning coupled with domain-constrained physical consistency, the framework can substantially boost predictive accuracy in the regime of data scarcity. Large-scale experiments performed on benchmark partial differential equation systems indicate that the proposed method has a higher accuracy, lower error metrics, and better generalization than traditional supervised models and physics-informed baselines, especially in the limited case of training data. The findings emphasize the efficacy of the suggested framework in facilitating data-efficient simulation learning, and this framework has been a promising solution in the next-generation scientific computing systems where the cost or data bandwidth is high or limited.
