MULTI-FIDELITY CLASSIFICATION USING DEEP LEARNING TO ACCELERATING THE PREDICTION OF LARGE-SCALE COMPUTATIONAL MODELS

The prediction of large-scale computational models plays a crucial role in various scientific and engineering domains. However, the computational cost associated with running these models can be prohibitively high, hindering their practical use and scalability. In this study, we propose a multi-fidelity classification framework that utilizes deep learning techniques to accelerate the prediction process of large-scale computational models. The framework leverages the concept of multi-fidelity modelling, where models of varying computational costs and accuracies are combined to achieve efficient and accurate predictions. We train a deep learning model on high-fidelity data, which provides the most accurate predictions but requires significant computational resources. we generate low-fidelity data using simplified models or reduced-resolution simulations, which are computationally less expensive but yield less accurate predictions. To train the deep learning model, we employ a combination of transfer learning and domain adaptation techniques. We initially pre-train the model on the high-fidelity data and then fine-tune it using the lowfidelity data, allowing the model to learn from both fidelity levels. This approach enables the deep learning model to capture the complex relationships present in the high-fidelity data while benefiting from the efficiency of the low-fidelity data. Through extensive experiments on a large-scale computational model dataset, we demonstrate that the proposed multi-fidelity classification framework significantly accelerates the prediction process while maintaining a high level of accuracy. The trained deep learning model achieves comparable performance to running the high-fidelity model alone, but at a fraction of the computational cost. The proposed multi-fidelity classification framework demonstrates the effectiveness of deep learning techniques in accelerating the prediction process of large-scale computational models. By combining high-fidelity and low-fidelity data, the framework achieves efficient and accurate predictions while significantly reducing the computational cost. This approach has implications for various scientific and engineering applications, enabling faster simulations, optimization, and decision-making processes. Future research can explore the application of this framework to different domains and investigate methods to further improve the fidelity-efficiency trade-off.