Can Neural Networks Enhance Physics Simulations?

Abstract

The primary objective of this research manuscript is to design, develop, and evaluate an artificial neural network architecture that is capable of emulating and predicting the dynamic interaction patterns manifested during the encounter between two distinct entities. This endeavor is primarily centered around computational learning and understanding of the associated physical impulses that emerge when these objects engage in contact, elucidating the complex physical interplays therein. This process incorporates the strategic use of an extant physics engine to generate the requisite training datasets, thereby providing a robust and comprehensive foundation for neural network training and subsequent performance evaluation. In order to scrutinize and substantiate the effectiveness of the proposed artificial neural network model, this investigation also embarks on a rigorous comparative analysis. The principal focus of this comparison is to juxtapose the results rendered by the trained neural network vis-a-vis those produced by the original physics engine. The goal here is to gauge the precision, reliability, and practicality of the trained model in accurately predicting the physical impulses, thereby demonstrating its potential to stand as a feasible alternative to the traditional physics engine. Despite the initial success of this endeavor, it is worth noting that the proposed neural network system managed to achieve a range of prediction rates, oscillating between 60% and 91%, contingent upon the specific test scenario. While these preliminary results are promising, they elucidate the necessity for further optimization and refinement to bolster the model's performance and prediction accuracy.