A microscopic process can take a direction forward or in reverse, and such a temporal direction is determined by the thermodynamic arrow of time. In fact, thermal fluctuations are a fundamental part of microscopic systems, as they occasionally cause a violation of the second law of thermodynamics, which is reflected in the work distribution. In this way, the arrow of time becomes blurred, making it difficult to distinguish between the forward process and the reverse process. For this reason, Machine Learning techniques have emerged as a strategy in which computers, based on mathematical models, are capable of learning information from a series of data and making predictions on new data. Based on the above, this project aims to explore the capacity of Machine Learning in the context of statistical physics and the reproduction of theoretical results.

A Machine Learning model based on logistic regression was used to predict the thermodynamic arrow of time in a spin chain system within a time-dependent magnetic field. Simulations were carried out at temperatures β⁻¹ = 10, 30, 50. In these cases, the theoretical estimation of the arrow of time was replicated by the model with an accuracy of 99.0%, 86.5%, and 73.9%, respectively. It was observed that accuracy decreased as the temperature increased due to the thermal effects on the spins. The model was partially able to relate the work of spin chains to the direction of the arrow of time. However, some additional normalization parameter needs to be adjusted so that the learning better fits the theoretical model.