Predicting Subsurface Pressure and Heat Flow Using Machine Learning on synthetic data

This study evaluates the efficacy of Multi-Layer Perceptron (MLP) and XGBoost machine learning models in predicting thermal flows and subsurface pressures using highly detailed synthetic data. This data meticulously simulates complex geological characteristics such as depth, pressure, temperature, and rock composition, offering a controlled environment for precise model testing. Our comprehensive results demonstrate that the MLP model significantly outperforms XGBoost, showing higher accuracy and a better fit, as evidenced by superior R-squared values and lower Mean Squared Error (MSE). The findings underscore the potential of MLPs to effectively handle complex, non-linear geological data and emphasize the critical importance of integrating mechanical engineering expertise with advanced machine learning techniques. This enhances geophysical predictions in vital applications such as geothermal energy development and earthquake risk assessment. This interdisciplinary approach not only optimizes the predictive models but also significantly enriches our strategic insights into subsurface phenomena, potentially revolutionizing how we explore and manage Earth's geological resources, driving forward innovations in both scientific understanding and practical applications.

Keywords:

Multi-Layer Perceptron , XGBoost , geothermal energy , machine learning