Thesis abstract:

«p»«p»In recent years, there has been rapid and exciting progress in modern machine learning techniques, with applications spanning various fields leading to significant successes. Specifically in scientific research, machine learning has shown promising potential to revolutionize multiple domains, ranging from bioscience to quantum physics. Aligned with this trend, we aim to contribute further from three perspectives in the thesis. Firstly, machine learning holds the potential to help scientists reduce experimental costs and significantly accelerate the research process. Collaborating with domain experts, we will engage in research within quantum physics and materials chemistry, employ advanced learning algorithms to find ground states of quantum systems, explore chemical reaction pathways, enhance fuel cell designs, and solve other intriguing problems. Secondly, although neural networks possess remarkable fitting capabilities, their lack of interpretability presents a significant challenge in understanding how decisions are made, thus hiding new scientific principles. To address this issue, we plan to propose new learning algorithms that better balance interpretability and flexibility. In particular, our focus will be on causal learning, which has effectively uncovered hidden mechanisms and aided decision-making. Thirdly, real-world scientific research data often exhibits complex structures, such as symmetric features, ambiguous labels, or even unknown factors. To tackle these challenges, we will develop novel pre-processing and learning algorithms tailored to complex environments, aiming to improve the robustness and effectiveness of scientific research methodologies.«/p»«/p»