INTRODUCTION OF LABORATORY

I’m investigating new frontiers in physics through the use of machine learning and artificial intelligence, specifically by applying advanced AI techniques to research in solid-state physics.
Physics has made tremendous strides over the 19th and 20th centuries, driven primarily by two branches: experimental physics and theoretical physics. In the latter half of the 20th century, computational physics emerged as a third branch, leveraging the power of computers to open up new avenues of discovery. Now, with the rapid expansion of AI and machine learning in the 21st century, a new “fourth branch”—“machine learning physics”—is beginning to take shape.
This new field centers on using machine learning to gain deeper insights into physics and, conversely, applying physics principles to enhance machine learning techniques. In the Nagai Lab, our main objective is to advance our understanding of physics by harnessing the power of machine learning.
In physics, theoretical models described by Hamiltonians are extremely precise in representing physical phenomena, but they often involve complex, demanding calculations. For instance, predicting electron dynamics through quantum mechanics is an especially challenging task. My research focuses on what might be possible if machines could handle these intricate calculations in place of human physicists.
We are developing methods where machines generate models and perform simulations autonomously. Although outcomes may vary, we have successfully developed a self-learning Monte Carlo method that produces highly accurate results. By combining machine-generated models with those developed by physicists, we aim to uncover new and previously unknown phenomena.

高次元の揺らぎが3次元空間に影響を与える様子の概念図
Credit: UTokyo ITC/Shinichiro　Kinoshita

アルミニウム原子の拡散経路（黄）と、高次元空間の揺らぎで出現する仮想アルミニウム原子（緑）

同変トランスフォーマーを用いた、スピン系に対するニューラルネットワーク

超伝導磁束近傍における局所状態密度の計算結果