AI for an inverse problem: Physical model solving quantum gravity

Koji Hashimoto, Koshiro Matsuo, Masaki Murata, Gakuto Ogiwara, Daichi Takeda

Keywords: Machine Learning, ICML, AdS/CFT, Runge-Kutta Method

Abstract:

Mathematical inverse problems of determining a governing differential equation for given solution data remain a fundamental challenge.
To find a working example of AI for math, we provide a concrete example using a physical setup of a quantum gravity problem.
We present a novel sparse Neural Network (NN) model which is interpretable, to solve the inverse problem: the AdS/CFT correspondence.
According to the conjectured correspondence, a special condensed matter system on a ring is equivalent to a gravity system on a bulk disk. The inverse problem is to reconstruct the higher-dimensional gravity metric from the data of the condensed matter system.
We use the response functions of a condensed matter system as our data, and
by supervised machine learning, we successfully train the neural network which is equivalent to a scalar field equation on an emergent geometry of the bulk spacetime.
The developed method may work as a ground for generic bulk reconstruction, i.e. a solution to the inverse problem of the AdS/CFT correspondence.
From a technical perspective, to achieve better numerical control, our neural network model incorporates a novel layer that implements the Runge-Kutta method.

Submission Number: 11