Discover physical concepts and equations with machine learning

Abstract: Machine learning can uncover physical concepts or physical equations when prior knowledge from the other is available. However, these two aspects are often intertwined and cannot be discovered independently. We extend SciNet, which is a neural network architecture that simulates the human physical reasoning process for physics discovery, by proposing a model that combines Variational Autoencoders (VAE) with Neural Ordinary Differential Equations (Neural ODEs). This allows us to simultaneously discover physical concepts and governing equations from simulated experimental data across various physical systems. We apply the model to several examples inspired by the history of physics, including Copernicus' heliocentrism, Newton's law of gravity, Schrödinger's wave mechanics, and Pauli's spin-magnetic formulation. The results demonstrate that the correct physical theories can emerge in the neural network.

• The paper demonstrates a novel ML framework integrating VAEs with Neural ODEs to autonomously rediscover physical laws from observed data.
• It reconstructs established theories—from heliocentric motions to Newtonian gravitation and the Schrödinger equation—using minimal latent dimensions.
• The research paves the way for AI-driven autonomous theory formulation, suggesting transformative applications in both classical and quantum physics.

Discover Physical Concepts and Equations with Machine Learning

Introduction

The paper "Discover Physical Concepts and Equations with Machine Learning" (2412.12161) discusses an innovative integration of machine learning techniques specifically designed to discover physical concepts and equations. This research extends SciNet by leveraging Variational Autoencoders (VAEs) alongside Neural Ordinary Differential Equations (Neural ODEs) to uncover underlying physical laws from experimental data. It sets out to demonstrate the capability of this model across a range of classical and quantum physics scenarios, showcasing the model's ability to reconstruct established theories without prior physical knowledge.

Methodology

The foundation of this approach lies in aligning machine learning processes with the human strategy of physical theory formulation, which traditionally begins with observing data and crafting theoretical models. By extending the SciNet architecture, this paper integrates VAEs to decode observational data into latent space representations and employs Neural ODEs to simulate the evolution of these representations as governed by differential equations.

Figure 1: Neural Network Architecture. Our model consists of two parts: Variational Autoencoders and Neural ODEs.

The encoder maps observational data to latent variables, which are then used by Neural ODEs to predict future states. This setup mimics the questioning mechanism of physicists, analyzing how different variables influence physical systems over time.

Critical Applications and Results

The efficacy of this model is demonstrated across four pivotal examples in physics:

Copernicus' Heliocentric Solar System

The model successfully rediscovered the heliocentric angles and their motion laws from geocentric angle data by employing a two-dimensional latent space optimal for system representation.

Figure 2: Heliocentric solar system. (a) Positions and motions of Earth and Mars relative to the Sun. (b) Impact of latent dimensions on loss, showing a two-dimensional space as optimal.

Newton's Law of Universal Gravitation

In simulating gravitational dynamics, the model identified two physical concepts as latent representations, corresponding to distance and velocity, and deduced Newton's fundamental gravitational equations from varied initial conditions.

Figure 3: Law of Universal Gravitation. (b) Distribution of distances $r(t)$ from the central body corresponding to different $r_0$. (c) Impact of latent dimensions on loss, showing that a two-dimensional space is optimal.

Quantum Wave Function and Schrödinger Equation

By analyzing trajectories of quantum particles in potential traps, the neural network autonomously discovered the concepts of wave function and its derivative, accurately capturing the Schrödinger equation.

Figure 4: The Schrödinger Equation Model. (b) Impact of latent dimensions on loss, showing a two-dimensional space as optimal. (c) Latent representations with a first-order differential equation store linear combinations of the wave function and its derivative.

Spin-1/2 and Pauli Equation

The experiment demonstrated AI's capability to reconstruct spin properties and related equations from experimental settings not directly revealing quantum behavior, exemplified by simulating Stern-Gerlach experiments under a uniform magnetic field.

Figure 5: The Pauli Equation Model. (b) Random potentials (green) and corresponding density distributions (red). (c) Impact of latent dimensions on loss, showing a four-dimensional space as optimal.

Implications and Future Directions

This research illustrates a significant step forward in using machine learning for physics discovery, reducing reliance on preconceived theories. By focusing on a data-driven approach, the model opens avenues for exploring uncharted physical theories and suggests potential improvements through enhanced model architectures and questioning mechanisms. Future extensions of this work could expand the framework to accommodate partial differential equations, enabling deeper exploration of complex systems such as in quantum field theory.

Conclusively, the paper posits that AI tools like the one proposed could redefine traditional paradigms of scientific research, potentially emerging as the fifth paradigm facilitating autonomous theoretical discoveries in physics.