Uncover the Black Box of Machine Learning Applied to Quantum Problem by an Introspective Learning Architecture

Ce Wang, Hui Zhai, Yi‐Zhuang YouarXiv 2019 · arXiv

Abstract

Recently there is an increasing interest in applying machine learning algorithm to physics problems. These applications provide new platforms to challenge the critical issue of how to uncover the black box of machine learning because there always exists well-defined rules under physics problems. The success of such efforts can offer great promise of discovering new physics from experimental data by artificial intelligence. As a benchmark and a proof-of-principle study that this approach is indeed possible, in this work we design an introspective learning architecture that can automatically develop the concept of the quantum wave function and discover the Schrödinger equation from simulated experimental data of the potential-to-density mappings of a quantum particle. This introspective learning architecture contains a translator and a knowledge distiller. The translator employs a recurrent neural network to learn the potential to density mapping, and the knowledge distiller applies the auto-encoder to extract the essential information and its update law from the hidden layer of the translator, which turns out to be the quantum wave function and the Schrödinger equation. We envision that our introspective learning architecture can enable machine learning to discover new physics in the future.

Introduction

Machine learning (ML) has become a powerhouse in physics, helping researchers predict material properties or simulate complex systems. However, ML models are notoriously "black boxes." We often don't know how a neural network arrives at an answer or if it has "learned" the actual physics or just found a clever statistical shortcut. This work addresses a fundamental question: Can a machine, starting with no prior knowledge of physics, look at experimental data and extract human-understandable concepts like "wave functions" or "differential equations"? This is a crucial step toward using AI not just as a calculator, but as a partner in scientific discovery.

Methods: The Schrödinger Machine

The authors propose a two-part architecture they call the "Schrödinger Machine":

The introspective learning architecture consisting of a Translator (RNN) and a Knowledge Distiller (RAE).
  1. The Translator (RNN): This is a Recurrent Neural Network designed to map a 1D potential profile $V(x)$ to a particle density profile $ ho(x)$. The RNN respects translational symmetry—the laws of physics shouldn't change just because you move left or right. As it moves along the spatial sequence, it maintains an internal "hidden state" ($h_i$) to carry information forward.
  2. The Knowledge Distiller (RAE): This is the "introspective" part. It is a Recurrent Auto-Encoder that monitors the hidden states of the Translator. Its job is to compress those complex neural activations into the smallest possible number of variables (the "latent space") while still being able to reconstruct the original hidden states. By forcing the Distiller through an "information bottleneck," the machine is pressured to find the most parsimonious mathematical representation of the physics.

Datasets

The study uses simulated experimental data for a single particle in 1D:

  • Quantum Dataset: 15,000 pairs of potential $V(x)$ and density $ ho(x)$ profiles, including smooth, rough, shallow, and deep potentials.
  • Classical Dataset: For comparison, data from a high-temperature thermal gas (Boltzmann distribution) was used, where the density is a local function of the potential ($ ho \propto e^{-\beta V}$).
  • Training/Validation: 10,000 samples for training and 5,000 for validation.

Key Insights and Results

The architecture successfully rediscovered fundamental quantum mechanical principles through its internal optimization.

Analysis of the Knowledge Distiller showing reconstruction loss vs. latent dimension and the emergent wave function components.
  • The Information Bottleneck as a Discovery Tool: The researchers found that the Distiller's reconstruction error ($L_{RAE}$) only dropped to near-zero when the latent space dimension ($\tilde{d}$) was at least 2. This suggests the machine "realized" it needed exactly two variables to describe the system's state at any point.
  • Emergent Wave Functions: When examined, those two variables corresponded perfectly to the real and imaginary parts of the quantum wave function $\psi(x)$. Furthermore, by analyzing the learned weights, the authors showed the machine's internal "update rule" was mathematically equivalent to a discretized version of the Schrödinger equation.
  • Quantum vs. Classical "Memory": In the classical gas test, the Distiller found it needed zero variables to be passed along the sequence. This confirms the machine correctly identified that classical thermal density is local, whereas quantum mechanics is inherently non-local and requires a "wave" to propagate information.
  • Symmetry and Scalability: Because the RNN architecture respects translational symmetry, a model trained on a tiny "window" of space (50 units) could successfully predict the physics of a much larger system (400 units) without retraining.

Limitations and Future Work

While a landmark proof-of-concept, the study is currently limited to 1D single-particle systems. Scaling this to 3D or many-body systems (where electron-electron interactions occur) will be significantly more complex. The authors envision applying this "introspective" approach to experimental data where the underlying physics is currently unknown, potentially allowing AI to propose new equations for exotic states of matter or complex biological systems.

Summary updated

Abstract

Recently there is an increasing interest in applying machine learning algorithm to physics problems. These applications provide new platforms to challenge the critical issue of how to uncover the black box of machine learning because there always exists well-defined rules under physics problems. The success of such efforts can offer great promise of discovering new physics from experimental data by artificial intelligence. As a benchmark and a proof-of-principle study that this approach is indeed possible, in this work we design an introspective learning architecture that can automatically develop the concept of the quantum wave function and discover the Schr\odinger equation from simulated experimental data of the potential-to-density mappings of a quantum particle. This introspective learning architecture contains a translator and a knowledge distiller. The translator employs a recurrent neural network to learn the potential to density mapping, and the knowledge distiller applies the auto-encoder to extract the essential information and its update law from the hidden layer of the translator, which turns out to be the quantum wave function and the Schr\odinger equation. We envision that our introspective learning architecture can enable machine learning to discover new physics in the future.
BibTeX
@misc{wang2019uncover,
  author = {Ce Wang and Hui Zhai and Yi‐Zhuang You},
  title = {{Uncover the Black Box of Machine Learning Applied to Quantum Problem by an Introspective Learning Architecture}},
  year = {2019},
  eprint = {1901.11103},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1901.11103},
}
Figure