A schematic. Three points labelled x1, x2 and x3 at the top, joined in a triangle, feed a neural network drawn as rows of nodes labelled 16,000, 60, 30 and 2. The two output nodes at the bottom are labelled conformal, in rose, and scale only, in grey.

Applied AI · August–December 2020

Machine learning applied to conformal field theory

A proof of concept that neural networks can tell conformal field theory data apart, with 95% test accuracy on spin-correlator data.

The short version

Chen, He, Lal and Zaz (2020) introduced a machine-learning approach to classifying conformal field theories from synthetic data. This computer-science-minor capstone, co-authored with a fellow student in autumn 2020, set out to reproduce it, see how well it held up, and push it further.

Role
Co-author, with Joydeep Naskar
Period
August–December 2020
Repository
github.com/naskarjoydeep/ml-cft
PythonNumPyKeras/TensorFlowscikit-learnNeural networks (MLP)matplotlibJupyterGit

View the repository →

My CS-minor capstone at NISER Bhubaneswar, done alongside physics coursework. The reference paper is arXiv:2006.16114.

How

  • Generated synthetic conformal-field-theory data with NumPy as the training data for every task.
  • Trained Keras/TensorFlow multilayer perceptrons (MLPs) on it, with scikit-learn for data splitting and matplotlib in Jupyter for the results. The spin-correlator classifier, two dense layers with dropout, reached 95% accuracy on 1,000 held-out test samples.
  • Kept the pipeline in Git on separate main, data-gen and model branches.

A line chart of accuracy in percent against training epoch, 1 to 25, for the spin-correlator classifier. Validation accuracy, a rose line with dots, starts at 67 and climbs unevenly and stays between 94 and 97 from epoch 17 on. Training accuracy, a dashed grey line, climbs from 56 to 93. A dotted horizontal line marks the test accuracy, 95.1% on 1,000 held-out samples. The spin-correlator classifier over its 25 training epochs, ending at 95.1% on held-out test data.