M-ENIAC: A machine learning recreation of the first successful numerical weather forecasts

TL;DR

This paper recreates the first successful numerical weather forecasts from 1950 using physics-informed neural networks, demonstrating improved accuracy and ease over the original ENIAC solver.

Contribution

It introduces a novel application of physics-informed neural networks to replicate and enhance early numerical weather prediction methods.

Findings

Physics-informed neural networks outperform ENIAC in accuracy.

Neural networks simplify the solution process for meteorological equations.

Recreation of historic weather forecasts using modern ML techniques.

Abstract

In 1950 the first successful numerical weather forecast was obtained by solving the barotropic vorticity equation using the Electronic Numerical Integrator and Computer (ENIAC), which marked the beginning of the age of numerical weather prediction. Here, we ask the question of how these numerical forecasts would have turned out, if machine learning based solvers had been used instead of standard numerical discretizations. Specifically, we recreate these numerical forecasts using physics-informed neural networks. We show that physics-informed neural networks provide an easier and more accurate methodology for solving meteorological equations on the sphere, as compared to the ENIAC solver.