AutomataGPT: Transformer-Based Forecasting and Ruleset Inference for Two-Dimensional Cellular Automata.

Area of Science:

• Computational Science
• Artificial Intelligence
• Complex Systems

Background:

• Cellular automata (CA) model complex spatiotemporal patterns from simple local rules.
• Discovering CA rules and predicting their behavior from data remains a significant challenge.

Purpose of the Study:

• To develop an AI model, AutomataGPT, capable of inferring CA rules and forecasting their dynamics.
• To assess the generalization capabilities of large-scale pretraining for CA rule inference and state forecasting.

Main Methods:

• Pretraining a decoder-only transformer (AutomataGPT) on ~1 million simulated CA trajectories across 100 distinct 2D binary CA rules.
• Evaluating AutomataGPT on previously unseen CA rules for one-step state forecasting and rule matrix reconstruction.

Main Results:

• AutomataGPT achieved 98.5% accuracy in one-step state forecasts on unseen CA rules.
• The model demonstrated up to 96% functional accuracy and 82% exact match in reconstructing CA rules.
• Large-scale pretraining enabled significant generalization for both forward (forecasting) and inverse (inference) CA problems.

Conclusions:

• Transformer models can accurately infer and execute CA dynamics solely from data.
• AutomataGPT paves the way for interpretable CA surrogates in various scientific domains.
• This work opens new avenues for AI-driven scientific discovery in biology, physics, and engineering.