AI-Newton

This is a proof-of-concept implementation of AI-Newton, a concept-driven discovery system that can formulate general physical laws in an unsupervised manner without any prior physical knowledge. AI-Newton integrates an autonomous discovery workflow, powered by a knowledge base (KB) and knowledge representation. The KB is responsible for storing and managing structured knowledge, including symbolic concepts, specific laws, and general laws. The knowledge representation is a physical domain specific language (DSL) that allows for the representation of physical concepts and laws in a structured and formalized manner. Given a collection of physical experiments, AI-Newton can formulate symbolic general laws applicable across a wide problem scope without neither supervision nor any prior physical knowledge. As a proof-of-concept implementation, it can rediscover Newton's second law, law of gravitation, conservation laws and others in classical mechanics.

Installation

The program currently only supports Linux. CUDA is also currently required.

Pre-requisites:

Python 3.11.9 or higher
Rust 1.84.0
Maple 2024

Conda is strongly recommended for managing an independent python environment for AI-Newton.

Installation via Conda and Cargo:

1. Clone the repository:

git clone https://github.com/Science-Discovery/AI-Newton.git
cd AI-Newton

2. Create a virtual environment and install the dependencies:

conda create -n ainewton python=3.11
conda activate ainewton
pip install -r requirements.txt

3. Compile Rust libraries:

cargo build --release
cp target/release/libcore.so aiphy/core.so

4. Run the bench-marking test: You can run aiphy/test_human_operating-benchmark.ipynb to test the installation. It will take about 10-20 minutes to run. This test is not just for verifying the installation, but more importantly, it serves as a minimal working example to demonstrate how AI-Newton discovers and constructs important physical laws from Newtonian mechanics. An example output of this test can be found in the data/human/example_output directory.
5. Run the test case: Since it may take several days to run the test case, it is recommended to run in the background:

nohup python -m test.examples.test_example_1 > logs/test_example_1/test_example_1.log 2>&1 &

Some of our test results can be found in the data/test_cases/example_1 directory while logs can be found in the data/test_cases/logs directory. In which the *_knowledge.txt files are the most important. They are used to store the core part of the KB, including physical concepts and general laws discovered. The test results were generated on two configurations: (1) an Intel Xeon Platinum 8370C (128 threads @ 3.500GHz) with an NVIDIA A40 GPU, and (2) an Intel Xeon Silver 4314 (64 threads @ 3.400GHz) with an NVIDIA GeForce RTX 4080 GPU.

Citation

@article{Fang:2025fmv,
    author = "Fang, You-Le and Jian, Dong-Shan and Li, Xiang and Ma, Yan-Qing",
    title = "{AI-Newton: A Concept-Driven Physical Law Discovery System without Prior Physical Knowledge}",
    eprint = "2504.01538",
    archivePrefix = "arXiv",
    primaryClass = "cs.AI",
    month = "4",
    year = "2025"
}