Artificial intelligence has reached a point where it can outperform humans in pattern recognition and prediction across enormous datasets. Yet even the most advanced systems often fail when tasked with uncovering broad theoretical principles, such as universal physical laws. A research team in China now claims to have taken a meaningful step toward bridging that gap. Their new system, called AI-Newton, is designed not just to analyze data but to independently infer underlying scientific rules — including some of the most fundamental concepts in classical physics.
Researchers describe AI-Newton as an effort to replicate a key part of human scientific reasoning: the ability to move from observations to general principles. According to Yan-Qing Ma, a physicist at Peking University and one of the creators of the system, the model gradually builds its own internal library of concepts and relationships as it processes experimental information. Instead of being programmed with physics laws, it attempts to construct them from scratch, using only raw data. Ma notes that this capacity to identify and store meaningful abstractions could enable future AI systems to generate new scientific insights without being explicitly guided by humans.
A central technique behind AI-Newton is symbolic regression, a method that searches for mathematical expressions capable of describing a given dataset. Keyon Vafa, a computer scientist at Harvard University, explains that symbolic regression encourages the model to articulate physical phenomena in clear mathematical relationships rather than opaque numerical patterns. This makes it particularly suitable for scientific discovery, where the simplicity and interpretability of equations matter as much as predictive accuracy.
To evaluate AI-Newton, Ma’s team created a simulator that produced data from 46 classic physics scenarios. These included systems involving free-falling objects, elastic collisions, oscillations, spring-based motion, pendulum dynamics and more. The researchers deliberately added statistical noise to approximate the imperfections of real-world measurements. AI-Newton then worked through these problems one by one, using each result to enrich its internal knowledge.
In one example, the model was given data tracing the position of a ball over time. Without any prior information, it formulated an equation linking time and position, allowing it to derive the concept of velocity. This newfound understanding became a stepping stone for its next challenge: determining the mass of the ball using Newton’s second law, which connects force, mass and acceleration. The system successfully recovered this law from the noisy experimental data. Although the findings have not yet undergone peer review, they demonstrate how a model might incrementally reconstruct a coherent physical worldview from simple observations.
The idea of using AI to uncover planetary principles is not entirely new. In 2019, researchers at the Swiss Federal Institute of Technology in Zurich introduced a neural network called AI Copernicus. That system learned to produce mathematical expressions for planetary trajectories based on Earth-based observations. However, while the AI could generate equations, human scientists still needed to interpret those patterns to relate them to orbital mechanics.
More recently, Vafa and a team at the Massachusetts Institute of Technology performed their own experiments using modern foundation models — including large language models such as GPT, Claude and Llama. These models were trained on vast quantities of data, making them powerful tools for pattern recognition across many domains. Vafa’s group trained the models to predict planetary positions over time and then challenged them to infer the physical forces behind those motions.
The results were mixed. In their preprint findings, the MIT researchers reported that while foundation models performed well at predicting trajectories, they could not transform their learned patterns into meaningful physical laws. When the models attempted to derive a gravitational law from the orbital data, they produced equations that were mathematically incorrect and physically meaningless. Vafa argues that this reveals a key limitation of current AI architectures: although they can approximate solutions, they often encode knowledge in highly complex and non-intuitive ways. As a result, the internal representations they learn do not necessarily translate into human-readable scientific concepts.
David Powers, a computer and cognitive science researcher at Flinders University in Australia, emphasizes that uncovering laws from data is only one component of genuine scientific discovery. For AI to operate independently, it would need to participate in every stage of the process — identifying promising questions, designing experiments, determining relevant variables, collecting data and validating predictions. Powers notes that experimental science is fundamentally about choosing what to measure and why, something current AI systems are not equipped to do.
Ma agrees that AI-Newton is still far from functioning as a fully autonomous scientist. However, he sees the project as an early demonstration of what future models might achieve. By showing that an AI system can reconstruct foundational principles from experimental inputs, his team hopes to inspire the development of more capable scientific reasoning tools. The group is now testing whether AI-Newton can be applied to more complex domains, including quantum theory, where relationships are far less intuitive than those governing classical motion.
The emergence of systems like AI-Newton highlights a growing interest in combining machine learning with symbolic reasoning to advance scientific discovery. Many researchers believe that the next generation of AI tools will need to integrate intuitive abstraction with powerful numerical computation. Rather than simply predicting outcomes, these models will be expected to explain them using simple, interpretable structures. This could open new avenues for research in physics, chemistry, biology and other scientific fields where data is abundant but fundamental understanding remains incomplete.
If AI models eventually gain the ability to generate robust, general scientific theories, they could accelerate the pace of discovery in ways that were previously unimaginable. For now, systems such as AI-Newton remain proof-of-concept technologies. But their progress underscores a significant shift: artificial intelligence is beginning to move beyond pattern recognition and toward genuine conceptual reasoning. Whether these systems will someday rival human scientists in creativity and insight is still uncertain. What is clear, however, is that the frontier of scientific AI is expanding rapidly, and each new development brings us a step closer to machines that can uncover the hidden laws shaping the universe.
Leave a Reply